Key takeaways
- A mention tells you that AI named you. A narrative tells you what it said. Tracking the individual claims ("expensive", "no mobile app", "good for small teams") is the only way to catch misdescription.
- Answers vary run to run, so claims must be counted across many runs, prompts and models, not read off a single answer.
- Wrong claims usually come from a source, either your own stale pages or a third-party comparison page. Find the source and you can fix the claim.
- ChatGPT leans on product and landing pages for branded prompts, so old pricing copy on your own site is the first place to look.
- You can start manually in an afternoon, but claim-level tracking at scale needs a tool.
Why mentions are the wrong thing to stare at
Most AI visibility advice stops at "are we mentioned?" That is a fine first question and a bad last one. A buyer who asks ChatGPT "is [your brand] worth it for a startup?" doesn't care that you were named. They care that the answer said you cost $199 per seat when you actually dropped to $99, or that you "lack integrations" you shipped a year ago.
The 01net guide on tracking brand mentions makes the same point with a real example: one tracked brand was still described as "expensive" in Gemini answers six months after a price cut, and nobody on the marketing team had noticed because nobody was looking.

I call the set of claims AI repeats about you your brand narrative. It is made of individual statements, each with a frequency, a set of models that say it, and a source it came from. That structure is what makes it trackable.
Mentions, citations and claims are three different things
- A mention is your brand named in an answer.
- A citation is a link to a source the AI used. Promptwatch has a citations vs mentions explainer if you want the distinction in depth.
- A claim is a specific statement about you inside the answer, such as "pricing starts at $99/month".
You can be mentioned with no citations, cited while the answer recommends a competitor, and described wrongly in both cases. Claims are the layer where reputation damage actually happens.
Why you can't just ask ChatGPT what it says about you
It is tempting to open ChatGPT and type "what do you know about [brand]?" The answer feels authoritative. It isn't. As the LLM Pulse guide puts it, ChatGPT has no record of what it said to other people and no view of its own output distribution. You get one sample.

Two more problems make a casual check unreliable:
- Non-determinism. The same prompt can produce different brands, different order and different descriptions on the next run.
- Personalisation. Location, account history, memory and device change what you see. The API also isn't the same as the consumer app: OpenAI's API gives you a model, while the ChatGPT app wraps memory, personalisation and search tooling around it.
So a claim only counts as part of your narrative if it shows up repeatedly. One odd sentence is noise. The same sentence in 18 prompts and 41 responses is a narrative.
Where the claims come from
Before you track anything, it helps to know how a claim gets into an answer. Layer3 Labs breaks ChatGPT's brand knowledge into training data, live search results, cited sources, entity data and conversation memory. For narrative work, two of those matter most.
Live search builds claims from short queries
Promptwatch's ChatGPT query fanout data shows average web searches per response between roughly 1.0 and 2.15 over recent months, with query length dropping from about 117 characters in December to around 53 in April. The short, keyword-style searches mean claims about you are built from whatever ranks for things like "[brand] pricing", "[brand] vs [competitor]" and "[brand] reviews". That is my inference rather than something the report states, but it gives you a sensible set of prompts to test. The query fan-outs guide explains how to see them.
Your own product and landing pages carry a lot of weight
Promptwatch Data on which content types ChatGPT cites (Sep 28 to Oct 4, 2026) found product pages get 22.4% of classified citations and landing pages 16.8%. For branded prompts like "is X good for startups", product pages get 23.3% and landing pages 22.8%, almost half of branded citations together. Reviews rise to 5.6% and social posts to 3.0% for branded prompts.
So what ChatGPT says about you leans on your own pages first. Stale pricing copy or an old feature list there will be repeated back to buyers.
Wrong claims stick around
Promptwatch's content half-life report tracked pages ChatGPT cited on July 6, 2026 for 90 days. After 30 days ChatGPT still cited 78% of landing pages and 76% of docs; after 60 days, 69% and 67%. Comparison pages held at 72% after 30 days. Pricing pages are classified as landing or product pages. News fades faster, reaching half-life at 36 days.
Practically: a wrong claim sourced from a landing page, doc or comparison can persist for months. One sourced from a news story will probably fade on its own.
A manual method you can run this week
This won't scale, but it will show you whether you have a problem. It borrows from the manual routines in the 01net and Frac.tl guides.
- Open an incognito window and log out, in ChatGPT, Gemini and Perplexity.
- Write 10 to 15 prompts a buyer would type when choosing: "is [brand] worth it", "[brand] pricing", "[brand] vs [competitor]", "[brand] alternatives", "best [category] for [use case]", "what are the downsides of [brand]".
- Write each prompt at least three different ways and run each version at least three times. The Searchable guide recommends exactly this to cope with answer variation.
- Paste every answer into a spreadsheet with columns for model, prompt, run, and cited sources.
- Extract the claims. Read each answer and write every statement about you as a short line: "pricing starts at $99", "no free plan", "strong customer support", "limited integrations".
- Group claims that mean the same thing and count how many answers contain each one.
- Mark each claim accurate, inaccurate or unclear against your own site and product.
The limits are obvious. It takes an afternoon, covers one moment in time, and you won't repeat it every week. But even this will probably surface two or three claims you didn't know about.
Keep mentions and citations in separate columns
Searchable's guide warns that ChatGPT might know your brand well and still cite other sites when it talks about you. If an answer describes you accurately but cites a competitor's comparison page, the claim is fine and the source is a risk. Record both.
What to record for each claim
| Field | Why it matters |
|---|---|
| Exact quote | Lets you find the source and prove the claim existed |
| Models that say it | A claim only in Gemini points to a different source than one in ChatGPT |
| Number of prompts and responses | Separates a stray sentence from a repeated narrative |
| Theme (pricing, features, support, reputation) | Lets you score and compare areas |
| Status (accurate, inaccurate, unclear) | Tells you what needs fixing |
| Cited source | The page you will actually edit or contact |
| Date first and last seen | Shows whether a fix worked |
Tracking claims at scale with a tool
Once your prompt set passes a few dozen prompts across several models, the spreadsheet falls apart. Here is how the tool categories split up.
Mention and sentiment trackers
Most AI visibility tools report whether you were mentioned, where, and with what overall sentiment. That is useful, and several do it well: Peec AI is a clean, affordable option, and Otterly.AI covers basic monitoring.

