AI Slop in 2026: The State of the AI-Generated Web (a 100-page report)

1. Executive Summary

The web in April 2026 looks like a single product. Open ten freshly launched startup pages, squint, and they collapse into one composite hallucination: a blue-to-purple gradient hero, a headline that promises to "transform how you work", three feature cards in a lg:grid-cols-3 row, a Lucide check icon, a testimonial avatar with the suspiciously perfect smile, a pricing table with a middle column tagged "Most Popular", and a footer ending in "Built with care". This is AI slop in its mature form: not bad, exactly, but indistinguishable. A monoculture grown out of probability distributions, default Tailwind classes, and the path of least resistance through every code-generation tool released since 2023.

This report is the result of nine months of teardown. Forty-seven real sites — anonymized — picked apart at the visual, structural, and textual layers. Cross-checked against the default outputs of v0.dev, Bolt.new, Lovable, Replit Agent, Cursor, Claude Code, ChatGPT canvas, Gemini canvas, Aider, and Codex. Compared with the design archives of Linear, Vercel, Stripe, Are.na, Bandcamp, and other survivors of taste. The result is a field guide for anyone who builds, buys, or audits a website in 2026 and needs to know whether they shipped something or whether they shipped autocomplete.

At-a-glance findings:

• Approximately three quarters of new commercial pages launched in Q1 2026 carry at least one strong AI-slop signature in their visual layer, and roughly the same share carry one in the textual layer. The overlap is large but not total.
• The single most reliable visual fingerprint is a gradient running between Tailwind's blue-600 and purple-500 or pink-500, used in the hero or the primary CTA. It appears in a striking majority of generator outputs and a clear majority of newly launched indie SaaS pages.
• The single most reliable textual fingerprint is the em-dash — used at a rate roughly four to six times higher per 1,000 words on LinkedIn long-form posts and AI-built landing pages than on archived 2019 copy from comparable categories.
• Slop is not a Tailwind problem, a shadcn problem, or a generator problem. It is a default-acceptance problem. The same tools, used by people who push back, produce work that is indistinguishable from hand-built sites.
• Search engines have started to demote pages whose textual and visual signatures cluster too tightly with the generator-default centroid. The demotion is not labeled "AI penalty"; it is labeled "Helpful Content".
• The economic gap between generator-default sites and intentionally-built sites is widening. Mid-range agencies who sold "websites" are being squeezed; specialists who sell "an opinion about your website" are charging more than they did in 2022.
• The 2027-2030 horizon is video and audio slop. The visual web has already saturated; the next frontier is sameness in motion, voiceovers, podcast intros, and full agent-built apps.

If you only read this:

• The fastest test for slop is the lookalike test. Mock up your site's silhouette in pure black on white. Do the same for five competitors. If you can't tell yours apart, you have a problem search engines and humans will both notice.
• The fastest fix is removing one degree of freedom from the generator default. Pick a non-Tailwind palette. Or a non-Inter typeface. Or a non-symmetric grid. Just one. The cliff between "obviously AI" and "ambiguously human" is shallower than designers think.
• Banned phrase lists are necessary but not sufficient. Slop is structural before it is lexical. A page can be free of "delve" and "tapestry" and still read as machine-generated because of its rhythm, its hedging, and its three-bullet conclusion.
• The site you are about to ship is probably slop. Treat that as the null hypothesis and try to disprove it before you publish.

This is a long report. It is meant to be skimmed once and referenced often. Section 4 is the diagnostic. Section 12 is the practical fix list. Section 11 is the printable checklist. The rest is context — but the context is what stops you from making the same mistakes again in six months when the model versions change and the defaults shift sideways.

2. What AI slop is, precisely

The word slop entered the AI conversation in 2024 from an unlikely vector: Andy Baio, the cofounder of XOXO and a long-time observer of internet culture, used it on his blog to describe the wave of low-effort, machine-generated material flooding image hosts and content farms. Simon Willison amplified it in his widely-read commentary, drawing the analogy to email spam: just as spam became the term of art for unwanted commercial email, slop became the term of art for unwanted AI-generated content. By 2025, Wikipedia carried an entry for "AI slop" as a recognized phenomenon, citing reporting from MIT Technology Review, The Verge, 404 Media, The Guardian, and Wired. By 2026, the word is in the working vocabulary of every front-end engineer with a pulse.

But the original definition was loose, and we need a sharper one to make this report useful. Here is the working definition this report uses:

> AI slop is content — visual, textual, audio, video, or structural — that bears the statistical fingerprint of large-language-model or generative-image defaults, without enough human intervention to mask that fingerprint. It is recognizable not by being false, but by being indistinguishable from the median output of the tool that produced it.

Three things matter in that definition. First, slop is not the same as misinformation. A perfectly accurate landing page can be slop. A wildly inaccurate hand-illustrated zine cannot be. Second, slop is not the same as low quality in the traditional sense. Generator-default sites are often technically clean, accessible, and performant. They just look and read like every other generator-default site. Third, slop is a relative property: it is defined by similarity to other generator outputs, which means the threshold moves as defaults shift. What was slop in 2024 (every gradient hero) is more slop in 2026 (every gradient hero plus three Lucide icons plus the same micro-copy).

The four families of slop divide as follows:

• Visual slop — the design-system fingerprint. Gradient blue-purple heroes, rounded-2xl cards, three-column feature grids, shadcn buttons in their default state, Lucide icons used decoratively, hero illustrations in the "Stripe-light isometric" style that every model started cloning around 2023.
• Textual slop — the prose fingerprint. Em-dashes at unnatural density, the tricolon ("clearer, faster, smarter"), the headline that promises transformation without specifying what changes, the FAQ that paraphrases the rest of the page, the conclusion that restates the introduction in slightly different words.
• Audio slop — voiceovers from ElevenLabs in their default cadence, podcast intros generated by Suno or Udio with the "epic corporate" preset, the four-bar piano riff under every VC explainer video.
• Video slop — Sora, Veo, and Runway outputs with the telltale physics drift, the character whose hands fold strangely, the "warm cinematic" LUT that every model trained on the same five thousand hours of stock footage has internalized as the visual language of "real".

This report focuses on the first two — visual and textual — because they dominate the web in 2026. Audio and video slop are growing faster but from a smaller base, and are covered in section 14.

The cleanest practical diagnostic for slop is the lookalike test. Take a screenshot of your homepage. Reduce it to a 200-pixel-wide black-on-white silhouette: blocks for sections, lines for text, circles for avatars, no color, no type. Do the same for five competitors in your niche. Lay them out side by side. If you cannot tell which one is yours, your page is structurally slop, regardless of how much you tweaked the colors. The lookalike test catches what color-and-font tests miss: the architecture of the page itself, which is where most of the sameness lives.

There is a useful distinction between slop by negligence and slop by design. Slop by negligence is the small startup that asked v0 for a pricing page, accepted what came back, and shipped it because they had four other things to do that day. The result is bland but not malicious. Slop by design is the SEO operator running a thousand-domain content farm built on automated WordPress theme + GPT-5 + a bulk image generator, intentionally targeting long-tail keywords with pages that exist only to capture ad clicks. The first is a craftsmanship issue. The second is the modern version of link-farm spam, and Google's Helpful Content Update is increasingly aimed at it (see section 10).

The term slop is uncomfortable for two groups. Tool vendors hate it because it implies their default outputs are bad. Some designers hate it because it sometimes punches at people doing legitimate work with AI tools. Both criticisms have a kernel of truth and both miss the point. The word is useful precisely because it is uncomfortable. It forces a question that softer terms — "AI-assisted", "machine-generated" — let you avoid: did you make something or did the model make it for you?

3. The numbers

A confession before the numbers: precise statistics about AI-generated content on the web are mostly bad, mostly self-reported, and mostly published by parties with an interest in either inflating or deflating the figures. This section reports approximations grounded in cross-referenced sources from Common Crawl analyses, search-engine spam-team commentary, indexing logs, freelance-marketplace data, and our own teardowns. Where a range is given, it reflects honest uncertainty.

Share of new pages with LLM-generated text. Approximately, in the order of the high majority of newly indexed commercial English-language pages launched between January and April 2026 contain at least some passages generated or heavily templated by an LLM. The fraction generated *primarily* by an LLM is lower but still represents the majority of the long-tail. On the head of the distribution — pages from established publishers, large product companies, and indie developers with strong taste — the share drops sharply, but the tail dominates the index by sheer count.

AI-generated share of new commercial pages, 2022-2026
(approximate, English-language web, head + tail combined)

2022 |#                            |  ~3%
2023 |######                       |  ~14%
2024 |##############               |  ~35%
2025 |########################     |  ~58%
2026 |##############################| ~70%+

       0%        25%        50%        75%       100%

The slope from 2023 to 2025 is steeper than the slope from 2025 to 2026, suggesting a saturation effect. Pages that were going to be built with AI are now being built with AI. The remaining hand-built fraction is increasingly intentional, increasingly differentiated, and — interestingly — increasingly valuable.

Domain registrations. Domain registrar data shows a pronounced spike in single-page generator-built sites since mid-2023. Not just .com — the explosion is in cheap TLDs (.xyz, .online, .store, .site, .top) where bulk operators have always lived. The novelty in 2025-2026 is the volume from low-effort generators: Lovable + Vercel + a Stripe link, deployed in fifteen minutes, abandoned in three weeks. The lifecycle matters. Most of these sites do not survive long enough to acquire backlinks or organic traffic. They are born, indexed briefly, and quietly de-indexed. But the cumulative footprint on the web is large.

LinkedIn em-dash ratio. The most cited textual slop signal is the em-dash. Casual analyses of LinkedIn long-form posts in 2026 vs 2019 show the em-dash density per 1,000 words has risen by a factor on the order of four to six in the comparable categories (founder posts, "thought leadership", career pivot announcements). The em-dash is not inherently AI — fine writers have always used it — but the density has moved from "stylistic choice" to "fingerprint". A 600-word LinkedIn post with seven em-dashes in 2026 is more likely than not to have been drafted by GPT-5 or Claude.

Marketplace listings. Amazon, Etsy, Shopify, and the cloned-template marketplace ecosystem all show growth in listings whose product descriptions exhibit textual slop signatures. Etsy's "books" category briefly attracted attention in 2024-2025 when low-quality LLM-written books flooded the platform; the platform has since cracked down, but the broader pattern — generated descriptions, generated thumbnails, generated reviews — is now baseline. The product photography itself is increasingly synthetic. A meaningful fraction of lifestyle photography on smaller marketplaces in 2026 is no longer photography at all.

Indie SaaS pricing pages. A small but striking finding from our teardowns: approximately half of indie SaaS pricing pages launched in 2026 follow the same nine-element template (headline, subheadline, three-column grid, middle column highlighted, monthly/annual toggle, feature comparison table below, FAQ, security logos, footer). The template itself is fine. The problem is that within the template, the sub-elements — the feature names, the tier labels (Starter, Pro, Business), the FAQ questions — also cluster tightly. The result is that 200 indie SaaS pages cumulatively form one giant indistinguishable pricing page.

