How to see what AI claims about your brand, claim by claim

A sentiment number is a weather report. It says it is gloomy out and nothing about why. When ChatGPT describes your product as "too expensive for small teams" or says "its mobile app lacks features", the score might read 62 out of 100 and tell you nothing you can act on. The claim is what a buyer reads, and the claim is what you can change.

This guide covers how to see what AI says about you one claim at a time, how to check whether your own site is feeding the wrong ones, and which tool does it with the least manual work.

Why one score hides the problem

Sentiment tools score a whole response. That is fine for trend lines. It fails in two common cases.

First, a response can be warm overall and still contain one damaging sentence. "A solid choice with a clean interface, though pricing is steep for small teams" scores as mostly positive. The buyer with a small budget stops at the last clause.

Second, the same claim repeats across dozens of answers and models. A single score smears it into the average. You never see that fourteen different responses say the same thing about support.

What you want is the claim as the unit: the sentence, how often it appears, which engines say it, and whether it is true.

The manual method

You can do this without any software for a small prompt set. It is slow but it teaches you what to look for.

Collect responses. Run 20 to 40 brand and comparison prompts ("is [brand] worth it", "[brand] vs [competitor]", "who is [brand] best for") in the engines your buyers use. Save the full text.
Pull out statements. Copy every checkable assertion into a sheet. One row per claim.
Group by theme. Pricing, features, support and reputation cover most of it. Merge near-duplicates so "affordable for startups" and "good value for small teams" become one row.
Count and mark. Note how many responses say it and in which engine. Then mark it true, false or outdated against your own source of truth.
Search your site for the claim and its opposite. If AI says you lack a feature, look for a page that says you do. More often than people expect, an old comparison page, a stale FAQ or a help article still says the opposite of what you sell today.

Step five is the one that pays. AI engines fetch your pages. If one of them states the wrong thing, correcting it is faster than any outreach campaign. Our guide on how to fact-check what AI says about your brand walks the wider process, including third-party sources.

The trouble is scale. Twenty prompts across five engines is a hundred responses. Run it weekly and the spreadsheet dies in a month.

What to look for in a tool

A tool worth paying for in this job should do four things:

Extract claims from stored responses rather than only scoring them.
Group repeats so you see frequency, not just examples.
Score by theme so you can compare yourself with competitors on pricing perception or reputation.
Connect a claim to the pages on your own site that say it.

The fourth is the rare one. Most tools stop at "here is what AI says". Few tell you "here is the page on your site that says the opposite".

Where Promptwatch fits

Promptwatch added this in October. According to the changelog entry for Fact Check and Brand Narrative, two pages now sit under Brand perception.

Brand Narrative pulls the individual claims AI makes about you and your competitors out of every tracked response. 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 0 to 100 score you can compare against competitors, and sentiment is shown per competitor. Brand Narrative is on every plan, so you can start there.

Fact Check is the step that closes the loop. You mark a claim Accurate or Inaccurate, and Fact Check reads the pages crawled from your site to find the ones that state it or the opposite. Each check lists the pages and quotes it found, plus a recommended next step. This runs on Business, Agency Scale and Enterprise, each check uses agent credits, and you see the estimate before it runs. So it is not free to run casually, but you control the spend.

That tier split is honest advice to give: Brand Narrative tells you what is being said on any plan. Proving that your own pages are the cause needs Business at $579 a month, or Agency Scale at $799 for agencies. If your budget sits lower, use Brand Narrative to find the claim and do step five by hand with a site search. That is slower but entirely workable.

Because the claims come from the same tracked responses that feed visibility, citation and crawler data, you can also check the other direction. If a wrong claim shows up in answers that cite a third-party page, Fact Check will not find it on your site, and that tells you the fix is outside it. Pair this with your Brand Book so the tool agrees with you on what counts as a mention.

A weekly routine that takes 30 minutes

Open Brand Narrative and sort themes by score against your top competitor.
Pick the lowest theme and read the three most repeated claims.
Mark each true or false against your source of truth.
For false or outdated claims, run Fact Check if you have the plan, otherwise search your own site for the claim and its opposite.
Fix the page, add the clear statement where it was missing, and note the date.
Next week, see whether the claim's frequency dropped.

That last step is the one to be careful about. A changed page does not change an answer overnight, because engines refetch on their own schedule. Treat frequency over several weeks as the signal, not a single run.

For a wider comparison of products that analyse brand mentions, see our roundup of tools that analyze how LLMs talk about your company. I would start with Promptwatch because it is the one that goes from the claim to the page on your site. The others in the catalog are documented as monitoring or action tools, and we have no sourced evidence that they extract claims at this level, so treat that as unknown rather than absent.