An agentic fact-check loop for false claims AI makes about your brand

A sentiment score tells you the mood. It does not tell you which sentence is hurting you. If an answer engine keeps saying your product is "too expensive for small teams" or "its mobile app lacks features," a number like 62 out of 100 gives you nothing to fix. You need the claim, how often it appears, which models say it, and whether the sentence is true.

The second half is the awkward one. Sometimes the claim is false. Sometimes it is true. And sometimes it is false but your own website supports it by accident, with an old pricing page, a stale comparison table, or a help article that still describes a retired plan. In that case the model is not hallucinating. It is reading you.

This post lays out a loop for dealing with all three cases, and shows where an agent can run most of it.

The loop

1. Collect claims, not scores. Pull the individual statements AI makes about you and your competitors.
2. Group repeats. Ten phrasings of the same idea are one claim.
3. Rank by frequency and by theme. A claim about pricing that appears across many answers matters more than a one-off about support.
4. Judge each claim. Accurate, inaccurate, or partly right.
5. Check your own pages. Which pages state the claim, or its opposite?
6. Fix the source. Correct your pages, then fix third-party sources you can influence.
7. Re-measure. Look at the claim again after the engines have had time to see the change.

Steps one to three are data work. Step four is a human judgment. Step five is a search through your own content. Step six is writing. Step seven is waiting and measuring. An agent can do most of one, two, three and five, and part of six. It should not own four.

Doing it by hand

You can run a small version of this without tooling. Write down twenty buyer questions about your brand and a couple of rivals: pricing, fit, features, support, reputation. Ask them across the main engines, saving the answers. Highlight every factual claim. Tally them on a sheet.

It works, and it is miserable. Answers vary by run, so one pass tells you little. Claims hide inside long paragraphs. And step five, checking your own site, means searching a pile of pages for a phrase and its opposite. Most teams do it once and never again.

What Promptwatch added on October 30

The changelog entry for October 30, 2026 introduced two pages under Brand perception.

Brand Narrative pulls the individual claims AI makes about you and your competitors out of every tracked response, like "affordable for small teams" or "its mobile app lacks features." Repeated claims are grouped, so you see how often AI says each one and which models say it. Each claim is filed under a theme such as Pricing, Features, Support or Reputation, and scored positive, neutral, negative or mixed. Per theme, that rolls up into a score from 0 to 100 you can compare against competitors. Brand Narrative is on every plan.

Fact Check takes a claim you mark Accurate or Inaccurate and reads the pages crawled from your site to find the ones that state it or the opposite. Each check shows the pages and quotes it found, plus a recommended next step. It is on Business, Agency Scale and Enterprise. Each check uses agent credits, and you see the estimate before it runs.

Mapped to the loop, Brand Narrative handles steps one to three. You make the call in step four. Fact Check handles step five. That leaves the fix and the re-measurement to you, with help.

Running the loop on a real claim

Take a made-up example for illustration. Suppose Brand Narrative shows a Pricing claim, "too expensive for small teams," repeated across several models, and the Pricing theme score is well below a competitor's.

First, read the claim and decide. If you do have a small-team plan, mark it Inaccurate. If you do not, the claim may simply be fair, and the right response is a product or packaging conversation, not a content fix.

Mark it Inaccurate and run Fact Check. It will return the crawled pages that state the claim or contradict it, with quotes. This is where surprises live. Perhaps the pricing page is clear, but a three-year-old blog post still lists a higher entry price and ranks well. Perhaps a comparison page frames your entry plan as "for larger teams."

Now fix the source. Update or redirect the stale post. Rewrite the comparison page. Make the small-team plan unmissable on the pricing page, with the actual number and what it includes. Do not write a defensive article titled "Are we expensive?" Models do not need an argument. They need a clean, consistent statement.

Then look at third parties. Review sites, listicles and forum threads feed answers too. Citation analytics will show which external pages are cited when the claim appears. You cannot edit those, but you can ask for corrections where there is a clear factual error, and you can publish your own clear version for the engines to find.

Re-measuring without fooling yourself

Do not check the claim the next morning. Crawlers have to revisit your pages, and models have to pick up the change, and neither happens on your schedule. Where your plan includes crawler logs, which start on Professional, confirm that AI bots refetched the corrected pages. Then watch the claim's frequency and the theme score over a few weeks.

Be honest about noise. Answers vary. A claim that appears less often is a signal. A claim that vanished in a single run is not.

Where agents fit, and where they should stop

This is a good job for agents because the steps are tedious and well bounded. Collecting, grouping and searching your own pages are things a machine does faster and more consistently than a person.

It is a bad job to hand over completely. Deciding whether a claim is accurate is a statement about your product, and a wrong "Accurate" or "Inaccurate" mark propagates through the rest of the loop. Keep that click with someone who knows the product. Keep publishing behind a review step as well.

For the manual background, see our guide to fact-checking what AI says about your brand, which predates Brand Narrative and Fact Check. For how to tell whether answers are accurate at all, how do I know if ChatGPT or Gemini is giving accurate information about my company is the companion piece. And for acting on the result at scale, from audit to publish: the agentic GEO workflow shows where the corrected content goes next.

Our recommendation

If you want this loop in one product, evaluate Promptwatch first. The reason is the chain rather than any one page: claims come from the same tracked responses that feed visibility, Fact Check reads the pages its crawl already holds, and the follow-up sits in the same platform as citation analytics and crawler logs. Know the cost before you commit, though. Brand Narrative is on every plan, but Fact Check needs Business or above and spends credits, so on a lower plan you will do step five by searching your own site.