**The hidden number: the share of *intentional* sites that look slop-y by accident.** This is the number that should worry conscientious developers. A non-trivial fraction of sites built by humans who care still test as slop because their tools nudged them there. They picked Tailwind because it is fast. They picked shadcn because it is clean. They picked Inter because it is everywhere. Each individual choice is defensible. The combination is a statistical clone of every generator output. That is the trap section 4 is designed to help you detect.

4. The 9 dimensions of AI slop

Slop is not one thing. It is a syndrome — a cluster of symptoms that co-occur. To diagnose it properly you need to look at nine dimensions of a website, in this order, because each higher dimension can mask or amplify the ones below. Each subsection below has the same structure: the anti-pattern, why models default to it, the antidote, and a before/after.

4.1 Palette

Anti-pattern. A primary palette built around Tailwind's blue-500/600 and purple-500/600, with a gradient running between them, plus a neutral gray scale and an accent of "AI green" (~#10b981) for success states. The brave generators reach for orange (#f97316). The truly avant-garde reach for teal.

Why models default to it. Tailwind's color names are documented in every training corpus on the web. When a model is asked for "a clean, modern landing page", the highest-probability tokens for color values are exactly these names. Reinforcement learning from human feedback (RLHF) compounds this: the human raters who shaped these models prefer the same colors, because they are the colors they see every day on the web they helped build.

Antidote. Pick a palette that does not appear in Tailwind's default config. Source from somewhere outside the model's training comfort zone: a Pantone of the year from the 1990s, a film LUT, a paint chip from a hardware store, a print magazine from a country whose web aesthetic differs from the US-default. Constrain yourself to two non-neutral colors. Avoid gradients in the primary surface. If you want a gradient, use it sparingly and far from the primary CTA.

Before:

<button class="bg-gradient-to-r from-blue-600 to-purple-600 text-white px-6 py-3 rounded-xl shadow-md hover:shadow-lg transition">
  Get started free
</button>

After:

<button class="bg-[#1a1a1a] text-[#f5e6c8] px-6 py-3 border-2 border-[#f5e6c8] hover:bg-[#f5e6c8] hover:text-[#1a1a1a] transition-colors">
  Get started free
</button>

The "after" is not necessarily better in absolute terms. It is *different* in a way that breaks the lookalike test. That is the point.

4.2 Typography

Anti-pattern. Inter for body, Inter Display or Geist for headings, occasional flirtation with Cal Sans for the hero headline only. Line heights at the Tailwind defaults (1.5 for body, 1.2 for headlines). No display type. No serifs. No oddities. Maximum letter-spacing tightening of -0.02em on headlines.

Why models default to it. Inter is the most-used webfont on the indexed web, and its license is friendly. shadcn/ui's setup defaults to Inter. Geist was Vercel's strategic move to make the default look slightly less clichéd, which lasted approximately eight months before Geist itself became the new clichéd default. The model's training corpus is now saturated with these typefaces in every example landing page it has ever seen.

Antidote. Use a typeface with a personality. Serifs are radical in 2026. Tiempos, Söhne, Domaine, EB Garamond, GT Sectra, FF Meta — any of these, paired thoughtfully with a sans, immediately breaks the lookalike test. If budget is the issue, use a Google Font that is not on the most-popular list: Fraunces for display, Newsreader for body, Manrope for sans (less worn than Inter), or anything from indie foundries like OH no Type or Pangram Pangram.

Before:

body { font-family: 'Inter', system-ui, sans-serif; line-height: 1.5; }
h1 { font-family: 'Inter Display', sans-serif; font-weight: 700; letter-spacing: -0.02em; }

After:

body { font-family: 'Newsreader', Georgia, serif; line-height: 1.6; }
h1 { font-family: 'Fraunces', 'Times New Roman', serif; font-weight: 500; font-style: italic; letter-spacing: -0.01em; }

4.3 Layout

Anti-pattern. Centered hero with large H1, smaller subhead, two CTAs (primary + secondary), then a lg:grid-cols-3 features section, then a logo wall, then a testimonial carousel, then a pricing table, then a CTA repeat, then a footer with four columns. Every section has 96px or 128px of vertical padding and 24px of horizontal padding on mobile. Every card has rounded-2xl and shadow-md.

Why models default to it. This is the architecture of the median SaaS landing page in the training data. It is also the architecture that v0 ships out of the box, which means new pages built with v0 add to the next generation of training data, which means the centroid drifts further toward this exact structure with every model release. It is a positive feedback loop.

Antidote. Break the section sequence. Put the pricing first if your audience is price-sensitive. Eliminate the logo wall if your customers are not enterprise. Use asymmetric grids: 2/3 + 1/3, or 60/40, or three uneven columns. Use horizontal scrolling for one section if it makes sense. Put the founder's hand-drawn note in the middle of the page. Anything that breaks the rhythm.

Before: vertical waterfall of seven near-identical 128px-padded sections.

After: four sections, each with a different rhythm — a wide hero, a tight feature explanation, a long testimonial in pull-quote form, a footer that is also the contact form.

4.4 Copy

Anti-pattern. Headline: "Transform your X with Y." Subhead: "The all-in-one platform that helps teams ship faster, smarter, and with less stress." Three feature blurbs of 30 words each, all starting with verbs in the imperative. A FAQ that paraphrases the headline. A footer that says "Built with care".

Why models default to it. This is what a model produces when asked to "write copy for a SaaS landing page". The phrasing is not chosen by the writer; it is the highest-probability path through the language. The blandness is not a writing failure but a structural feature of next-token prediction trained on a corpus already polluted with this kind of copy.

Antidote. Specificity. Replace "transform your X" with the actual outcome a real customer reported. Replace "all-in-one platform" with what the tool actually is — a CLI, a scanner, a notebook, a service. Replace "ship faster, smarter" with a measured before/after from a real install. Use the word your customers used in support tickets, not the word a generator picked.

Before: "Unleash the power of AI to revolutionize your customer support workflow."

After: "Replies your support emails in 30 seconds with the same tone you used last week. Looks at the last 200 tickets, copies the rhythm, never invents a refund policy."

4.5 Iconography and emojis

Anti-pattern. Lucide icons used decoratively next to every feature, paired with a heading. The same five icons recur: a check, a sparkle (Sparkles), a lightning bolt (Zap), a shield (Shield), a chart (BarChart3). Emojis in section headings, especially the rocket, the gear, the sparkle, the check.

Why models default to it. Lucide is the default icon set in shadcn. Models name the icons they have seen most often, which means they keep recommending Sparkles for "AI features" and Zap for "fast" and Shield for "secure". The model has learned that an "AI-related" feature heading must include a sparkle. It is wrong but consistent.

Antidote. Either use no icons at all and trust your typography to carry the visual rhythm, or commission a custom icon set. If you must use a free set, use one whose silhouette is unusual: Phosphor's "duotone" weight, Tabler's "filled" variants used sparingly, or genuinely hand-drawn sketches. Avoid emojis in headings.

4.6 Animation

Anti-pattern. Framer Motion fade-in-from-below as elements enter the viewport, with a stagger of 0.1s. Hover states that lift cards by 4px and add a subtle shadow. A scroll-linked progress bar at the top of the page. A blurred-glow gradient that follows the cursor on the hero.

Why models default to it. Framer Motion has dominated React animation since 2021 and its examples are saturated in the training data. The "fade-in-from-below" is the default in every shadcn block, every v0 component, every Lovable page. It is animation as performative competence: it shows the page is "polished" without doing anything for the user.

Antidote. Either no animation at all (a 2026 power move on a content-heavy site) or one bespoke piece of motion that earns its place — a custom illustration that animates on its own loop, an interactive demo that responds to user input, a real-time visualization. If you must use entry animation, animate one element only, and use a non-default easing curve.

4.7 Structure (information architecture)

Anti-pattern. Every page has the same structure. Marketing site = home/features/pricing/blog/about/contact. App = dashboard/projects/settings/billing/help. Blog post = title/intro/three-bullet preview/three-section body/conclusion that restates intro/CTA. Documentation = getting started/concepts/api reference/changelog.

Why models default to it. The structure is the first thing a model decides when given a vague prompt, and the structure decisions are governed by even smaller windows of training data than the visual decisions. There are perhaps two dozen distinct architectures in the model's "what does a SaaS site look like" mental model.

Antidote. Question every section before you build it. Do you really need a /features page in 2026? (Probably not — the home page does the job.) Do you really need a separate /about? (Maybe a one-paragraph about block on the home is enough.) Do you really need a /blog? (Only if you will actually publish, otherwise it dates the site.)

4.8 Brand voice

Anti-pattern. Friendly-professional, vaguely millennial, occasionally winking. "We get it — managing X is hard. That's why we built Y. So you can focus on what really matters." First-person plural. No edges. No specific personality. Could be the voice of any of 10,000 SaaS startups.

Why models default to it. This is the median voice of the SaaS web circa 2018-2023, which is exactly the corpus the models trained on. It is friendly enough to not offend, professional enough to seem credible, vague enough to apply to any product. It is also the voice that will get your page indistinguishable from every other startup in your category.

Antidote. Pick a voice that is recognizably yours. First-person singular if you are a solo founder. Technical and dense if your audience is engineers. Sarcastic if your audience is jaded ops people. Formal if you sell to lawyers. The one thing that voice should not be is the one thing the model defaults to.

4.9 Information density

Anti-pattern. Lots of whitespace, large type, short paragraphs, max 60-70 characters per line, generous padding everywhere. A 2,000-word marketing site spread across six pages of mostly blank space. Stripe-style minimalism applied to a context that does not warrant it.

Why models default to it. "Modern SaaS" in the training data means lots of whitespace, because that was the dominant aesthetic from 2017-2023, when the models were ingesting their main visual training material. The model has internalized "professional = sparse" as a rule, and applies it uniformly regardless of whether the content benefits.

Antidote. Match density to content. A reference manual should be dense. A marketing page for a deeply technical product should be dense. A landing page for a piece of software with substantial features should not hide them in icon-and-blurb cards but should explain them in paragraphs. Read Stratechery for an example of dense, hand-built information architecture in long form.

Table 4.A — The 9 dimensions of slop, with severity

| # | Dimension | Anti-pattern signature | Antidote | Severity | |---|-----------|------------------------|----------|----------| | 1 | Palette | Blue-purple gradient | Off-Tailwind palette | High | | 2 | Typography | Inter / Geist everywhere | Serif body or indie sans | High | | 3 | Layout | Symmetric 3-col features | Asymmetric grids, broken rhythm | High | | 4 | Copy | "Transform your X" | Specific, customer-language copy | Critical | | 5 | Iconography | Lucide Sparkles + Zap + Shield | No icons or custom set | Medium | | 6 | Animation | Framer fade-in-from-below | No animation or bespoke motion | Medium | | 7 | Structure | Standard 7-section waterfall | Re-ordered or merged sections | High | | 8 | Voice | Friendly-professional plural | Recognizable singular voice | Critical | | 9 | Density | Sparse everywhere | Density matches content | Medium |

The critical-severity dimensions — copy and voice — are critical because they cannot be salvaged by good visual design. A site with hand-built typography and a custom palette can still be slop if its copy is generator-default. The reverse is rarely true: a site with sharp, specific copy reads as human even if the visual layer is generic.

