"Evidence or it didn't happen: the product bar that shapes everything we ship"

There's one rule in our repo that outranks every other product decision: a scan result must prove it read your actual site. Real browser screenshots, with stored provenance. Quoted text from your pages tied to every finding. The exact pages inspected, listed. If we can't show the evidence, we don't show the claim.

The rule exists because the failure mode of AI-powered audit tools is seductive: plausible, confident, generic. A language model will happily tell any business to "improve your calls-to-action" and "build trust with reviews" — advice that sounds diagnostic and is actually a horoscope. The moment a reader suspects the tool never really read their site, the product is dead. And our product's entire premise is judging what's verifiable — a tool that judges verifiability had better be verifiable itself.

The rule has teeth, and today it bit us. Our category inference — the part that says "this is a plumbing company" — was matching generic medical words and confidently labeling unrelated businesses as clinics. Wrong, specific, and confident: the worst combination. The fix we shipped wasn't just better matching. We added a display guard so that when stored evidence doesn't support a stored category, the result says "Business type not confidently inferred" — an honest shrug — rather than a confident lie. We'd rather show uncertainty than fake precision.

That principle cascades into decisions that would otherwise be debatable:

• Unreadable sites don't get scores. If a site's bot protection blocks us, we say so and refuse to produce a number. A score computed from a challenge page is fiction with a confidence interval.
• Screenshots are captures, never mockups. Each one stores the requested URL, final URL, viewport, and capture time. If the capture fails, the result says "unavailable" — it never substitutes a stand-in.
• The public gallery only shows real customer scans — actual scores, actual screenshots. Nothing staged. If eligible data doesn't exist, the section hides instead of rendering placeholders.
• Our own marketing gets the same bar. This blog quotes real numbers from our operations log, failures included, because a build-in-public blog that curates away its mistakes is just slower advertising.

There's a cost. Honest failure states mean sometimes delivering less than a customer hoped for — a "we couldn't read your site" instead of a shiny 76. We think that trade wins over any horizon longer than a week: the customers worth having are the ones deciding whether to trust you with $19 based on whether your free result proved itself.

See what an evidence-backed result looks like on your own site — free, about a minute: agentsitescan.com/scan.