Semrush's AI Visibility Toolkit adds a Questions report showing the exact prompts where you and competitors are named, which gets closer to "what does AI say about us compared with them".
Searchable, LLM Pulse and Profound sit in the same monitoring-first group, with different depth on sources, sub-queries and enterprise reporting.

If you want to browse more of this category, the GEO software directory at bestgeosoftware.com lists many of them side by side.
Claim-level tracking and fixing
This is where Promptwatch is different. Its Brand Narrative feature pulls the individual claims AI makes about you and your competitors out of every tracked response, groups repeated claims, shows how often and which models say each one, and files them under themes like Pricing, Features, Support and Reputation, each with a 0-100 score you can compare against competitors. Brand Narrative is on every plan.

Fact Check, available on Business, Agency Scale and Enterprise plans, takes the next step. You mark a claim Accurate or Inaccurate, and Promptwatch reads the pages crawled from your site to find the ones that state the claim or its opposite, with quotes and a recommended next step. That answers the question you really have: which of my pages is AI learning the wrong thing from?
Because Promptwatch collects answers from the real interfaces of ChatGPT, Gemini, Perplexity, Google AI Overviews and others, rather than only the API, you are looking at what buyers see. Pricing is on promptwatch.com/pricing and there is a free AI visibility report if you want a first look.
Comparison of approaches
| Approach | Catches claims? | Traces sources? | Helps fix them? | Best for |
|---|---|---|---|---|
| Manual spreadsheet | Yes, by hand | Partly, via cited links | No | A first audit |
| Mention trackers (Peec AI, Otterly.AI) | Mostly sentiment level | Citation lists | No | Basic monitoring |
| SEO suites (Semrush, Ahrefs Brand Radar) | Prompt-level mentions | Some | No | Teams already in those suites |
| Promptwatch | Claim-level with themes and scores | Yes, plus crawler logs | Content Agent and Action Board | Teams who want to fix, not just watch |
Tracing a wrong claim to its source
This is the step most guides skip. A claim you can't trace is a claim you can't fix.
Promptwatch's Fact Check page describes an illustrative example (demo data, not a real customer): the claim "Valley pricing starts at $99/month" was repeated across 18 prompts and 41 responses by ChatGPT, Perplexity and Gemini. The brand's own pricing page said $199 per seat per month, and a third-party comparison page was spreading the claim. The origin of the error was neither the model nor the brand's own site. It was an outside page.
That pattern is the main thing to take from this section. Wrong claims come from three places:
- Your own outdated pages. Check pricing, feature lists, FAQs and old blog posts first.
- Third-party comparison and review pages. Common for pricing and feature claims.
- Community content such as Reddit or forum threads, which matter more for reputation claims. Keep in mind that ChatGPT's use of these has shifted a lot: Promptwatch Data shows reddit.com's share of ChatGPT citations fell from about 4% to 0.5% on August 14, 2026, while Google AI Overviews gets 12.2% of its citations from social platforms versus 0.46% for ChatGPT in the October 2026 social citations report. So the same reputation claim may come from different places in different models.
To trace manually, open the cited sources for each answer containing the claim, and search each page for the wording. For a systematic version, a page-level view of what AI cites helps. Promptwatch's Page Tracker and citation reports serve that purpose, and its AI crawler logs show whether bots are even reading your corrected page.
Fixing a wrong claim
If the source is your own page