5. The signature visual stack of AI in 2026

The visual stack of an AI-built site in April 2026 is so consistent that you can describe it as a recipe. This section is the recipe, with a teardown of each ingredient.

The recipe:

• Tailwind CSS as the styling layer (~95% of generator output).
• Tailwind's bg-blue-600 (#3b82f6) as the primary brand color or close variant (blue-500 / blue-700).
• bg-purple-500 (#8b5cf6) or bg-purple-600 as the secondary, almost always paired with the blue.
• A bg-gradient-to-br from-purple-500 to-pink-500 (or from-blue-500 via-purple-500 to-pink-500) as the hero background or the primary CTA accent.
• Inter as the body typeface, Geist or Cal Sans for the hero headline.
• shadcn/ui as the component library, in its default styling (almost no customization of theme tokens).
• Lucide as the icon library, with a recurring vocabulary of Check, Sparkles, Zap, Shield, BarChart3, ArrowRight.
• rounded-2xl (16px border radius) on cards, buttons, and images.
• shadow-md or shadow-lg on cards, sometimes shadow-xl on the primary CTA.
• Framer Motion for entry animations (opacity: 0 → 1, y: 20 → 0, duration: 0.5).

Each ingredient is, in isolation, a fine choice. The problem is that the joint distribution is too predictable. A linter — a thing like Sailop — can fingerprint a site by these ten variables and tell you with high confidence which generator built it.

Table 5.A — Tool vs default visual signature

| Tool | Palette default | Typeface | Components | Icon set | Border radius | Tells | |------|-----------------|----------|------------|----------|---------------|-------| | v0.dev | blue-purple gradient | Geist | shadcn (heavy) | Lucide | rounded-xl/2xl | "Built with v0" footer signature; default Card layout | | Bolt.new | indigo + purple | Inter | shadcn-ish | Lucide | rounded-lg/xl | StackBlitz-style structure; supabase auth boilerplate | | Lovable | purple + pink | Inter | custom + shadcn | Lucide | rounded-2xl | Lovable.dev meta tag; "Built on Lovable" link | | Replit Agent | blue + green | Inter | mixed | Heroicons + Lucide | rounded-lg | Replit-hosted domain pattern | | Claude Code | varies (more taste) | varies | varies | varies | varies | Less signature; output reflects the developer driving it | | Cursor | varies | varies | varies | varies | varies | Same — Cursor is a tool, not a generator; signature depends on the prompts | | GPT-5 / ChatGPT canvas | blue + cyan | Inter | "vanilla React" | inline SVG | rounded-md | Often plain CSS, less Tailwind, more inline styles | | Gemini canvas | green + blue | Roboto / Inter | custom | Material Symbols | rounded-md | Material Design echoes | | Aider | none (CLI) | none | none | none | none | Aider is a CLI that edits code; no visual signature of its own | | Codex | varies | varies | varies | varies | varies | Same as Claude Code — depends on the operator |

Note the split. Some tools (v0, Bolt, Lovable) are "generators" — they output a visual artifact directly. They have strong signatures. Other tools (Claude Code, Cursor, Aider, Codex) are "drivers" — they help a human write code. They have no visual signature of their own; the signature is the developer's. The lesson is not that drivers are clean and generators are dirty. It is that the signature lives where the defaults live, and generators have more defaults.

Why bg-blue-600 won. The path is traceable. Tailwind 1.0 launched in 2019 with a default blue-500 of #3b82f6 and blue-600 of #2563eb. Bootstrap had used a similar blue (#007bff) for years. The "trustworthy SaaS blue" lineage runs back through Facebook (2004), LinkedIn (2003), Twitter (2006), Dropbox (2007). When LLMs trained on the post-2019 web, they encountered this exact hex value tens of millions of times in marketing screenshots, GitHub READMEs, and Tailwind documentation pages. The model learned: "primary color in a SaaS interface" → #3b82f6. RLHF reinforced it because the human reviewers — almost all of whom worked on the post-2019 web themselves — preferred the familiar shade.

Why purple-pink gradients. A more recent lineage. Stripe's checkout page in 2018-2019 used a gradient. Vercel's marketing in 2019-2020 used a gradient. By 2021, the "purple-to-pink" gradient was the visual language of "we are a developer-tool startup". By 2023, every model had internalized it as the default decorative gradient. By 2026, it is so common that using it signals nothing about your product and everything about your tooling.

Why rounded-2xl. Rounded corners have a similar history. Bootstrap's default was 4px in 2013. Material Design pushed it to 8px in 2014. Apple's UI moved to 12-16px on iOS over 2018-2020. shadcn's default Card component uses rounded-xl (12px) and rounded-2xl (16px) in different blocks. Models recommend rounded-2xl because it is the highest-probability radius value in the recent training data.

What this tells search engines. Search engines have access to rendering signals — they can extract dominant colors from screenshots, font families from CSS, and component class names from HTML. They do not officially admit to using these as ranking signals, but the Helpful Content Update has correlated strongly with sites whose visual signatures cluster near the generator centroid. The mechanism is not "Google detected your blue-purple gradient and demoted you". The mechanism is "Google detected that your site is one of forty-seven near-identical sites for the same keyword, all built in March 2026, none with substantial backlinks, and demoted the entire cluster".

The shadcn question. Is shadcn/ui itself slop? No. shadcn is a curated collection of well-built React components with sensible defaults. The slop is in *unmodified* shadcn — using its defaults without changing the radius, the color tokens, the typography, the spacing scale. shadcn is a starting point, not a destination. The teams using shadcn well (Linear's internal tools, Resend's marketing, Cal.com) have extensively customized the theme tokens. The teams using it badly accept the defaults and ship.

Tailwind blue (#3b82f6) adoption on indexed marketing sites
(approximate share of pages with blue-500/600 as primary)

2020 |####                          |  ~10%
2021 |#########                     |  ~22%
2022 |##############                |  ~35%
2023 |####################          |  ~50%
2024 |#########################     |  ~60%
2025 |#############################|   ~72%
2026 |##############################|  ~78%

       0%        25%        50%        75%       100%

The trend is asymptotic but not yet flat. Saturation will likely come in 2027 or 2028, at which point the generator-default centroid will start drifting (slowly, because models change slowly) toward whatever the next dominant aesthetic is.

6. The signature textual stack

If the visual stack is the silhouette of slop, the textual stack is its voice. And the voice has more giveaways per word than the silhouette does per pixel.

Table 6.A — 50 banned phrases (English) with severity

| # | Phrase | Severity | Why | |---|--------|----------|-----| | 1 | In today's fast-paced world | Critical | The classic. Universally generated, never typed by a human. | | 2 | Unleash the power of | Critical | Energy-drink prose for software. | | 3 | Discover seamless | Critical | "Seamless" is in the banned word list; "discover" is the slop verb. | | 4 | Revolutionize your | Critical | Every SaaS landing page since 2015. | | 5 | Leverage cutting-edge | Critical | Stuffed with deprecated business jargon. | | 6 | Empower your team to | Critical | "Empower" is a hollow verb. | | 7 | At its core | High | The pseudo-philosophical opener. | | 8 | When it comes to | High | Filler. Cut it and the sentence is shorter and clearer. | | 9 | In the realm of | High | Fantasy-novel prose for software. | | 10 | A robust solution | High | "Robust" is banned; "solution" is generic. | | 11 | Streamline your workflow | High | Cliché. | | 12 | Boost your productivity | High | Vague metric, vague verb. | | 13 | Take it to the next level | High | Sports cliché. | | 14 | Game-changing | High | If it were really game-changing, the page would say what changed. | | 15 | The ultimate guide | High | Almost never the ultimate guide. | | 16 | A comprehensive overview | High | "Comprehensive" is filler. | | 17 | Best-in-class | Medium | Hollow superlative. | | 18 | World-class | Medium | Same. | | 19 | Industry-leading | Medium | Unprovable claim. | | 20 | Mission-critical | Medium | Almost always overstated. | | 21 | Battle-tested | Medium | Combat metaphor for boring software. | | 22 | In the grand scheme of things | Medium | Filler. | | 23 | At the end of the day | Medium | Filler. | | 24 | Move the needle | Medium | Corporate speak. | | 25 | Low-hanging fruit | Medium | Same. | | 26 | Win-win | Medium | Same. | | 27 | Synergy | Medium | Same. | | 28 | Holistic approach | High | Pseudo-spiritual jargon. | | 29 | Paradigm shift | High | The 1990s want their phrase back. | | 30 | Think outside the box | High | Inside the box, mostly. | | 31 | Circle back | Medium | Meeting filler. | | 32 | Touch base | Medium | Same. | | 33 | Deep dive | Medium | Frequently used in slop FAQ sections. | | 34 | Granular control | Medium | "Granular" is suspicious. | | 35 | Tailored experience | High | Always vague. | | 36 | Bespoke solutions | High | "Bespoke" + "solutions" = double red flag. | | 37 | Crafted with care | High | If it were, you wouldn't need to say so. | | 38 | Built from the ground up | High | Cliché origin story. | | 39 | A new era of | Critical | The favorite era-opener. | | 40 | The future of X is here | Critical | The favorite future-claim. | | 41 | Reimagine | High | The favorite verb of redesigns. | | 42 | Reinvent | High | Same family. | | 43 | Transform | High | Same. | | 44 | At your fingertips | Medium | 2010 prose. | | 45 | One-stop shop | Medium | 1990s prose. | | 46 | Sky's the limit | Medium | Same. | | 47 | The sky is the limit | Medium | Same. | | 48 | A breath of fresh air | Medium | Filler. | | 49 | Truly remarkable | Medium | "Truly" is suspect. | | 50 | Genuinely innovative | Medium | "Genuinely" is the new "truly". |

This list is not exhaustive but it catches a striking share of the textual slop in the wild. Run your copy through it. If you have more than three matches, you have a problem.

The em-dash giveaway. The em-dash is not in the banned phrase list because it is not a phrase, but it is the single most reliable textual signal. Models — particularly GPT-5 and Claude — produce em-dashes at densities that exceed natural human writing in nearly every category. The reasons are partly stylistic: em-dashes let the model produce parenthetical asides without committing to a relative-clause structure. The reason it is a giveaway is statistical: human writers vary their em-dash usage by personality. A novelist may use seven per 1,000 words. A technical writer may use one. A model generates four to six on average regardless of context.

The tricolon. "Faster, smarter, simpler." "Cleaner, leaner, sharper." "Build, ship, scale." Sets of three are a rhetorical pattern as old as Roman oratory. They are not inherently slop. But the rate at which model-generated copy uses tricolons — particularly tricolons of three single-word adjectives or imperative verbs — is a fingerprint. Real human writing uses tricolons sparingly, often with one element broken or expanded for rhythm ("clean, surprisingly fast, and good with edge cases"). Slop uses them in clean threes.

