AI Visibility
Fact Check: Is What AI Says About You True?
Check the factual claims AI engines make about your pricing, specs, features, availability and policies against your own verified content — and trace each wrong answer back to the page that caused it.
AI engines do not just answer the question they were asked — they volunteer detail. That is how wrong pricing, superseded specs, and policies you changed two years ago end up in front of buyers. Fact Check finds those claims before your customers do.
How it works
- 1We pick the prompts worth checking. Prompts with a factual answer ("how much does X cost", "does it integrate with Y", "what is the refund policy") are selected automatically. Opinion prompts are skipped — the perception scorecard already covers those. Override any prompt in the "Prompts being checked" panel.
- 2We read your tracked answers. Every claim an engine made about your brand is pulled out verbatim, never paraphrased, and tagged as pricing, specs, features, availability, policy or company fact.
- 3We check each claim against your ground truth. Your project knowledge base — scraped pages, uploaded files, and approved Knowledge Studio documents — is the only source used. Approved documents outrank scraped pages.
- 4We trace the wrong answers back. For each inaccurate claim we look at the citations sitting nearest to it in the answer, and label each domain as your site, a competitor, or a third party.
You need a knowledge base first
Fact Check will not run against an empty knowledge base. Without verified content there is nothing to compare against, so every claim would come back "could not verify" — and the run would spend credits to tell you nothing. Scrape your site from the Fact Check tab, or curate documents in Knowledge Studio, then run it.
Reading the three verdicts
- Accurate — your content supports what AI said.
- Wrong — your content *explicitly contradicts* what AI said. The correction shown is drawn from your own knowledge base, with the source excerpt beside it.
- Could not verify — your knowledge base does not cover this claim. This is not an accusation; it is the honest answer, and it is why coverage is always shown next to the score.
Missing evidence is never counted against you
A thin knowledge base must not make a brand look like it is being widely misrepresented. Unverifiable claims are excluded from the accuracy score entirely and reported separately as coverage. If coverage is low, the fix is to add verified content — not to distrust the score.
The accuracy score
Accuracy is the share of verifiable claims that were correct, smoothed toward neutral so a run with only two or three verified claims cannot read as a perfect 100. The unsmoothed figure appears next to it as "raw". Once a fact check has run, accuracy also becomes one of the five inputs to your overall Alignment score, weighted at 25%. Projects that have never run one are scored exactly as before — you are never penalized for not having used the feature.
Fixing a wrong claim
The right fix depends entirely on who owns the page AI read, which is why every source is labelled:
- Your site — the page itself is out of date. Update the stated facts, delete the superseded number rather than burying it further down, and mirror structured facts in schema. This is the highest-priority fix: your own page is the one misinforming AI.
- A competitor — their page is the most quotable answer to the question, so AI uses their framing of you. Publish a better one of your own.
- A third party — an independent site is carrying stale facts and AI trusts it. Use "Request a correction" on the claim to open an outreach opportunity prefilled with the claim, the correction, and your source link.
Tracking improvement
Each completed run adds a point to the accuracy trend chart. Run a check, fix the sources it names, then run it again once those changes have been crawled — the line is your proof that the work moved AI answers.
Curated documents beat scraped pages
Approved Knowledge Studio documents rank first as evidence and carry drift detection, so they stay correct over time. If one claim type keeps coming back unverifiable — pricing is the usual culprit — writing a single approved document covering it is the biggest lever you have on the accuracy score.