Fact-checking what AI says about your brand

Build the fact sheet first

You cannot fact-check an AI answer until you have written down what is true: a short, dated list of checkable claims about your brand, each with the page that proves it. Fact-checking without a fact sheet is two opinions arguing.

Four kinds of claim earn their place, because they are the ones models get wrong most often and the ones a buyer checks first: pricing shape (tiers, trial, no free plan), who you serve (company size, market, industry), founding facts (year, founders, headquarters), and integrations (which tools you connect to, and which ones you deliberately do not).

Keep every fact atomic and dated. “Founded in 2019, headquartered in Lisbon” is checkable; “a trusted global leader in workflow software” is not, and judging it would be an opinion dressed as a verdict. Revisit the sheet each quarter, because a fact sheet nobody updates becomes the source of the staleness you are trying to catch in everyone else.

The four verdicts, with an example of each

Every atomic claim found in answers that touches a registered fact gets exactly one verdict: Contradicted, Supported, Inconclusive or Not covered. Contradicted sorts first, because a false claim repeating across engines is the one costing you deals.

Contradicted

The answer states something your fact denies outright. This is the costly error, so the bar for declaring it is deliberately high.

Supported

The answer matches the registered fact. Nothing to do beyond keeping the fact current.

Inconclusive

The answer says something adjacent to a fact but not comparable as written: a number with no unit, a date the fact does not cover, a claim half-matching the register.

Not covered

The answer claims something about you that no registered fact speaks to. No verdict is possible, and none is invented.

Trace a contradicted claim back to its source

A Contradicted verdict is only actionable if you can see where the model got the idea, so each one links to the pages the answer cited. Open them before you argue with the model.

The first suspect is your own site. A stale pricing page from two redesigns ago, a forgotten landing page promising an integration you dropped, a press release from the acquisition you later reversed: models read what is published, and half of contradicted claims turn out to be your own archaeology. That page is the page to fix, and fixing it is faster than any conversation with an engine.

When the cited source is a third party, a review site or a directory listing, the fix is outreach rather than editing: correct the listing, pitch the reviewer, add a dated facts page the next crawler can read. Either way the trail matters, because without it a wrong claim ends in a hallway argument about what the model probably meant.

What to do with each verdict

The verdict tells you the move: fix the source, protect the fact, sharpen your register, or ignore it for now.

• Contradicted: update or remove the source page, then re-run. Expect a lag before engines pick up the correction, and re-check the claim in later waves rather than the next morning.
• Supported: no action. Note which sources earned the support, because those are the pages worth keeping alive.
• Inconclusive: make the underlying fact more specific so a comparison becomes possible, for example adding the since-when date the fact was missing.
• Not covered: decide whether buyers care. If they do, register the fact; if they do not, let it go and keep your sheet short.

What fact-checking cannot tell you

Only the facts you register are judged, so Not covered is the most common verdict by a wide margin, and that is the design working rather than a gap to panic about.

Answers make dozens of loose claims about any brand and a fact sheet holds a few dozen lines, so most claims land outside it. The tool never grades a claim against a fact you did not write, because an invented standard would produce confident verdicts nobody can defend. Grow the sheet as Not covered clusters reveal which loose claims keep recurring.

Also worth remembering when you read any verdict: a model's certainty is not a measure of its accuracy. OpenAI puts it plainly, that confidence is not reliability and the model may express high confidence even in incorrect answers (Does ChatGPT tell the truth?). A confidently worded wrong claim and a confidently worded right one look identical in the answer, which is exactly why the check runs against your register and not against the model's tone.