Skip to content
RankX AI
RankX AI Docs
AI Visibility

Answer accuracy

Whether what AI assistants say about your business is true. Seven verdicts, judged in code against the facts you have published rather than by a second model.

Answer accuracy asks a different question from the rest of AI visibility. Visibility asks whether you were named. Accuracy asks whether what was said about you was true. An assistant that recommends you confidently and quotes a price you have never charged is a visibility success and a commercial problem, and only this check can tell the two apart.

RankX AI runs it over the answers it has already collected for one brand prompt, pulls out every factual claim those answers make about your business, and judges each claim against what you have published in Brand Book. It costs 5 credits per run, which covers one extraction pass over every stored platform answer for that prompt.

The seven verdicts, worst first

RankX AI grades every extracted claim into one of seven verdicts. The order is the answer to one question: how wrong is the buyer left?

VerdictWhat it means
Says something you have bannedThe answer used a phrase you recorded in Brand Book as one your business must never be described with
Contradicts your own factsA confident claim RankX AI can disprove from what you have published
Describes a different businessA fluent, confident answer about somebody else with a similar name
Makes it upA confident claim with nothing behind it, in a category where you have published something
Does not know youThe assistant said it has not heard of you
Says something you have never publishedThe claim may be right, but nothing on your side backs it
Describes you correctlyThe claim matches a fact you published

"Does not know you" ranks below the three above it on purpose. An assistant saying it has not heard of you leaves the buyer with no answer, which is bad. It does not leave them with a wrong one, which is worse.

"Says something you have never published" is a finding, not a degraded state. It is the cleanest action in the product: an assistant told a buyer something about your business, and there is nothing anywhere you control that either supports or corrects it. The remedy is yours, and RankX AI names the exact Brand Book field that would close it rather than telling you to add more detail.

What gets judged, in seven categories

The model that reads the answers classifies each claim into one of seven categories, and nothing else:

CategoryThe claim is about
identitywhat your business is
capabilitywhat you do
pricingwhat you cost
audiencewho you are for
provenanceyour background
riska warning about you
integrationwhat you work with

The category decides which of your published facts the claim is checked against, and it decides what the remedy names when nothing backs the claim. A pricing claim with no ground truth behind it sends you to the price on one of your services; an integration claim sends you to your integrations list.

The model extracts, the code judges

This is the design decision the whole feature rests on, and it is worth understanding before you read a result.

The model is asked for claims only. For each stored answer it returns the claim, its category, and the verbatim span of the answer the claim came from. It never sees your published facts, never returns a verdict, and never scores anything.

Every verdict is computed in ordinary code, from the claim plus your published facts. RankX AI does not ask one model to mark another model's homework, because two models share a failure mode: a judge model agreeing with an extractor model is not evidence that either was right. Reading prose and saying "this sentence claims the company charges fifty pounds a month" is the one part a model is genuinely better at than code, and it is the only part it is asked to do.

A quote that is not in the answer cannot survive. If the verbatim span the model returned does not appear in the stored response, the claim is discarded before any verdict is computed. There is no code path on which a fabricated quote reaches your account. The number discarded is reported rather than swallowed, because an exclusion that does not report itself reads as "we covered everything".

Uncertainty always resolves the same way

The two available errors are not symmetrical, and RankX AI picks its error deliberately.

Telling you an assistant is lying about your business when it is not is a claim about a third party that RankX AI cannot substantiate. Telling you that you have not published something clearly enough when you have costs you a look at your Brand Book. So every uncertain band resolves to "says something you have never published", and the strongest accusations require the strongest evidence. "Makes it up" needs the claim to share not one meaningful word with anything you have published in that category.

If you think a verdict is too soft, that is the design working as intended.

What makes the check possible

Answer accuracy needs two things, and it is worth checking both before you spend the credits.

Stored answers. The check reads answers RankX AI has already collected, so a prompt with no runs behind it has nothing to judge. See AI Prompts.

Published facts. A claim can only be contradicted by something you have written down. An empty Brand Book produces a page of "says something you have never published", which is technically correct and not useful. Filling in your services, your audience and a price on at least one service is what turns the check from a list of gaps into a list of errors. See Brand Book.

Only facts that are approved, public and not confidential are used, which is the same predicate that governs what reaches your generated llms.txt. A confidential note in Brand Book never becomes ground truth and never becomes published text.

Where it sits, and what it costs

Answer accuracy runs from the AI Readiness page, as a band beneath the readiness score. It is 5 credits per run, per brand prompt, covering one extraction pass over every platform answer stored for that prompt.

It is a metered check, so it is not part of the free band that stays available after a trial ends.

Where to go next

Last updated