Several GEO (generative engine optimization) tools now offer a feature that checks AI answers for wrong facts about your brand: an old price, a feature you don't offer, an office you closed. They report the result as an accuracy score or a list of flagged claims.

That score is easy to misread. It does not measure how often AI is wrong about you. It measures how often AI answers agree with the facts you gave the tool, on the prompts it checks, divided by whatever the vendor counts. And each verdict is itself a judgment that can be mistaken. This article takes those four factors in turn, using vendor documentation read on 12 October 2026, and ends with what to check before relying on the number.

What a fact-check feature actually does

The features work in the same basic way. The tool takes AI answers to a set of prompts and splits each answer into single factual claims. It then compares every claim with a fact base you supply. Each claim gets a verdict, such as "accurate", "contradicted" or "not verifiable".

This is a different measurement from sentiment, which scores tone: a warm answer can still state the wrong price. As our guide to mentions, citations, sentiment and share of voice notes, tone is not accuracy. The details differ by tool, and they decide what the score means.

Which tools document it, and how they differ

We found four vendors that describe a claim-versus-facts feature in their product documentation: Profound (FactCheck), Peec AI (Fact-checking, part of Brand Perception), AthenaHQ (Oracle) and Genezio (grounded Perceptions).

Profound FactCheck Peec AI Fact-checking AthenaHQ Oracle Genezio Perceptions
Where facts come from One connected Knowledge Base; can recrawl your website on a schedule Facts you add; AI can draft them from your website for review Verified facts in your Knowledge Base Knowledge Base; website crawled automatically, plus documents, URLs and text
Setup All plans; not enabled by default; set up per workspace; 100–200 prompts recommended, 200 maximum without contacting Profound Plan limits apply but are not published Oracle access, at least 25 verified facts and a completed prompt run Knowledge Base built at brand creation
Prompts checked Prompts you tag; branded recommended but not required Tracked prompts you select; only answers naming a tracked brand Responses from your latest completed prompt run New perceptions from your tracked conversations
What gets a verdict Single claims Single claims Single claims; the score counts responses with at least one confirmed finding Single claims ("perceptions")
History Results after 7 days; nothing from before setup Only answers generated after a prompt is selected; claims can take up to a day to appear Each scan covers the latest run; earlier scans kept in Run history; imported older runs have no accuracy rate Not stated in the pages we read
Correcting a verdict Report a wrong verdict; the claim is removed from your view Change a verdict; it is marked "Edited" and can be restored No verdict override documented; Acknowledge or Ignore changes review status only Not stated in the pages we read

Scrunch announced a Fact Check Agent as part of Knowledge Studio, a free beta for enterprise customers, in a blog post on 11 June 2026, but we found no product documentation for it. Otterly.AI lists detecting inaccurate mentions as a use case without describing how. For Semrush, Ahrefs Brand Radar and Rankscale, we found no documentation of a claim-versus-facts feature, which is not proof the capability is absent.

The table shows the inputs differ. The first input, and the one that limits everything else, is the fact base.

Factor 1: the facts you supplied

A claim can only be judged against a fact that speaks to it, and the vendors say so directly.

Peec AI's documentation states: "A claim no fact of yours speaks to isn't given a verdict at all, so it doesn't show up as a finding." It adds that this is "why AI can be wrong about you for months without anything appearing here." Profound labels such claims "not relevant", and Genezio labels them "not grounded."

Partial facts can be worse than missing ones. Profound gives its own example: a claim about an old pricing tier "might be marked accurate if the Knowledge Base doesn't include the updated pricing page." Genezio's docs make the same point: "Old truth is worse than no truth."

What this means: a high score with a thin fact base means "little was checked", not "little is wrong".

Factor 2: the prompts that are checked

Facts decide which claims can be judged. Prompts decide which claims are collected at all, and here the tools split into two groups.

