How accurate is AI about your brand? Scoring accuracy per model
AI engines state claims about your brand all day. Here is how to turn that into one accuracy score, split by model and theme, and what to do with the weak spots.
Picture a head of marketing reading an AI answer about their company. It gets the pricing model wrong, says a feature does not exist, and describes a product the company retired two years ago. They fix the website. A week later, a different engine says something else that is wrong. Where do they even start?
Visibility numbers do not help here. Being mentioned is one thing. Being described correctly is another. Accuracy needs its own measurement, and since November 4, 2026, Promptwatch reports it as a score.
From claims to a score
The chain has two links: Brand Narrative and Fact Check.
Brand Narrative pulls the individual claims AI makes about you and your competitors out of tracked responses. It groups repeats, noting how often each claim appears and in which models. It files them under themes such as Pricing, Features, Support, and Reputation, scores each claim positive, neutral, negative, or mixed, and scores each theme from 0 to 100. Brand Narrative is on every plan.
Fact Check takes it further. You mark a claim Accurate or Inaccurate, and it reads pages crawled from your site for ones that state the claim or the opposite, returning quotes and a next step. Fact Check uses Agent Credits and shows an estimate first. It is available on Business, Agency Scale, and Enterprise.
The new accuracy score sits on top. It is the share of your brand's checked mentions that are accurate, shown as one score and as a heatmap per model and theme.
What the score counts
The rules are worth knowing, because they decide how far you can trust a number.
- A claim you marked counts as you marked it. Your judgment wins.
- Otherwise, the latest completed fact check decides.
- Inconclusive claims are left out.
- Unchecked claims are left out.
That last pair matters. A score of 90% on a handful of checked claims is a different animal from 90% on hundreds. Before you quote the score, look at how many claims it rests on. Because unchecked claims drop out, running more checks changes what the score covers, not only what it says. If the number jumps after a round of checking, the denominator may have changed.
Reading the heatmap
One score hides the interesting part. The heatmap splits accuracy by model on one axis and theme on the other.
| What you see | A reasonable first reading |
|---|---|
| One model is low across every theme | That model is working from an outdated or thin source set |
| One theme is low across every model | A common source probably carries the wrong claim |
| One cell is low, others fine | A single claim worth tracing |
The second pattern is the most useful. If Pricing is weak everywhere, the likely culprit is a widely cited page with an old figure. You can trace it from the claim back to the responses and citations behind it. Brand Narrative lists the best-scoring theme first and puts the most negative claims at the top, so the worst problems sit where you will see them first.
Keep in mind that models differ in how they collect information. Promptwatch tracks some through real-interface monitoring and others through API, which answers from training data. A low score on an API-tracked model points to what that model learned, which you cannot edit directly. A low score on a live-search model is more likely to trace back to pages it fetched.
What to do with a weak spot
Work in order.
- Open the lowest cell and read the actual claims.
- Mark them. Your marks override the automatic check.
- Use Fact Check's quotes to see whether your own site states the claim clearly. Sometimes the wrong claim exists because your site never says the right one in plain words.
- Trace the claim through the citations to the third-party page that repeats it.
- Fix what you control, and watch the same cell after the next runs.
Do not promise yourself that fixing a page will change an engine's answer. It might, and you can watch whether it does. What the score gives you is a before and after on the same measure.
Where we land
Most trackers in this directory tell you whether you were mentioned. They stop short of asking whether what was said was true. Promptwatch is where we would start for accuracy work because the claim extraction, the fact check against your own crawled pages, and the per-model heatmap all live together. If you are on Essential or Professional, Brand Narrative still works; Fact Check and the accuracy score need Business or above, so factor that into the plan choice. Business is $579 a month, with a 7-day free trial on monthly billing for the self-serve paid plans.
Next step
Book a demo at promptwatch.com/book-a-demo and ask to see the accuracy heatmap on your own brand. It is the fastest way to find out whether the weak cell is a pricing claim or a retired product.