Accuracy log
Every time we find and fix a false positive, a missed case, or add a new signal, we write it up here — including the ones that make us look bad. A detector that hides its own mistakes isn't one worth trusting.
Three false positives found by scanning real, well-known sites
While gathering material for a blog post, we ran our own detector against 20 well-known sites and checked the evidence before publishing any of it. Good thing we did: figma.com scored 100% "Wix ADI" because of class="wix-madefor-display" — the unrelated, open-source "Wix Madefor" Google Font, not the Wix ADI builder. The same scan also matched "Duda" via class="dm-mono", the unrelated "DM Mono" Google Font. Separately, shopify.com and framer.com both scored inflated as "Shopify Magic" purely for referencing Shopify's own CDN — evidence present on every ordinary Shopify store regardless of whether its AI feature was ever used. We removed both risky class-name patterns, and removed the Shopify Magic rule (and catalog entry) entirely, since none of its possible evidence can actually distinguish "used Shopify Magic" from "is a Shopify store." Added regression tests for all three. Full writeup on the blog.
Fixed a false positive: outbound links were being read as attribution
Several coding-tool and builder rules matched on a platform's bare marketing domain (bolt.new, lovable.dev, v0.dev, replit.com, plus a few builder-catalog domains) instead of requiring an actual attribution claim. That meant a page simply linking to one of these tools — to recommend it, review it, or list it in a comparison — scored the same as a page that platform actually generated. We found this on our own homepage first. Fixed by requiring real attribution language next to the tool's name, and added regression tests asserting a plain reference link produces zero signals. Full story on the blog.
Added Claude Code to the catalog and detection
Claude Code was missing from our builder catalog despite being a common AI coding agent. Added a catalog entry and a weak trace rule matching explicit attributions. Also documented in the FAQ why this rarely flips a verdict: local agentic tools edit a real codebase like a human developer would and don't watermark the shipped output the way hosted builders do.
Found something we got wrong? Tell us at info@aiwebsitedetector.net — if it's a real issue, it'll end up on this page too. Read the full methodology for how detection works today.