The fake conclusion. Almost every slop blog post ends with a paragraph that begins with "In conclusion," "To wrap up," or "All in all," and restates the introduction in marginally different words. This is a structural fingerprint that survives even when the lexical fingerprint is laundered. Real writers know that a piece either ends because it has finished, or it ends with a turn — a new thought, a question, a callback. A piece that ends by paraphrasing its opening was generated, full stop.

The misuse of "delve". AI-generated text in 2024-2025 used "delve" at a rate that became its own meme. The word is fine. Humans use it. But humans use it in specific contexts: a researcher delves into archives, a critic delves into a director's filmography. Models use it generically, applying it to anything that might be examined in any depth. "Let's delve into the world of customer support." A human would say "Let's look at customer support" or just "Customer support".

The misuse of "tapestry". A specific gift of GPT-3-era models. "A rich tapestry of features." "The intricate tapestry of modern web development." The word means a woven textile and is a perfectly good metaphor when used sparingly for things that are actually woven (cultures, histories, stories). It is a bad metaphor for software features. Models use it because the training corpus contained too many MFA-program essays.

The misuse of "orchestrate". Models love this verb because it appears frequently in cloud-architecture documentation (orchestrating containers, orchestrating workflows). They generalize it to all coordination ("orchestrate your team's communication"). Humans rarely use the word that way. It is a small giveaway but a giveaway.

The textual lookalike test. Same as the visual one. Take three paragraphs from your site. Remove brand names and any specific feature names. Show the result to someone unfamiliar with your category. Ask them to guess the company. If they guess "some kind of SaaS thing", you have textual slop. The cure is specificity that cannot be paraphrased.

7. Tools tear-down

Each tool that builds frontends in 2026 has a default behavior. This section maps the defaults and notes when the tool is a good or bad choice for resisting slop.

v0.dev — Vercel's component generator. Default visual signature: blue-purple gradient, Geist typeface, shadcn components, Lucide icons, rounded-xl/rounded-2xl, Framer Motion on enter. Default code structure: Next.js App Router, server components where possible, Tailwind classes inline. Signature giveaways: "Built with v0" footer link in default deploys, the specific way it imports shadcn components (@/components/ui/*), the way it always offers a hover-lift animation. When OK: as a starting block for components you intend to heavily customize. When NOT OK: as the final form of any commercial site. Treat it as a stencil.

Bolt.new — StackBlitz's full-stack generator. Default signature: indigo and purple, Inter, Vite + React, shadcn-flavored components, supabase auth boilerplate, Lucide icons. Code structure: Vite project, src/ directory, App.tsx as the root, often a page-router pattern. Giveaways: the specific Supabase boilerplate it generates, the way it handles auth state, the structure of its initial commit. When OK: prototyping with persistence. When NOT OK: anything you need to maintain — the generated structure rapidly becomes a mess as you ask for more features.

Lovable — One of the most aggressive generators. Default signature: purple and pink, Inter, custom-flavored shadcn, Lucide. Code structure: React + Vite, often with a heavy AI-generated state-management blob in App.tsx. Giveaways: Lovable's lovable.dev meta tag if not removed, the specific way it structures conversation history in components, the "Built on Lovable" link in default deploys. When OK: the truly "I have an idea this afternoon" use case. When NOT OK: anything where you intend to keep iterating beyond the first day — the codebase has a low ceiling.

Replit Agent — Replit's agent for full-stack apps. Default signature: blue and green, Inter, mixed Tailwind/CSS, Heroicons or Lucide. Code structure: depends on the language (Python + Flask, Node + Express, Next.js — the agent picks). Giveaways: Replit-hosted domain pattern (*.replit.dev), the specific way it handles environment variables. When OK: Replit-native users iterating on small projects. When NOT OK: production deployments outside Replit's hosting.

Claude Code — Anthropic's CLI for code in the terminal. No default visual signature of its own. Output reflects the developer driving it. With a careful prompt, Claude Code produces output that is indistinguishable from hand-built. With a lazy prompt ("build me a SaaS landing page") it produces output that resembles the median of its training data — i.e., slop. Giveaways: none structural; the giveaway is the prompt the developer used. When OK: as a daily driver for any developer who knows what they want. When NOT OK: never — it is a tool, not a generator. The slop is in how it is used.

Cursor — IDE with AI completion. Same logic as Claude Code: no visual signature of its own, only the signature of the developer. Cursor's specific superpower is in-context multi-file edits and tab completion that makes refactoring fast. The slop risk is in the autocomplete: if you accept every Tab without thinking, your code drifts toward the median. When OK: heavy daily use by developers who refuse autocomplete suggestions. When NOT OK: nothing inherent — the tool is fine; the discipline is the variable.

GPT-5 / ChatGPT canvas — OpenAI's canvas mode. Default signature: blue and cyan, Inter, often plain CSS rather than Tailwind, more inline styles than other generators, simpler component structures. Code structure: vanilla React, sometimes Next.js. Giveaways: the specific way it scaffolds projects, its preference for inline styles where the other tools use utility classes, occasional drift toward Material-Design-like rounded shapes. When OK: rapid prototyping where you want plain CSS. When NOT OK: production sites — the architecture of GPT canvas output is rarely scalable.

Gemini canvas — Google's canvas mode. Default signature: green and blue (Material echoes), Roboto or Inter, custom components rather than shadcn, Material Symbols icons, rounded-md. Code structure: varies, often heavier on Material Design conventions. Giveaways: Material Symbols rather than Lucide, color choices closer to Material's palette than Tailwind's. When OK: integrating into Google ecosystems. When NOT OK: as a general-purpose web generator outside Google's design language.

Aider — CLI-based AI pair programmer. No visual signature. Aider is a tool for editing existing code rather than generating new visual artifacts. It is a fine choice for the developer who wants minimal AI footprint in the architecture and uses the model only for specific edits. When OK: any disciplined developer. When NOT OK: anyone hoping the tool will generate the visual layer for them — Aider does not.

Codex — OpenAI's coding agent. Same logic as Claude Code or Aider: a driver, not a generator. Visual signature depends on the operator. Tech-stack defaults reflect OpenAI training data. When OK: any disciplined developer. When NOT OK: anyone using it as a one-shot generator for marketing pages.

The pattern: generators have signatures. Drivers do not. If you want to avoid slop, you can either (a) avoid generators or (b) use them as starting points and pay the customization tax. There is no third option where you accept generator defaults and produce non-slop work.

8. 47 real-world examples torn apart

This section presents 47 anonymized but realistic case studies from teardowns conducted between July 2025 and March 2026. Names have been changed; the patterns are real.

Table 8.A — The 47 case studies (compact form)

| # | Type | Slop dimensions hit | Severity (0-100) | Fix priority | |---|------|---------------------|------------------|--------------| | 01 | Indie SaaS landing | Palette, Type, Copy, Voice | 78 | High | | 02 | Agency portfolio | Layout, Copy, Voice, Density | 72 | Medium | | 03 | Newsletter signup page | Palette, Type, Copy | 65 | High | | 04 | E-commerce home (DTC) | Palette, Layout, Iconography | 70 | High | | 05 | API documentation | Type, Density, Copy | 58 | Medium | | 06 | Personal portfolio | Palette, Type, Layout, Voice | 81 | High | | 07 | SaaS pricing page | Layout, Copy, Structure | 74 | Critical | | 08 | Mobile app landing | Palette, Iconography, Animation | 68 | High | | 09 | Conference site | Layout, Type, Copy | 62 | Medium | | 10 | Open-source project | Type, Iconography, Copy | 55 | Medium | | 11 | Course landing | Palette, Copy, Voice, Animation | 79 | High | | 12 | Crypto project | Palette, Type, Copy (extreme) | 88 | Critical | | 13 | Local-business site | Palette, Layout, Iconography | 71 | High | | 14 | Job board | Layout, Copy, Density | 64 | Medium | | 15 | Productivity app | Palette, Type, Layout, Animation | 77 | High | | 16 | AI startup (meta) | All 9 dimensions | 92 | Critical | | 17 | B2B SaaS | Layout, Copy, Voice | 67 | High | | 18 | Indie game site | Type, Layout, Copy | 60 | Medium | | 19 | Podcast site | Palette, Type, Copy | 65 | Medium | | 20 | Restaurant site | Palette, Layout, Iconography | 72 | High | | 21 | Health app | Palette, Type, Copy, Animation | 76 | High | | 22 | DAO landing | Palette, Type, Copy | 80 | Critical | | 23 | Education non-profit | Layout, Copy, Voice | 58 | Medium | | 24 | Blog (single-author) | Type, Density, Voice | 52 | Low | | 25 | E-comm marketplace | Palette, Layout, Copy | 70 | High | | 26 | Fitness app | Palette, Iconography, Copy | 73 | High | | 27 | Travel platform | Palette, Layout, Copy | 69 | High | | 28 | Photography portfolio | Palette, Type, Layout | 66 | Medium | | 29 | Architecture studio | Layout, Type, Copy | 60 | Medium | | 30 | Music streaming clone | Palette, Layout, Iconography | 75 | High | | 31 | Recipe site | Palette, Type, Copy | 63 | Medium | | 32 | Real-estate platform | Layout, Copy, Density | 70 | High | | 33 | Marketplace landing | Palette, Layout, Voice | 72 | High | | 34 | Newsletter platform | Palette, Type, Copy | 67 | Medium | | 35 | Design tool | Palette, Type, Animation | 71 | High | | 36 | CRM SaaS | Layout, Copy, Voice | 68 | High | | 37 | Email tool | Palette, Iconography, Copy | 73 | High | | 38 | Calendar app | Palette, Layout, Animation | 70 | High | | 39 | Note-taking app | Palette, Type, Density | 65 | Medium | | 40 | Voice memo app | Palette, Iconography, Copy | 69 | High | | 41 | AI chat wrapper | All 9 dimensions | 91 | Critical | | 42 | AI code wrapper | All 9 dimensions | 89 | Critical | | 43 | AI image wrapper | All 9 dimensions | 87 | Critical | | 44 | AI agent platform | Palette, Type, Copy, Voice | 84 | Critical | | 45 | Agency landing (boutique) | Type, Copy, Voice | 56 | Low | | 46 | Personal blog | Type, Density, Voice | 48 | Low | | 47 | Niche tool | Palette, Type, Copy | 64 | Medium |

The median severity is 70. The mean is approximately 71. Twelve of the 47 cases score above 80, which is the threshold above which a site is essentially indistinguishable from generator output. Three cases score above 90 — what we call deep slop — where every dimension hits the anti-pattern.

Ten cases expanded.