Profound and Peec AI check only the prompts you tag or select. Profound's limit is 200, and Peec's limit depends on your plan. AthenaHQ and Genezio check answers from the prompts you already track, with no separate selection step. AthenaHQ's scan supports websites tracking up to 2,000 live prompts; above that, the scan fails. In every case, a wrong claim that appears in an answer to a question you don't track will not be seen.

The kind of prompt matters too. A branded prompt asks about you directly; Profound's examples include "How much does Profound's Pro plan cost?". A broad category question shows what AI says when your brand comes up on its own. Profound recommends branded prompts but does not require them. Peec suggests prompts where a wrong answer affects a buying decision, such as pricing questions and plan comparisons. AthenaHQ and Genezio check whatever mix of prompts you already track.

Genezio also documents a separate Fact Checker agent that asks engines about specific claims. Direct questions and organic answers measure different things, so check which kind feeds the number.

Factor 3: what the percentage divides by

Once claims are collected and judged, the tool turns them into a percentage. A percentage is only meaningful once you know its denominator, the thing it divides by.

  • Profound defines its Accuracy Score as accurate claims divided by accurate plus inaccurate claims. Claims that can't be checked against the Knowledge Base are left out.
  • Peec AI reports a contradicted share that "only includes claims that received a verdict." Its "By fact" view shows "Never comes up" for facts no claim has matched, so gaps stay visible.
  • AthenaHQ reports inaccuracy as "the share of scanned responses with a confirmed finding." Its accuracy by model is the share of sampled responses with no confirmed findings.
  • Genezio reports the share of perceptions that are correct, incorrect, or not verifiable, so the unverified share is shown next to the other two.

Profound's own glossary example shows how much can be left out. FactCheck extracts 50 claims, and 30 can be checked. Of those, 24 are accurate, so the score is 80%. The other 20 claims, 40% of the total, do not affect the score at all.

AthenaHQ judges single claims but scores whole responses. On our reading of its definition, a response with no checkable claim also has "no confirmed findings", so it counts toward accuracy. That is our interpretation, not a vendor statement, but it means AthenaHQ's percentage and Profound's are not the same measurement.

A published study shows how far the denominator alone can move the number. Seer Interactive, an agency that sells this kind of analysis, tracked 1,562 prompts about its own brand across six AI platforms between 1 May and 9 June 2026, and published the results on 27 August 2026. On average across the six platforms, 95.9% of answers were accurate when AI answered. Counted against every response, including the 31.2% where AI gave no answer, the average fell to 66%. The average inaccuracy rate was 2.8%. All three figures come from the same data.

What this means: don't compare accuracy percentages across tools, or against published studies, without checking that they divide by the same thing.

Factor 4: who makes the verdict

Even with the right facts, prompts and formula, each verdict is a judgment about whether a claim matches a fact, and that judgment can be wrong. AthenaHQ states that its findings "are generated using AI." Profound and Genezio describe the comparison as automatic but do not name the method; Peec AI does not say.

Profound and Peec AI both document a way to correct wrong verdicts, which shows they expect some. None of the four vendors publishes, in the pages we read, how often its verdicts are wrong. So someone has to review flagged claims before acting on them: the tool narrows down where to look, but it does not settle the question.

What to check before relying on the score

The four factors come down to a short list of questions to ask a vendor, or to check in a trial:

  1. Fact base. Where do facts come from, and how are they kept current when prices or products change? Can you see the claims that got no verdict?
  2. Prompts. Which prompts are checked, how many, and do they name your brand or ask broad category questions?
  3. Denominator. Does the score count claims or responses, and are unverifiable claims excluded?
  4. Verdicts. How are verdicts made, can you override them, and can you export claims with their verdicts?
  5. Sources. Does the tool show the cited pages behind a wrong claim? Treat them as leads, not proof.

None of these features changes what AI says. AthenaHQ notes that acknowledging a finding "does not edit the underlying fact, publish content, or change what an AI model says." An accuracy score is a way to find wrong claims worth fixing, and it is only as complete as the facts and prompts behind it.

For how these vendors compare on other capabilities, see GEO tools compared.