Rewrite the conflicting sentence, and make the corrected fact easy to quote: a plain sentence with the price or feature stated once, clearly. Don't leave old versions live. Then check that AI crawlers can actually reach the page. Promptwatch's AI crawler logs and the free Agentic Readiness Check are useful here.
In Promptwatch the fix flow is: open the conflicting pages list, review a before/after diff of what is true, use the Content Agent to rewrite the sentences, review, and publish to a connected CMS such as WordPress, Webflow or Framer. The Content Agent is on Business and Enterprise plans.
If the source is a third-party page
Contact the publisher with the correct fact and a link to your canonical page. Be specific and polite, and tell them which sentence to change. Where outreach is the job, tools like Respona can help manage it.
If the page is a review or comparison you can't change, publish your own authoritative comparison or FAQ that states the true fact so the model has a better page to cite. Don't expect speed. The half-life data above suggests comparison pages stay cited for a long time.
If it's a reputation claim
These are slower. Address the underlying issue (support response time, a bug, a missing feature), then create current, credible content and reviews that reflect the change. You can't edit your way out of a claim that is still partly true.
Setting up a repeatable narrative audit
The Clutch guide suggests starting with an audit you can repeat: save what the assistant says, separate answer mentions from sponsored placements, and verify the claims that could affect a customer's decision. Here is a simple cadence.
- Weekly: rerun your core prompts, compare claim counts, check for new claims.
- After any change: after a price change, launch or rebrand, check within a week that new claims are emerging and old ones falling.
- Monthly: review themes (pricing, features, support, reputation) against two or three competitors.
- Quarterly: refresh the prompt set, since buyer questions change.
Track inquiries or conversions that customers attribute to AI separately from your sampled results, as Clutch advises. Visitor analytics that attribute AI traffic help close that loop. One note: ChatGPT stopped sending a referrer on October 8, 2026, so tagging links with utm_source=chatgpt.com is now worth doing. The attributing AI traffic guide covers it.
Common mistakes
- Checking once. One answer is one sample.
- Using only the API. API answers and app answers aren't interchangeable.
- Tracking only the sentiment number. A score of 60 doesn't tell you which sentence to fix.
- Treating all models alike. A claim in Gemini and not ChatGPT has a different source.
- Fixing the model instead of the source. You can't edit ChatGPT. You can edit pages.
- Forgetting persona and location. Buyers phrase things differently, and answers differ by country.
A short example of the whole loop
Suppose you run a project management tool and your manual audit shows that "no free plan" appears in 14 of 40 ChatGPT answers. You have had a free plan for a year. You check the cited sources and find a third-party comparison page from two years ago plus your own old pricing FAQ. You update the FAQ, email the publisher, publish a plain "free plan" section on your pricing page, and rerun the same prompts in three weeks. If the count falls, the fix worked. If it doesn't, you look at the crawler logs to see whether bots reached the new page.
That sequence is dull, which is the point. Narrative tracking is bookkeeping with a clear target.
Where to go next
If you want a broader view of the tools available, the GEO software directory is a good starting point, and Promptwatch Data has more reports on what AI models cite and how. If you want help setting up an AI search program rather than doing it yourself, 1001 SEO Media runs GEO and AI search work for brands. For a first look at your own narrative, try the Promptwatch free trial or run the manual method above this week.



