Is your site AI slop?
Paste an address. We render the page, look for builder fingerprints and stock component patterns, and quote every piece of evidence we find. Free, no signup, about five seconds on top of the normal audit.
You land on the audit form with the AI check already ticked. Press Run.
What we look at
Builder fingerprints
Lovable, v0, Bolt, Base44, Replit and ChatGPT Sites generate a site from scratch and leave product markers in scripts, meta tags and asset hosts. Those count only where code lives, never in the text. A page that mentions v0 in a sentence was not built by v0.
Coding-agent residue
Claude Code, Codex, Cursor and similar tools edit a repository instead of generating a product. They leave instruction files, backup copies with the tool's name, "generated with" markers, and a recognisable signature in how the CSS was overridden. A fixed fingerprint list scores such a site zero. We look for the residue.
Stock component patterns
Forty named page elements, each weighed on 640 sites against pages built without AI. Some raise the odds: an accent-coloured word in the headline, one particular icon set, a dot-separated feature line, gradient text, numbered steps, three icon cards when they really are cards. Some lower them: customer logos, testimonials with faces, a four-column footer. The announcement pill above the headline, which almost everyone calls a tell, is neutral: designers reach for it just as often, so it does not count.
What the verdict means
- Confirmed
- A build artefact names the tool: a generator tag, a builder script, an instruction file served from the root. Right every time in our validation.
- Likely
- Strong stack signals, or a cluster of measured patterns with little authored work around them. Comes with a confidence number.
- Possibly
- Stock components were found. Designers use the same kits, so the verdict is capped here and no number is shown.
- No evidence found
- Nothing generated was found. This is not a certificate of human work: on our labelled set it was wrong 7% of the time.
- Not measured
- The page could not be fetched or rendered, a bot wall answered instead of the page, or the page has too little content to measure (a single-canvas app, an empty shell).
How often it is right
Measured on 181 sites with reliable labels, September 2026. Another 61 are hosts that refuse a scanner; they are excluded, not counted as clean. Only a new corpus run changes these numbers.
| Verdict | Measured |
|---|---|
| Confirmed, when we say it | correct 100% |
| Likely, when we say it | correct 98% |
| Confirmed or Likely, AI-built sites caught | 68% |
| No evidence found, sites that were in fact AI-built | 7% |
Sites whose makers said they used a coding agent are harder. Of 24 such sites (Cursor, Claude Code, Bolt), 15 could be read and the check caught 6 of those; the other 9 were unreachable or carried too little to measure. An agent working inside a designed product leaves less residue than a browser builder, so a clean result on that kind of site is worth less. We publish the number so you know.
Tool by tool
| Tool | What it leaves behind |
|---|---|
| Lovable | A builder script, the badge, the default social image, tagged elements. |
| v0 by Vercel | A generator tag and images on the Vercel blob host. |
| Bolt | The builder host and a server header. |
| Base44, Replit, ChatGPT Sites | The app host in a technical slot, runtime token names, the one-line app shell. |
| Pasted ChatGPT, Gemini or Perplexity text | Citation markers the chat left in the HTML. |
| Claude Code, Codex, Cursor, Windsurf | Instruction files served from the root, backup copies carrying the tool's name, "generated with" markers, the override signature in the CSS. |
| Framer, Wix, Webflow | Counted as counter-evidence. They are site builders, not AI. |
What it will not tell you
- It never says a site was made by a human. It reports the evidence it found, and its absence.
- It checks one page, the one you give it. Site-wide checks are part of the paid audit we are building.
- A quarter of AI-built sites are empty shells filled in by JavaScript. We render them, and the report says when a static fetch would have seen nothing.
- A well-executed site on shadcn or Tailwind lands in Possibly. That is the design, not a bug.
- It is built for websites: landing pages, marketing sites, product pages. Games and single-canvas apps show it almost nothing, and it says so rather than calling them clean: of 40 games from a vibe-coding jam, 22 came back as not measured and it caught 1 of the 18 it could read.
- A verdict is an automated estimate from public signals, not a statement of fact about a site or its authors.
If we got it wrong
Every report has two buttons under the verdict: wrong, and right. What you send goes into the next corpus run, and the numbers on this page change only when a run changes them.