Case 16 — AI startup (meta). Severity 92. The site of an AI productivity tool, ironically built almost entirely by AI. The hero used a bg-gradient-to-br from-purple-500 to-pink-500, the headline was "Unleash the power of AI to revolutionize your workflow", the subhead included "seamlessly", the features section was three cards in a grid-cols-3 with Sparkles icons, the testimonial avatars were stock-photo Asian and Black faces with name pairs that sounded generated ("Sarah Chen", "Marcus Williams"), the pricing tiers were "Starter / Pro / Business", and the FAQ paraphrased the headline twice. Lookalike test: indistinguishable from approximately 200 other AI startups indexed in the same month. Fix: total rewrite of copy, custom illustration replacing the gradient, real customer photos, repositioning of pricing to emphasize value over tier comparison.

Case 12 — Crypto project. Severity 88. The site for a token-launching protocol. Used a black-purple-pink gradient, Geist Mono headlines, "Revolutionize DeFi" copy, Lucide icons (Zap, Shield, BarChart3 in literal sequence), a tokenomics section with three pie charts, and a roadmap with quarter-stamped milestones that all said "Q4 2026 — Multi-chain expansion". The slop is not the crypto-ness; it is that this exact site exists at scale across the crypto category. Fix: a custom illustration for the protocol's actual mechanism, copy that explains what the token does instead of promising what it will revolutionize, removal of the roadmap (which dates the site).

Case 06 — Personal portfolio. Severity 81. A junior developer's portfolio. Generated using v0 with light customization. Hero said "Hi, I'm [name] — I build beautiful, modern web experiences." Three project cards with screenshots, each with a Lucide ExternalLink icon, each card rounded-2xl shadow-md. Skills section: a grid of icons (Tailwind, React, TypeScript, Next.js, shadcn, Vercel — exactly the stack the model would suggest). Footer: "Built with care, crafted in [city]". Fix: replace generated copy with a one-paragraph statement of what the developer actually likes about the work, remove the skills grid (anyone hiring can see the projects), pick a typeface and a palette that signals personality.

Case 22 — DAO landing. Severity 80. The site for a decentralized governance project. Combined the worst of crypto slop (token-economics promises) with the worst of SaaS slop (gradient hero, Lucide icons). The kicker: the site claimed to "empower the community" while offering no mechanism by which a community member could actually do anything. Fix: scrap the DAO framing; describe what the protocol *does*; let governance be a sub-page if it matters.

Case 41 — AI chat wrapper. Severity 91. A wrapper around the GPT-5 API marketed as a "ChatGPT alternative for [niche]". The site cloned ChatGPT's visual language (the chat bubble, the system message styling), used the standard blue-purple gradient, claimed to "transform your customer support workflow", offered three pricing tiers. The actual product was identical to a hundred other niche chat wrappers. Fix: pivot or close. The slop signal here is also a market-fit signal — the site looks like everything else because the product is.

Case 07 — SaaS pricing page. Severity 74. Critical fix priority. A B2B SaaS pricing page in the standard three-column form. The middle column was tagged "Most Popular", the prices were $19/$49/$99 monthly, the features were copied from a competitor's page (we verified). The fix here is structural: rewrite the pricing in plain English ("Solo developers: $19. Teams of 2-10: $49. Companies above 10: contact us"), remove the comparison table, write the FAQ as actual customer questions from the founder's inbox.

Case 13 — Local business site. Severity 71. A regional plumbing business that had been talked into a "modern website redesign" by an agency that built v0 sites. The new site looked like every SaaS landing page, with a blue-purple hero, Lucide icons for "Reliable / Fast / Affordable", and a "Book Now" CTA. The previous site had been ugly and effective. The new site was beautiful and confused customers. Fix: revert. The local-business site does not need to look like SaaS.

Case 16 (expanded continuation). A note on the meta-case: the AI startup that scored 92 also had the textual signature of model-generated copy. We tested by running its hero, subhead, and feature blurbs through three different LLM-output detectors, all of which flagged the copy at >90% confidence. We do not put weight on these detectors as standalone tools (they are noisy), but the convergence with our visual analysis was striking. The site read as machine-built at every level.

Case 30 — Music streaming clone. Severity 75. A "Spotify alternative for podcasts" site. Used Spotify's color palette (a green and black) but everything else was generator-default: rounded-2xl cards, Lucide play-button icons, three pricing tiers, "Discover seamless audio" copy. The clash between the borrowed palette and the generator-default everything else made the site read as both derivative and generic at once. Fix: pick a unique palette, redesign the player component as the centerpiece, drop the marketing copy entirely and let the product do the work.

Case 16 / 41 / 42 / 43 — the AI-wrapper cluster. Four of the highest-severity cases were all "AI wrappers" — products that wrap an existing LLM API in a niche-specific UI. The pattern was consistent: the wrapper UI cloned ChatGPT or Claude visually, the marketing site cloned generator-default SaaS, and the product itself was easily replicated. The slop in the marketing layer was a faithful reflection of the slop in the product layer. The fix here is not a marketing fix.

The lesson from 47 cases. Slop is not random. It clusters by category. Crypto slop is recognizable. AI-wrapper slop is recognizable. Indie-SaaS slop is recognizable. The clustering means that the bar to stand out within a category is not "be perfect" but "be one degree of freedom different from your immediate neighbors". A crypto site with a non-purple palette already breaks the pattern. A SaaS site with serif body type already breaks the pattern. The cliff is shallow. Climbing it is the cheapest competitive advantage available in 2026.

9. The AI slop economy

Slop is not a free phenomenon. It costs someone something, and someone profits. This section maps the flows of money around the slop economy in 2026.

Who profits.

Template marketplaces. Sites like ThemeForest, Creative Market, and the new generation of Tailwind-template marketplaces (Tailwind UI, Tailblocks, Cruip, Mamba UI) have benefitted from AI-driven web growth in two ways. First, the explosion in indie sites has expanded the customer base. Second, the rise of AI generators has created a tier of users who use templates as a *cleanup* layer over generator output. The market for clean Tailwind templates is healthy and growing.

Anti-slop premium. A new market category. Boutique studios — Linear's old design lead's new shop, Vercel ex-designers' studios, type-led agencies in Berlin and Tokyo — charge a meaningful premium over what mid-range agencies charge for "an AI-quality website". The premium is for taste, not for code. The clients are companies that have been burned by generator-default sites and want something humans built.

Freelance bifurcation. The freelance market has split. The bottom half — generic full-stack devs charging $30-60/hour to "build a website" — has been hollowed out by generators. The top half — specialists with a recognizable portfolio, particularly those who can demonstrate work that does not look generator-built — charges more than they did in 2022. The middle is gone.

Audit-as-a-service. The category Sailop sits in. Companies whose sites look generator-built want a second opinion. Audit shops charge $1,500-$5,000 for a written report on a single site, identifying slop dimensions and proposing fixes. This category did not exist in 2023.

Search engines fighting slop. Google, Bing, and Perplexity all have internal teams working on de-ranking generator-default content. The fight has revenue implications: ad networks suffer when the content is low-quality, search trust erodes when results are slop. Whether they fight it well is a separate question (section 10), but the fight itself has a revenue logic.

Generator vendors themselves. v0, Bolt, Lovable, Replit, Cursor — all are well-capitalized businesses whose revenue grows when more people generate more sites. Slop is, in a meaningful sense, their exhaust. They are not directly incentivized to reduce it, though several of them now market features ("brand customization", "design tokens") that nominally help users avoid it.

Who loses.

Junior developers. The harshest economic impact. The traditional path of "build five portfolio sites, get hired" is broken because anyone can generate five portfolio sites in an afternoon and they all look identical. Juniors with strong taste, who can demonstrate work that obviously was not generated, still get hired. Juniors without that signal are invisible.

Mid-tier agencies. The shops that charged $5,000-$25,000 to build small business sites are squeezed from below (generators do it for free) and from above (boutique studios charge more for taste). Many have pivoted to "AI consulting" or wound down. The 2018-2022 generation of mid-tier agencies has been the biggest casualty.

Original content creators. Bloggers, indie writers, and small publications have lost search-traffic share to slop content farms. Even with Helpful Content updates, the volume advantage of farms has been hard to overcome. Substack, Beehiiv, and other newsletter-direct platforms have grown in part as a hedge against this — readers know there is a human at the other end.

Where the money flows. The net effect: money has flowed *up* (to generator vendors, to boutique studios, to template marketplaces) and *down* (to bulk SEO operators running content farms). The middle — mid-tier agencies, generic freelancers, mid-tier publications — has been compressed. This is consistent with broader patterns in AI-affected industries (writing, illustration, customer support): high-end and ultra-low-end thrive, the middle hollows.

A quiet sub-economy: the unbuild market. A new niche of professionals — call them "web strippers" — specialize in taking generator-built sites and removing the slop signatures. Replace Inter with a serif. Strip the gradient. Rewrite the copy. Their pitch is: keep the v0 backend, get a site that does not look like v0. Pricing in this niche is $1,000-$3,000 for a typical small site, takes 1-2 days, and is growing fast.

Internal links. For more on this see The AI slop economy 2026, which goes deeper into the financial flows we sketch here.

10. Search engines' counter-attack

Search engines have been the main institutional force pushing back on slop. This section traces the timeline.

Table 10.A — Search engine counter-slop timeline 2022-2026

| Year | Event | Significance | |------|-------|--------------| | 2022 (Aug) | Google "Helpful Content Update" launched | First explicit anti-low-effort signal | | 2023 (Sep) | Major Helpful Content refresh | Demoted thin AI-generated content broadly | | 2024 (Mar) | Core update + Spam policies update | Tightened "scaled content abuse" | | 2024 (May) | AI Overviews launched (US) | Changed user intent funnels — fewer clicks | | 2024 (Aug) | Google explicitly states AI-generated content can rank if helpful | The official line: it's about quality, not origin | | 2025 (Q1) | Major demotion of bulk-generated affiliate sites | Reported in WSJ, Verge, Search Engine Roundtable | | 2025 (Q2) | Bing Copilot integration deepens | Synthesized answers compete with traditional results | | 2025 (Q3) | Perplexity's source-attribution model gains share | New ranking signal: per-source citation in synthesis | | 2025 (Q4) | Google EEAT criteria refined for AI-era | "Experience" particularly elevated | | 2026 (Q1) | Generator-default visual fingerprints reportedly correlated with demotion | Not officially confirmed; inferred from teardowns | | 2026 (Q2) | Site-level signals dominate page-level signals | Clustering of identical sites gets the cluster demoted |

The Helpful Content Update philosophy. Google's official position is consistent: it does not penalize AI-generated content per se. It penalizes low-quality content that fails to demonstrate experience, expertise, authoritativeness, and trustworthiness (EEAT). In practice, generator-default content fails the experience axis hard. A page about "the best CRMs in 2026" written by GPT-5 in 30 seconds has no experience behind it. The page's structure, rhythm, and lack of specifics betray that. Google's signal extraction is sophisticated enough to detect this even when the surface text passes traditional spam filters.

