How we audit what AI says about a client's brand

Most clients first notice the problem through a screenshot. Someone in sales pastes an AI answer into Slack: it says the product is "too expensive for small teams" or that "its mobile app lacks features." The founder wants to know if that is true, where the model got it, and what to do about it. A sentiment score cannot answer any of those questions. It tells you the mood, not the sentence.

So we audit claim by claim. This is the process we use on client programs, with the manual version first because it works without any tool and it teaches you what to look for.

Step one: collect the sentences, not the score

Start with a prompt set that reflects how buyers research the category: comparisons, "best for" questions, pricing questions, and a few direct brand questions. Run each on the engines the client cares about. Copy out every sentence that makes a claim about the client or a named competitor.

Then group the claims by theme. The groups that keep showing up are pricing, features, support, and reputation. Within a theme, merge duplicates. If six answers say a version of "expensive for small teams," that is one claim with a frequency of six, not six findings.

Record three things for each claim: the exact wording, which engines said it, and how often. Frequency is what separates an annoying outlier from something that is shaping buyer perception. One model saying something odd once is a note. Four models repeating the same line across twenty prompts is a work item.

Step two: decide whether the claim is true

This is where the client has to be in the room. For each repeated claim, someone with product knowledge marks it accurate or inaccurate. We keep a third state, "partly true," because the "expensive for small teams" claim is often true for one plan and false for another.

We do not argue with the model at this stage. We only decide what the truth is, in one sentence, so the next step has a target.

Step three: look for the source of the claim on the client's own site

This step gets skipped, and it is the one that most often pays off. Models do not invent claims in a vacuum. They read pages. If an AI answer says a client lacks a feature, we check whether the client's own site says so, or says the opposite on a different page, or says nothing at all.

The usual findings are boring and fixable:

• An old pricing page or blog post still states a retired price or plan.
• Two pages on the site contradict each other, for example a features table and a help article.
• The truth is true, but it only appears inside a PDF or behind a script, so crawlers never see it.
• The claim does not appear on the client's site at all, which means the source is somewhere else, such as a review site, a forum thread, or a competitor comparison page.

That fourth case needs offsite work, and it is slower. The first three are usually a one-day content fix.

Step four: choose the smallest fix that changes the evidence

We prioritise by frequency multiplied by how costly the claim is to the sale. A false pricing claim that appears in most comparison answers ranks above a vague reputation claim that appears twice. For each item we write one action: correct a page, remove a stale page, add a plain statement of the fact near the top of the relevant page, or request a correction on a third-party page.

We then re-run the same prompts after the fix has had time to be crawled, and compare the claim's frequency. We do not promise the sentence will vanish. Models mix sources, and a corrected page is one input among many. We report the change in frequency and leave it at that.

Where Promptwatch does the heavy lifting

Doing steps one to three by hand across a prompt set of any size is a day or two of copying and pasting. Promptwatch is the platform we run client programs on, and two pages under Brand perception map to this process directly, both introduced in the changelog entry for October 27, 2026.

Brand Narrative pulls the individual claims AI makes about the brand and its competitors out of every tracked response. Repeated claims are grouped, so you see how often each one appears 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 that you can compare against competitors, and the platform also shows sentiment per competitor, which is useful for spotting where AI talks about a rival more favourably. Brand Narrative is on every plan.

Fact Check covers step three. You mark a claim Accurate or Inaccurate, and Promptwatch reads the pages it has crawled from the client's site to find the ones that state the claim or its opposite. Each check lists the pages and quotes it found, plus a recommended next step. Fact Check is available on Business, Agency Scale, and Enterprise, each check uses agent credits, and you see the estimate before it runs. That tier detail matters for planning: on a lower brand plan you can still run the claim analysis, but the page-level check is a manual job.

How we use it in an engagement

In the first month of a program we run Brand Narrative on the client's tracked prompts and review the top claims with the client in a single call. We mark the true and false ones together, then send the false ones through Fact Check where the plan allows it. The output feeds our action list directly: each contradictory page becomes a ticket with the quote attached.

We also keep a record of claim frequency by month. It is a modest chart, but it is the most direct evidence we have that the client's perception in AI answers is moving, and it avoids arguing about a single screenshot.

Limits worth stating to the client

An audit like this describes what tracked answers said on the days we ran them. It does not describe everything every user saw, and it does not prove why a model said something. Fact Check shows which crawled pages state a claim, which is strong evidence of one possible source, not a complete causal map. We say that out loud so a client does not expect a corrected page to flip an answer overnight.

If a client asks where to start, the short version is this: pull the claims, mark them true or false, find the contradicting pages, fix the cheapest ones first, and measure frequency again. The tooling just makes that loop fast enough to repeat every month.