How do you measure brand accuracy in generative AI answers?

Brand accuracy measurement borrows its structure from factuality research. The method is simple. The discipline is in the denominator.

1. Write a verified fact sheet. List the facts an AI answer is likely to state: pricing tiers, founding year, headquarters, product names, integrations, ownership. Each entry needs one correct value and the page where you publish it.
2. Collect answers under logged conditions. Ask the same buyer questions on each platform. Record the platform, the date, and whether the answer used live web search (visible citations are your clue).
3. Split each answer into atomic claims. "Founded in 2019 in Denver, with a $49 starter plan" is three claims. Label each one supported, contradicted, or unverifiable against your fact sheet.
4. Divide by what returned. Your error rate is wrong claims over total claims, or answers with at least one error over answers returned. Never divide by answers you requested. A platform that timed out gave you no evidence.
5. Keep a worst-error view. Report the most serious single error next to the rate.

If you want the platform-specific version of this, see how to tell whether ChatGPT is accurate about your company.

Every established factuality metric normalises by a count of observations. None of them sums penalties.

Two lessons matter for brands. The SAFE authors note that precision alone "does not penalize systems that abstain from answering too often or generate text with fewer facts." That is why a single rate needs a companion view. Vectara reports an answer rate alongside accuracy, so a high refusal rate can't pass for good accuracy.

The practical consequence: a summed penalty score falls as you ask more questions, even when the error rate stays flat. Say you find 10% of answers wrong at 20 questions and at 200. A rate shows the same 10% both times. A penalty total punishes you for testing more.

Most brand teams want "acquired by a competitor" to count for more than a wrong headcount. That instinct is sound. It just isn't inherited from research.

FActScore states that "every atomic fact has an equal weight of importance." Vectara scores each summary as consistent or not. MQM, a translation-quality framework, is one of the few that weights severity, and it recommends an exponential spread rather than a linear one.

So if you use tiers:

• Define them in writing before you score anything.
• Keep the unweighted rate visible next to the weighted one.
• Don't change the weights mid-series, or your trend line breaks.

For how the most damaging errors arise, see AI hallucinations about brands.

The same question can produce different facts depending on conditions you don't control. Log them, or you'll mistake drift in the setup for drift in accuracy.

• Retrieval state. An answer given without web search comes from training data frozen at a cutoff. OpenAI's model docs (checked August 2026) list a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path. Older cutoffs mean older facts.
• Platform and surface. Google says AI Overviews and AI Mode "may use different models and techniques." Score each surface on its own.
• Personalisation. Memory and custom instructions can change what ChatGPT searches for. Test logged out, or note the setup.
• Date. As of October 2026, engines update models and defaults often. Timestamp every answer.

This is also why platforms differ so much on accuracy. The breakdown in which AI platforms are more accurate about brands covers that comparison.

Unverifiable is not the same as false

If an AI states your employee count and your fact sheet has none, the claim isn't wrong. It's unchecked. FActScore's "not supported" label covers both contradicted and unverifiable claims. Splitting them is a choice you make, and it usually improves the report.

Errors and refusals

Count answers that returned. Report refusals and failures separately.

Small samples

One error in two answers reads as a collapse. Report a platform's rate only when enough answers came back to mean something, and fall back to the overall rate otherwise.

Your own site disagrees with itself

If the pricing page and the footer state different prices, an AI repeating either one is not hallucinating. Fix the source before you score the answer.

Start small and repeatable. Write the fact sheet, pick three to five buyer questions, and score one platform by hand. The free AI hallucination check walks through a manual version. The free checker on this site gives a snapshot of how engines answer up to three buyer questions.

When you outgrow spreadsheets, judge tools against this framework: do they show per-claim evidence, a denominator of returned answers, per-platform results and a logged retrieval state? The criteria in choosing a brand accuracy monitoring platform go further, and alerts for false AI claims covers ongoing monitoring.

Should I combine all platforms into one accuracy score? Not without showing each platform underneath. Engines use different models, indexes and citation formats, and Google says even AI Overviews and AI Mode may use different techniques. A blended number can hide one engine that is badly wrong behind several that are fine. Report per platform first, and treat any combined figure as your own model rather than a measurement.

Does a citation prove the AI used my page correctly? No. OpenAI's API returns the full set of retrieved sources separately from citations, and the retrieved set is often larger. Mechanistic research on how models cite found citation behaviour can be changed without changing the sources. A citation tells you a page was linked. Only checking the claim against your fact sheet tells you whether the answer was right.

Can I stop AI platforms from getting my facts wrong? You can't control what a model says. You can remove blocks: make sure AI crawlers can reach your site, put key facts in the HTML before JavaScript runs, and state specific facts plainly on a fetchable page. Fix contradictions across your own pages. When engines cite other sites, being named accurately there is associated with appearing in answers, with no guaranteed effect.

How often should I re-measure brand accuracy? Re-measure after any change you make to your facts, such as a price change or rebrand, and after major model launches. Keep the questions, platforms and scoring rules fixed between runs so changes reflect the engines, not your method. Timestamp every run, because engines change models and defaults often and an undated score can't be compared with the next one.