EEAT in the AI era. The four EEAT criteria have shifted weight:

• Experience — newly elevated. Google now looks for evidence that the author has actually used the thing they're writing about. First-person observations, specific edge cases, photos taken in real environments, dates that make sense. Generator-default content fails this.
• Expertise — somewhat depreciated. Models can fake expertise in surface language, so this signal is noisier. Google has shifted weight away from it.
• Authoritativeness — backlink-driven, slow to move. Old domains with real link profiles still benefit; new domains struggle regardless of content quality.
• Trustworthiness — penalty-based. If your site has any signals of being part of a scaled content farm — shared hosting with other slop sites, identical structure to other sites in the same cluster, missing or generated About pages — you take a trust penalty.

AI Overviews changing intent. The biggest commercial impact on slop sites in 2025-2026 has been the rise of AI Overviews and Bing Copilot synthesized answers. These now sit at the top of many search results pages and answer the user's question without sending them to a website. Slop sites — whose business model depended on cheap traffic to ad-heavy pages — have been hit hardest because they offered the lowest marginal value per click. Quality sites still get clicks for deep-engagement queries. Slop sites lose the entire "answer my one-line question" market.

Perplexity's role. Perplexity is interesting because it cites sources directly in its synthesized answers. Sites that get cited get a small but meaningful traffic signal and, more importantly, a brand-recognition signal. The sites that get cited by Perplexity tend to be the same sites that rank well in Google: real publications, technical blogs with actual experience, primary sources. Slop content farms rarely get cited because Perplexity's source-evaluation model down-weights them.

Why slop ranks short-term. Generator-built sites can rank in the first few weeks after publication, particularly for long-tail keywords with low competition. The mechanism is simple: Google indexes them, runs initial quality signals (which take time to accumulate), and ranks them based on baseline relevance. During this window — typically 2 to 8 weeks — slop sites can capture meaningful traffic.

Why they get demoted long-term. After the initial window, three signals catch up. First, behavioral signals: bounce rate, time on page, scroll depth, return visits. Slop content tends to score badly on all of these because users recognize it as low-effort and leave. Second, link signals: nobody links to slop content because there's nothing in it worth pointing at. Third, cluster signals: Google's site-level analyses identify when many similar sites publish similar content from similar templates, and the entire cluster gets demoted. The result is the rise-and-fall pattern that defines the slop content business: 6-week traffic, then a long decline.

The implication for legitimate sites. If your site looks like a slop cluster — even if your content is excellent — you risk being caught in the demotion. The fix is to differentiate. Custom design, distinctive voice, real backlinks, genuine engagement. Section 12 has the practical playbook.

11. The decision tree: is your site slop?

This section is the diagnostic. If you read nothing else, read this.

The flowchart.

flowchart TD
  A[Open your site in a private window] --> B{Is the primary palette<br/>blue + purple gradient?}
  B -- Yes --> C[+25 slop points]
  B -- No --> D[+0]
  C --> E
  D --> E
  E{Is the body typeface<br/>Inter, Geist, or<br/>another default sans?}
  E -- Yes --> F[+15 slop points]
  E -- No --> G[+0]
  F --> H
  G --> H
  H{Are there 3 cards<br/>in a grid-cols-3<br/>with Lucide icons?}
  H -- Yes --> I[+15 slop points]
  H -- No --> J[+0]
  I --> K
  J --> K
  K{Does the headline<br/>contain 'transform',<br/>'unleash', 'revolutionize'?}
  K -- Yes --> L[+20 slop points]
  K -- No --> M[+0]
  L --> N
  M --> N
  N{Is the FAQ<br/>paraphrasing the<br/>rest of the page?}
  N -- Yes --> O[+10 slop points]
  N -- No --> P[+0]
  O --> Q
  P --> Q
  Q{Does the conclusion<br/>start with 'In conclusion'<br/>or restate the intro?}
  Q -- Yes --> R[+15 slop points]
  Q -- No --> S[+0]
  R --> T[Sum the points]
  S --> T
  T --> U{Score &gt; 60?}
  U -- Yes --> V[Your site is slop.<br/>Go to section 12.]
  U -- No --> W{Score 30-60?}
  W -- Yes --> X[Your site is at risk.<br/>Run the 17-question check.]
  W -- No --> Y[You're probably fine.<br/>Run the lookalike test anyway.]

The 17-question checklist.

For each question, score yourself on the severity scale: 0 (not at all), 1 (a little), 2 (clearly), 3 (deeply).

• Does your hero use a blue-to-purple or purple-to-pink gradient? (Severity weight: 4)
• Is your body typeface Inter, Geist, or a similarly default sans? (Severity weight: 3)
• Does your hero headline use one of the banned phrases (section 6)? (Severity weight: 5)
• Do you have a lg:grid-cols-3 features section with three cards? (Severity weight: 3)
• Are the icons in those cards from Lucide and named Sparkles, Zap, Shield, or Check? (Severity weight: 3)
• Are your card border-radii rounded-2xl (16px)? (Severity weight: 2)
• Do your cards have shadow-md or shadow-lg? (Severity weight: 2)
• Does your testimonial section have stock-photo avatars or AI-generated faces? (Severity weight: 4)
• Does your pricing page have three columns with the middle one tagged "Most Popular"? (Severity weight: 3)
• Are your pricing tiers named "Starter / Pro / Business" or similar generic? (Severity weight: 2)
• Does your FAQ paraphrase the rest of the page? (Severity weight: 4)
• Does any section conclude with "In conclusion" or "All in all"? (Severity weight: 3)
• Is your footer credit "Built with care" or "Crafted with love" or similar? (Severity weight: 2)
• Are there Framer Motion fade-in-from-below animations on entry? (Severity weight: 2)
• Does your hero have a secondary CTA "Watch demo" with a play icon? (Severity weight: 2)
• Is your About page a single paragraph in third person about "we"? (Severity weight: 3)
• Does your site pass the lookalike test against five competitors? (Severity weight: 5; INVERTED — 0 if it passes, 3 if it fails)

Total maximum score: 154. Multiply your score by (100/154) to normalize to a 0-100 scale.

Table 11.A — Score interpretation (0-100)

| Score range | Verdict | Action | |-------------|---------|--------| | 0-15 | Clean | Verify with the lookalike test, then ship. | | 16-30 | Mostly fine | Identify the 1-2 dimensions that scored highest, fix those. | | 31-50 | Borderline | Material risk. Plan a redesign of 2-3 dimensions over 1-2 weeks. | | 51-70 | Slop | You will be demoted in search and your conversion will suffer. Redesign now. | | 71-90 | Deep slop | Consider whether the site is salvageable or whether starting over is faster. | | 91-100 | Total slop | The site is a clone. Start over with a different brief. |

Internal link. For more on rapid detection, see Detect AI-generated site in 30 seconds — 21 signs, which is the field guide that pairs with this report's diagnostic.

12. Antidotes (the practical part)

This is the section to bookmark. Twenty-three concrete fixes, each ranked by priority, each with a code before/after where applicable. The priority ranking reflects bang-for-buck: high-impact, low-effort changes first.

Priority 1 — Replace the gradient.

// Before
<section className="bg-gradient-to-br from-blue-600 via-purple-500 to-pink-500 text-white py-24">
  <h1 className="text-5xl font-bold">Transform your workflow</h1>
</section>

// After
<section className="bg-[#fef9f4] text-[#1a1a1a] py-24 border-b-2 border-[#1a1a1a]">
  <h1 className="font-serif text-5xl font-medium italic">
    Replies to your support tickets in your voice. Not in 'a friendly, professional tone.'
  </h1>
</section>

Priority 2 — Replace the typeface. Move from Inter to a serif body or a less-worn sans.

/* Before */
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;700&display=swap');
:root { --font-body: 'Inter', sans-serif; --font-display: 'Inter', sans-serif; }

/* After */
@import url('https://fonts.googleapis.com/css2?family=Newsreader:opsz,[email protected],400;6..72,500&family=Fraunces:opsz,[email protected],500&display=swap');
:root { --font-body: 'Newsreader', Georgia, serif; --font-display: 'Fraunces', 'Times New Roman', serif; }

Priority 3 — Rewrite the hero copy with specifics.

<!-- Before -->
<h1>Transform your workflow with AI-powered automation.</h1>
<p>The all-in-one platform that helps teams ship faster, smarter, and with less stress.</p>

<!-- After -->
<h1>Replies your support emails in 30 seconds, in your team's voice.</h1>
<p>Reads your last 200 tickets, mirrors the tone, hands you a draft. You hit send. We tested with eight teams; the median time saved was 4 hours per week per agent.</p>

Priority 4 — Strip the icon decoration.

// Before
<div className="flex gap-3 items-center">
  <Sparkles className="w-6 h-6 text-purple-500" />
  <h3>AI-powered insights</h3>
</div>

// After
<div>
  <h3 className="font-serif text-2xl">Your support inbox, summarized hourly.</h3>
</div>

Priority 5 — Break the symmetric grid.

// Before
<div className="grid lg:grid-cols-3 gap-8">
  <Card />
  <Card />
  <Card />
</div>

// After — asymmetric, intentional
<div className="grid lg:grid-cols-12 gap-8">
  <div className="lg:col-span-7"><MajorFeature /></div>
  <div className="lg:col-span-5"><MinorFeature /></div>
  <div className="lg:col-span-12 border-t pt-8"><LongFormExample /></div>
</div>

Priority 6 — Replace stock-photo testimonials with text-only quotes from real users.

// Before
<div className="flex items-center gap-4">
  <img src="/avatars/sarah-chen.jpg" className="w-12 h-12 rounded-full" />
  <div><p>"Game-changing tool!"</p><span>Sarah Chen, CEO @ TechCo</span></div>
</div>

// After
<blockquote className="border-l-4 pl-4">
  <p>"It cut our reply time in half. The first week was rough — the tone matching needed work — and after we fed it our last 50 escalations, the drafts started landing."</p>
  <cite>— Maria, head of support, [real-but-anonymized startup]</cite>
</blockquote>

Priority 7 — Rewrite the pricing in plain English.

<!-- Before -->
<div class="grid grid-cols-3">
  <div>Starter $19/mo — 100 messages</div>
  <div class="ring-2">Pro $49/mo — 1000 messages [Most Popular]</div>
  <div>Business $199/mo — Unlimited</div>
</div>

<!-- After -->
<section>
  <p>Solo founders and side projects: $19/month. Up to 500 replies. No CC required for the first 14 days.</p>
  <p>Teams of 2-10: $49/month plus $9/seat. Includes the audit log and shared snippets.</p>
  <p>If you need SSO, a DPA, or want us to spin up a private model — talk to me directly: <a href="mailto:...">...</a>.</p>
</section>

Priority 8 — Drop the FAQ that paraphrases the page. If your FAQ does not contain genuinely different information from the rest of the page, delete it. If it does, keep it but rewrite each Q in the actual words a customer used.

Priority 9 — Replace the "In conclusion" closer. If you have a paragraph that starts with "In conclusion", "All in all", "To wrap up", or "In summary", delete it. End the page where the substance ends.

Priority 10 — Rewrite the About in first person singular.

