AI NativeAI answers can get things wrong. A model can state an outdated interest rate, quote the wrong eligibility age, or invert a statistic. For regulated and detail-heavy products, a wrong number in an AI answer is not a minor inconvenience. It shapes a buyer's first understanding of what you offer. The accuracy audit is how AI Native catches those contradictions before they go unnoticed.
What gates the accuracy audit
The audit only runs against facts that a strategist has marked as verified in the Brand Truth Studio. This is a deliberate constraint, not a limitation. An unverified number cannot reliably check an answer because the number itself might be wrong. Only facts a human has confirmed are accurate, authoritative, and current are allowed to gate the audit.
When you add facts about your product, they arrive as unverified candidates. The system AI-gathers a proposed set based on your product type and vertical, and you review, correct, and confirm each one. A confirmed fact becomes a truth anchor. An unconfirmed fact stays in the candidate list and never influences an accuracy verdict, because scoring an answer against a guess would produce meaningless results.
How the check works
For each branded prompt's answer, the engine looks for numeric claims the answer makes about your product. It scans the answer text for any claim that corresponds to a verified fact by matching on the fact's label, key, and any aliases you have defined. When it finds a match, it compares the claimed number to the truth.
A claimed number that falls within a narrow tolerance of the truth is recorded as accurate. A number that diverges beyond that tolerance is recorded as inaccurate and flagged with the detail: what was claimed, what the truth is, the unit, and the source URL you confirmed. A numeric claim that references a topic AI Native recognises but cannot match to a specific verified fact is recorded as unverified. If the answer makes no checkable numeric claim at all, the state is not applicable.
This deterministic check runs on every branded answer. On live scans, it runs alongside the LLM accuracy judge and the numeric check takes precedence: if a number is flagged as inaccurate by the deterministic check, that verdict is final regardless of what the LLM judge thought.
The four accuracy states
Accurate means at least one checkable claim was found and it matched the truth within tolerance. No contradictions were detected.
Inaccurate means at least one claim was found that contradicts a verified fact. The detail row shows you exactly which claim failed, what was said, and what the truth is.
Unverified means the answer stated a number that looked relevant but did not match any of your verified fact aliases closely enough to check. This is not a contradiction finding. It is a gap: either the fact needs an alias added, or the claim is about something not yet in your verified set.
Not applicable means the answer made no checkable numeric claim. This is the normal state for brand-presence answers on unbranded prompts, or for answers that describe your brand qualitatively without citing numbers.
What the audit does not cover
The audit checks numeric claims only. Non-numeric claims, things like channel descriptions, eligibility criteria stated in prose, or product positioning language, are assessed by the LLM judge in the classification step, not the deterministic accuracy check. That distinction matters because the deterministic check is auditable and source-anchored. For numeric claims it is more reliable than a language model reading the sentence, which is why contradicted numbers always take precedence.
The audit also only fires on branded prompts. Unbranded category questions are not where your specific facts are likely to appear, and fact-checking an answer about a category against one brand's verified numbers would produce false positives.
Acting on an inaccurate verdict
When an answer is flagged inaccurate, the detail view shows you the specific claim, the truth, and the source URL from your Brand Truth Studio. The right next step depends on whether the answer was grounded or parametric. A grounded answer that gets a fact wrong is drawing on a source that has incorrect or outdated information about you. Updating that source, or making the correct information more prominent and citable, is the direct fix. A parametric answer that gets a fact wrong reflects training data that the model has not updated; entity work and high-authority primary-source publication are the slower but necessary path.
Questions
Can the accuracy audit run without verified facts?
No. The audit requires at least one verified fact to have something to check against. If your Brand Truth Studio has no verified facts, every branded answer will return not applicable for accuracy. Adding and verifying facts is the prerequisite for meaningful accuracy scores.
What does "unverified" mean as an accuracy state?
It means the answer stated a number that looked like it might relate to your product, but the number did not closely match any verified fact label or alias. It is a gap in coverage rather than a contradiction finding. Adding an alias to the relevant verified fact, or verifying a new fact that covers that claim, will convert future unverified states into accurate or inaccurate verdicts.
Does accuracy checking run in demo mode?
Yes. The demo fixture engine synthesises branded answers that include real numeric claims drawn from your verified facts, and on a deterministic fraction of them it deliberately introduces a wrong number. The accuracy check then runs on those answers exactly as it would on live answers. This gives you a realistic view of what the audit surfaces before you spend credits on a live scan.
How is the numeric tolerance set?
A claimed number is counted as accurate if it falls within twelve percent of the truth. This tolerance handles rounding differences, currency conversions, and minor display variations. A number that is ten percent higher than the truth is still a reasonable representation. A number that is two or three times the truth is always flagged, even though some models do invert or scale numbers significantly. The tolerance is applied symmetrically.
Why does a contradicted number override the LLM judge's verdict?
The deterministic check is source-anchored: it has a specific truth to compare against and a specific claim the answer made. A language model reading the whole answer can be confused by context, hedging language, or an answer that qualifies the number correctly. The deterministic check does not care about context; if the number is wrong, the number is wrong. Letting a clear factual contradiction be overridden by a more lenient qualitative reading would undermine the audit's value.
Where can I set up verified facts?
Verified facts are managed in the Brand Truth Studio. See the Brand Truth Studio guide for how to review candidates, confirm facts, and add aliases so the check covers the specific ways your facts get stated in answers.