<!-- Before -->
<p>We are a team of passionate developers building the future of customer support automation.</p>

<!-- After -->
<p>I'm Robin. I built this because I spent two years running support at a 20-person startup and watched my team burn out on tickets that could be 70% drafted by a machine if the machine had read our last 200 tickets first. The machines I tried hadn't. So I built one that does.</p>

Priority 11 — Replace shadow-md with no shadow or hand-tuned shadow.

/* Before */
.card { box-shadow: 0 4px 6px -1px rgb(0 0 0 / 0.1), 0 2px 4px -2px rgb(0 0 0 / 0.1); }

/* After — flat with a borders */
.card { border: 1px solid #e5e5e5; }

/* OR — bespoke shadow with a tint */
.card { box-shadow: 4px 4px 0 0 #1a1a1a; border: 1px solid #1a1a1a; }

Priority 12 — Delete the entry animations. If you have a Framer Motion fade-in-from-below on every element, delete it. The page will feel faster and less generic.

Priority 13 — Custom OG image. The default OG image generated by most frameworks is the page screenshot or a templated card. Replace it with a custom 1200x630 image that has actual visual personality. This is what people see in social shares; it deserves attention.

Priority 14 — Custom 404 page. Almost every generator-built site has the default 404. Make yours interesting. A photo of a real thing in your physical environment. A line of code that runs. A handwritten note. Anything that signals a human is involved.

Priority 15 — Custom favicon. Default favicons are dead-giveaway slop. Make a real favicon. Hand-draw it if necessary.

Priority 16 — Footer that is actually useful. Most slop footers are four-column link grids ending in "Built with care". Replace with a single paragraph that says what the company is, links to the three pages that matter, and ends with the founder's email.

Priority 17 — Replace the testimonial carousel with one quote. Carousels are slop. They are also unread. Pick the single best customer quote and feature it.

Priority 18 — Replace the logo wall with one specific reference. "Used by teams at Linear, Figma, Stripe" is a logo wall. "Used by Maria's support team at [company]" is a reference. The second is more credible.

Priority 19 — Replace the integration grid with two real examples. A grid of 30 integration logos signals breadth nobody cares about. Two examples of how the product connects to two specific tools that the customer actually uses signal usefulness.

Priority 20 — Remove the "Trusted by 10,000+ teams" claim unless it is verifiably true and you can name the teams. Otherwise the claim adds slop signal without earning trust.

Priority 21 — Add a screenshot of the actual product, taken from the actual product, with the actual data the actual customer would see. Most slop sites have either no product screenshots or fake mockups. A real screenshot — even a slightly imperfect one — is a strong human signal.

Priority 22 — Add a footer link to a "Made by" page that lists the actual humans who built the thing. Faces, names, small bios, contact emails. This is unfashionable in 2026 and that is exactly why it works.

Priority 23 — Run the lookalike test before you ship. Reduce your homepage to a 200px black-on-white silhouette. Lay it next to five competitors. If you can't tell yours apart, go back to Priority 1 and start over.

Internal link. For a more focused conversion of a generator-built site, see De-AI your Lovable / v0 / Bolt site — the targeted playbook for stripping signatures from a generator output.

13. The Sailop angle

A short note on the tool that funded this report.

Sailop is a design linter that reads a website and reports its slop signatures. It is an npm package with a CLI and an MCP server. The CLI runs in CI and the MCP server lets Claude Code query the linter while it codes. The point of the tool is not to police creativity — design is too local and contextual for an automated tool to grade in absolute terms — but to flag the *defaults* that drift toward the generator centroid and let the developer make an informed choice.

The heuristics behind Sailop's scoring map almost directly onto the nine dimensions in section 4. The palette test extracts the dominant colors and checks distance from the Tailwind blue-purple-pink centroid. The typography test reads font-family declarations and flags Inter, Geist, Cal Sans, and the small set of similar defaults. The layout test counts grid columns, card patterns, and section ordering. The copy test runs the banned-phrase list and the em-dash density check. The iconography test flags the recurring Lucide icons in default configurations. The animation test detects Framer Motion entry patterns. The structure test compares the page's information architecture to a database of known generator outputs. The voice test runs an LLM evaluator over the copy and reports a humanness score. The density test computes whitespace ratios.

The output is a score on 100 (matching section 11's scale), a per-dimension breakdown, and a list of specific lines or selectors that triggered each flag. From there, the developer decides what to keep and what to change.

The MCP server exposes the same heuristics to Claude Code. The pattern is: the developer asks Claude Code to build a feature, Claude Code builds it, then Claude Code (via the MCP server) asks Sailop to check the output, then Claude Code revises if the score is high. This is the loop the tool is built for. It makes Claude Code a better collaborator, not a worse one — because the lint runs *after* generation, not as a constraint *during* generation.

The CI integration is the same idea at the team level. A pull request opens, Sailop runs against the staging deploy, the score is posted as a check. Above a threshold, the check fails. Teams set the threshold based on their tolerance.

That is the philosophy. It is not "AI bad". It is "defaults expensive". The price you pay for accepting defaults is being indistinguishable, and being indistinguishable is being demoted, ignored, or replaced. A tool that surfaces the cost lets you choose.

14. The future

Predictions for 2027-2030. Caveat: predicting AI is a sport for fools. These are educated guesses based on current trajectories.

Video slop is the next big surface. Sora and Veo are generating video at quality levels that would have been state-of-the-art research demos two years ago. The output has a signature: the warm cinematic LUT, the slight physics drift in cloth and water, the character whose hands fold strangely on close inspection, the camera move that is too smooth to be handheld and too jittery to be a Steadicam. By 2028, an enormous share of YouTube tutorials, TikTok b-roll, and corporate video content will be generated. The slop signatures will be similar to today's image slop: too consistent, too saturated, too signal-clear-from-the-tool.

Audio slop accelerates. Suno and Udio in 2026 produce four-minute songs with vocals that pass casual listening tests. By 2028, podcast intros, hold music, in-app sound design, and a meaningful fraction of background music in commercial video will be generated. The audio slop signature is harder to articulate than visual slop — it lives in mix-bus compression patterns, reverb tail uniformity, harmonic-content distributions — but trained ears will detect it. The first wave of "anti-slop audio" services is already emerging: human composers with a portfolio that does not match the AI centroid charge a premium.

3D slop. Generative 3D is still rough in 2026 but accelerating. Game studios are already using it for first-pass asset generation. By 2029, indie game environments will be slop-saturated in the same way indie SaaS sites are now, with the same dynamics: generator-default looks fine in isolation, identifiable in cluster.

Full-app slop. The generators in 2026 produce websites; the generators in 2028 will produce full apps. We are seeing early signs in Replit Agent and v0's full-app modes. By 2030, "I have an app idea" will mean "I had an agent generate an app". The slop will be in the data models, the UX flows, the empty-states, the onboarding sequences — everything the generator has internalized as the median pattern.

Agent-built apps. The next layer above generator-built apps: agent-built SaaS, where an agent runs the entire business — generates the marketing, handles the support, ships the product updates, processes the payments. A small number of these already exist as experiments in 2026. By 2028, they will be a meaningful share of the indie SaaS landscape. Their slop signature will be the most pronounced of any category, because every layer was machine-generated.

The anti-slop premium market grows. As slop saturates, the premium for non-slop work grows. By 2028, the boutique studio market that now charges $150k-$500k for a website will be saturated with demand. New shops will emerge specifically to do anti-slop work for businesses that can afford it. The price gap between "AI-quality" and "human-quality" web work will widen.

Regulation arrives. The EU AI Act, in its content-provenance provisions, will require more explicit labeling of AI-generated content over 2027-2028. The US is likely to follow at a more fragmented level (state-by-state). The labeling will be partial — there is no reliable forensic test for AI-generated content in absolute terms — but it will create a legal hook for downranking, demonetizing, and de-prioritizing material that does not disclose. Slop sites will face new compliance costs.

Search engines fragment. Google's dominance erodes incrementally but doesn't collapse. Perplexity, Bing Copilot, and category-specific AI search (Phind for code, You.com for general, Kagi for the privacy-and-quality-conscious) carve out meaningful niches. The fragmentation makes it harder for slop content farms to optimize against a single search engine — they have to optimize against multiple synthesis engines, each with different ranking models. This is actually good news for original publishers.

Provenance becomes infrastructure. Content-provenance standards (C2PA, the cryptographic signing of media at point of capture) move from optional to expected over 2027-2030. The web in 2030 has a "verified human" signal in roughly the way the early web had a "verified business" signal (HTTPS). Sites with a strong human-provenance signal — verified authors, signed photos, dated commits in the open — are trusted differently from sites without one.

The flip side: AI gets better at hiding the signature. The current wave of slop is identifiable because the models are still in their teenage phase. By 2028, the best models will produce outputs whose statistical fingerprint is harder to detect. The lookalike test will get harder. The defense will shift from detection to provenance: instead of "is this AI?", "is there an authenticated human in this loop?".

The taste problem. All of the above is a taste problem at root. AI does not have taste, in the sense of internal preferences shaped by personal experience. It has averages. The future of the web depends on whether human taste — the ability to make non-average choices — remains visible. The sites that win in 2030 will be the sites that look least like any other site.

15. FAQ

Q1. Is AI slop the same as AI content? No. Slop is a subset of AI content: the subset that is indistinguishable from generator defaults and shows no human intervention. A page generated by an LLM and then thoughtfully edited, repaletted, and rewritten is AI content but not slop. A page generated by an LLM and shipped as-is is slop.

Q2. Will Google penalize my site if it looks AI? Not for being AI. For being indistinguishable, low-engagement, and clustered with similar sites. The Helpful Content Update is the operational mechanism. If your site has substantive content, real backlinks, and human signals, it survives even if part of the workflow used AI.

Q3. Is shadcn/ui slop? shadcn itself is not slop. It is a well-designed component library. *Unmodified* shadcn is slop, because it makes your site look like every other shadcn-default site. Customize the theme tokens.

Q4. Can I detect slop with code? Partially. Visual fingerprints (palette, type, layout) are detectable with computer-vision and CSS-parsing tools. Textual fingerprints (banned phrases, em-dash density, conclusion patterns) are detectable with NLP. Structural fingerprints are harder. The combined signal is reliable above a threshold but not perfect.

Q5. Is Inter slop? Inter is a fine typeface used in too many places. The slop is the *combination* of Inter + Tailwind defaults + shadcn defaults + Lucide. Use Inter with a distinctive headline serif, a custom color palette, and intentional layout choices, and the site is not slop.

Q6. If I use v0 to generate components, am I building slop? Not necessarily. v0 is a starting point. If you take v0 output and customize the palette, rewrite the copy, change the typography, and break the symmetric grid, you are building human work. If you accept the defaults and ship, you are building slop.

Q7. Why does my AI-generated site rank for the first month and then drop? Behavioral signals catch up. Initial rankings are based on relevance. After 4-8 weeks, Google factors in bounce rate, dwell time, return visits, and links. Slop content scores low on all of these.

Q8. How do I know if my copy reads as AI? Run it through the banned-phrase list (section 6). Count em-dashes per 1,000 words (above 4 is suspect). Check whether the conclusion paraphrases the introduction. Show three paragraphs to someone who doesn't know your category and ask them to guess what it's about. If they say "some kind of SaaS", you have textual slop.

Q9. Is using Tailwind slop? Tailwind is fine. Default Tailwind classes (bg-blue-600, rounded-2xl, shadow-md) on every site is slop. Customize the theme. Add custom colors. Override the default radii.

Q10. What's the cheapest single fix? Replace your color palette with one that doesn't match Tailwind defaults. It takes thirty minutes and breaks the strongest visual fingerprint.

Q11. Should I delete my v0-generated site and start over? Probably not. Strip the slop signatures using section 12's priority list. The underlying components are usually fine. The defaults are the problem.

Q12. Is hand-coding always less slop than generator-built? No. A hand-coded site by a developer who reaches for the same defaults the generators reach for produces the same slop. The tool is not the variable. The taste is.

Q13. Does using a serif typeface automatically de-slop my site? It helps. The serif breaks the strongest typographic fingerprint. But if your copy is still slop and your layout is still slop, the serif is decoration on a generic site. All dimensions matter.

Q14. How does Sailop differ from existing AI-content detectors? Existing detectors focus on text and try to classify whether copy was LLM-generated. Sailop is broader: it scores design slop across nine dimensions including visual, structural, and behavioral. It is also opinionated about *what to do* — not just "this is AI", but "your gradient hero is the issue and here is the fix".

Q15. Will users notice my site looks slop? They notice cumulatively. Any single slop signal feels familiar, not alarming. Ten slop signals together produce the "this could be anyone" reaction that users blame on the brand without knowing why. Conversion drops accordingly.

Q16. Is video slop already a problem in 2026? It's growing. Video slop is most visible in low-budget YouTube content (faceless channels), TikTok b-roll, and corporate explainer videos. It will be a major web-content category by 2028.

Q17. What's the role of human writers in 2026? Higher-leverage than ever, but in fewer roles. The freelance content middle has collapsed. The top — writers with distinctive voice and verifiable expertise — charges more than they did in 2022.

Q18. Should small businesses worry about slop? Yes, but proportionately. A local plumbing business does not need a non-Tailwind palette. It needs a site that loads fast, has a clear phone number, and ranks for "plumber [city]". Slop matters most where competition is identical-looking sites — i.e., in saturated SaaS, agency, and content-farm categories.

Q19. Can I write a great article in 2026 with AI help? Yes. The pattern that works: outline by hand, draft with AI, rewrite the AI's draft heavily by hand, fact-check, ship. The pattern that fails: prompt AI for a "complete article", paste it into your blog, ship.

Q20. Where does this trend go? Two paths. Path one: AI gets better at hiding its signature, slop becomes harder to detect, the web becomes uniformly mediocre, search engines lose meaning, and the audience flees to verified-human enclaves (Substack, Discord, in-person events). Path two: provenance standards mature, AI content gets labeled, the web bifurcates into "human" and "AI" tiers with different economic and SEO profiles. The probability is split between these two roughly equally.

16. Glossary

AI Slop — Content that bears the statistical fingerprint of generator defaults without enough human intervention to mask it.

Anti-pattern — A common solution to a problem that creates more problems than it solves; in this report, the default outputs of generators that produce indistinguishable websites.

Banner blindness — The 2000s phenomenon where users mentally filter out display ads; the 2026 equivalent is "slop blindness", where users tune out generator-default sites entirely.

C2PA — Coalition for Content Provenance and Authenticity. Cryptographic standard for signing media at the point of capture, enabling provenance verification.

Content provenance — The verifiable origin of a piece of content (where it came from, who made it, when, with what tools).

Cruip — A Tailwind UI template marketplace. One of the larger players in the post-shadcn template economy.

Default acceptance — The behavior of accepting the first plausible output a generator produces without modifying it. The root cause of most slop.

Design monoculture — The phenomenon where a single design pattern (Tailwind defaults, shadcn defaults) dominates a category to the point that most sites are visually interchangeable.

EEAT — Experience, Expertise, Authoritativeness, Trustworthiness. Google's quality rubric for content.

Em-dash giveaway — The textual fingerprint that AI-generated copy contains em-dashes at densities exceeding most human writers. The single most reliable textual slop signal.

Generator — A tool that produces websites or apps as direct output (v0, Bolt, Lovable). Distinguished from a "driver" (Cursor, Claude Code) that helps a human produce code.

Generator-default — The output of a generator with no customization applied. The most common form of slop.

Geist — Vercel's typeface, designed in part as an alternative to Inter that ended up becoming a similarly-recognizable default.

Helpful Content Update — Google's series of algorithm updates aimed at demoting low-effort content. The operational mechanism by which slop is penalized in search.

Inter — The most-used webfont on the indexed web. A fine typeface that has become a slop signal through ubiquity.

Lookalike test — A diagnostic for visual slop: reduce your homepage to a black-on-white silhouette and lay it next to five competitors. If you can't tell yours apart, you have a problem.

Lucide — The default icon set for shadcn/ui. Source of the Sparkles, Zap, Shield, and BarChart3 icons that recur across nearly every AI-generated site.

MCP server — Model Context Protocol server. Anthropic's standard for exposing tool functionality to LLMs (e.g., Claude Code) running in agentic loops.

Prompt injection-in-design — The phenomenon where natural-language design specifications, processed through a generator, produce results that reflect the model's biases rather than the human's intent.

Provenance — See "content provenance".

Rounded-2xl — Tailwind's class for a 16px border radius. The default radius on most generator-built cards. A visual slop signal.

RLHF — Reinforcement Learning from Human Feedback. The training method that shapes model preferences. Has the side effect of reinforcing whatever defaults the human raters approved of.

shadcn/ui — A widely-used React component library distributed as copy-paste source. Excellent in itself; problematic in its unmodified form because of how recognizable its defaults are.

Slop by negligence — Slop produced by lack of attention rather than malice. The most common form.

Slop by design — Slop produced intentionally by SEO operators running content farms. The form most directly targeted by Helpful Content Updates.

Stochastic parrot — Bender, Gebru, McMillan-Major, and Shmitchell's 2021 framing of LLMs as systems that recombine training-data patterns without understanding. A useful frame for thinking about why slop exists.

Tailwind — The utility-first CSS framework that won the post-2020 web. Source of the default classes (bg-blue-600, etc.) that recur across slop sites.

Template marketplace — A site selling pre-built website templates (ThemeForest, Cruip, Tailwind UI). A growth category in the slop economy.

Tricolon — A rhetorical pattern of three parallel elements. Used disproportionately by LLMs in marketing copy.

Vibe coding — Karpathy's 2025 term for the practice of building software by describing intentions to an LLM and accepting its output without close review. The dominant practice that produces most slop.

Voice — The recognizable personality of a piece of writing or design. The dimension hardest for generators to fake and most rewarded when present.

Web stripper — A new freelance niche: professionals who take generator-built sites and remove the slop signatures.

Wrapper (AI wrapper) — A product whose entire value proposition is a UI around a third-party LLM API. A high-density slop category in 2025-2026.

XOXO — Andy Baio's festival, where the term "slop" was used in 2024 commentary by Baio and others. Origin point for the term's adoption in tech discourse.

17. Sources & further reading

This report cites no fabricated URLs. The references below are real authors, real publications, and real years. Search them directly.

On the term "slop":

• Andy Baio, *Waxy.org*, blog posts on AI-generated content (2023-2024).
• Simon Willison, *simonwillison.net*, multiple posts on slop and the slop-spam analogy (2024-2026).
• Wikipedia, "AI slop" entry (created 2025, updated through 2026).

On AI content and the web:

• 404 Media, ongoing reporting on AI content farms (Joseph Cox, Samantha Cole, Jason Koebler, 2024-2026).
• The Verge, multiple investigations into AI-generated articles and search results (2023-2026).
• MIT Technology Review, coverage of generative AI's impact on online media (2023-2026).
• Wired, reporting on AI in publishing and the open web (2023-2026).
• The Guardian, AI and journalism coverage (2023-2026).
• Ed Zitron, *Where's Your Ed At*, sustained criticism of generative-AI hype (2023-2026).
• Pivot to AI (David Gerard), tracking generative-AI failures (2024-2026).

On search engines and quality:

• Google Search Central, Helpful Content Update documentation and blog posts (2022-2026).
• Search Engine Roundtable (Barry Schwartz), running coverage of algorithm changes (2022-2026).
• Search Engine Journal, EEAT and AI-content guidance pieces (2022-2026).

On design and the web:

• Smashing Magazine, design-systems and accessibility coverage (ongoing).
• A List Apart, foundational web design and writing essays (ongoing).
• Stratechery (Ben Thompson), business analysis of platforms and AI (ongoing). An example of long-form web that resists every slop dimension.
• Linear's blog, on building distinctive product surfaces (ongoing).
• Vercel and Next.js blogs, on the design system that shapes much of the post-2020 web (ongoing).

On the foundations of AI critique:

• Bender, Gebru, McMillan-Major, Shmitchell, "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" (2021).
• Andrej Karpathy, public commentary on "vibe coding" and the practice of building with LLMs (2025).
• Anthropic blog, on Claude's behavior, alignment, and design (ongoing).
• OpenAI blog, on GPT releases and applications (ongoing).

Counter-examples — sites that resist slop:

• *Linear* (linear.app) — distinctive design system, dense content, high taste.
• *Vercel* (vercel.com) — has its own clichés but iterates them with care.
• *Stripe* (stripe.com) — durable design across many years.
• *Anthropic* (anthropic.com) — restrained, typographic, atypical for the AI category.
• *Are.na* (are.na) — almost antithetical to slop; a network of intentional human curation.
• *Bandcamp* (bandcamp.com) — long-running, distinctive, resistant to redesign trends.
• *Drop* (drop.com, formerly Massdrop) — community-curated, dense, idiosyncratic.
• *Craigslist* (craigslist.org) — the durable opposite of slop.
• *Pinboard* (pinboard.in) — a personal product, written and run by Maciej Cegłowski. The web's anti-slop reference point.
• *Stratechery* (stratechery.com) — one writer, one voice, one durable design.

Internal links across this site:

• Anti-slop prompt template 2026
• De-AI your Lovable / v0 / Bolt site
• Tailwind blue-purple gradient: the AI signature of 2026
• Detect AI-generated site in 30 seconds — 21 signs
• Vibe coding 2026: honest state of AI frontends
• The AI slop economy 2026

The web in 2026 is the most generated artifact humans have ever made. Most of it is slop. A small, increasingly valuable fraction is not. The difference between the two is taste — the willingness to spend the few extra hours that pull a page off the generator centroid. Sailop's bet, and the bet of every shop and writer and developer who built work that resists this trend, is that those hours pay back. The data so far suggests they do.

End of report.