Key takeaways
- AI perception is what ChatGPT, Gemini, Perplexity and Claude say about you. Brand perception is what real customers believe. They overlap, but they are built from different inputs.
- AI answers are stitched together from your own pages, reviews, comparisons and third-party content, so outdated or thin sources produce outdated or thin descriptions.
- Promptwatch Data shows that for branded prompts, product pages and landing pages make up almost half of ChatGPT's citations. Your own site is the biggest lever you have.
- Audit with real buyer questions, repeated many times, across several engines. One screenshot of one answer proves nothing.
- Fix the sources, not the symptom. Wrong claims persist for as long as the pages feeding them stay cited.
Two perceptions, one buying decision
A prospect used to form an opinion of you from your homepage, a few reviews, maybe a colleague's offhand comment. Now a growing share of them type "is [your brand] any good for a team of 20?" into ChatGPT and read the paragraph that comes back.
That paragraph is AI perception. It might say you are "reliable but expensive." It might say you are "best for enterprises" when half your revenue comes from solo founders. It might not mention you at all.
Brand perception is different. It lives in the heads of people who have used your product, seen your ads, talked to your support team. It is shaped by experience. AI perception is shaped by documents.
That is the whole problem in one sentence. A model has never used your product. It has read about it.
Why the two diverge
Models read sources, customers live the product
An LLM's view of your brand is an aggregation: press articles, review sites, forums, comparison posts, your own pages, LinkedIn posts. The vurto.ai write-up on brand perception in AI puts it well: the model synthesizes all of that into one consolidated view, and that view can drift a long way from your official positioning.
Your happiest customers rarely write anything. Your angriest one wrote a 2,000-word forum post in 2023. Guess which one the model has read.
Time lag and partial memory
What a model knows depends on its training data and on what it finds in live search. A pricing change from last quarter may not be reflected. A bad event from two years ago can still show up if the pages covering it are well indexed.
Promptwatch's content half-life report makes this concrete. It tracked pages ChatGPT cited on July 6, 2026 for 90 days. After 60 days, ChatGPT still cited 69% of landing pages and 67% of docs. Product pages, reviews and how-tos fell to about 55% at day 60, with a half-life of roughly 74 to 77 days. News was the exception: half of it was gone after 36 days.
Translation: an inaccurate pricing page or a stale comparison does not quietly disappear. It keeps getting cited for months.
Each engine is built differently
ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews do not share sources or answer architecture. Your brand can read as positive in one and flat in another. Refractia, one of the vendors in this space, separates chat engines that answer from what the model knows from Google AI Overviews, which summarizes what ranks for a search. The fix differs depending on which one is misdescribing you.
Promptwatch's social media citations data (Oct 11 to Nov 9, 2026) shows how different they are. Social platforms were 13.16% of Google AI Overviews citations but only 0.64% of ChatGPT's. YouTube alone was 6.71% of AI Overviews citations. So ChatGPT's description of you leans on your own pages, listicles, guides and comparisons, while AI Overviews leans much more on video and social chatter.
Synthetic authority
A Trustpilot review reads as one person's opinion. An AI answer reads as a neutral summary. People trust it more than it deserves, and it rarely shows its uncertainty. That is the part I find genuinely uncomfortable: the model can be confidently wrong about your refund policy and the reader has no way to tell.
What actually feeds ChatGPT's description of you
Promptwatch Data on which content types ChatGPT cites (Sep 28 to Oct 4, 2026) is the most useful starting point I have seen for branded questions.
| Content type | Share of all ChatGPT citations | Share for branded prompts |
|---|---|---|
| Product pages | 22.4% | 23.3% |
| Landing pages | 16.8% | 22.8% |
| Listicles | 13.6% | n/a |
| How-tos | 8.7% | n/a |
| Documentation | 7.5% | n/a |
| Reviews | 1.7% | 5.6% |
| Social posts | under 1% | 3.0% |
| FAQ pages | 3.3% | 4.8% |
For branded prompts, such as "is HubSpot good for startups," product and landing pages are almost half of citations. Reviews climb to more than three times their usual share. Social posts climb almost four times.
Two things follow. First, you own more of your AI perception than you think, because your own pages are the biggest single input. Second, the third-party pieces that do matter on branded queries are reviews, so which review sites you show up on is not a side project.
One more detail: almost half of ChatGPT responses cite ten or more unique domains, per Promptwatch's unique domains per response report. Your description is a blend, and a single outlier source can tilt it.
The four dimensions worth measuring
Most guides on this topic, including those from Sight, Moistur and Brantial, converge on similar layers. Here is the version I would actually use.
Sentiment
Is the tone positive, mixed or negative? This is the easy number to put in a slide. Some vendors score it 0 to 100. Refractia, for example, treats 70+ as positive, 40 to 69 as mixed and below 40 as negative. That is the vendor's own banding, not an industry standard, so treat it as a convention rather than a fact.
Attributes
Which adjectives keep showing up? "Reliable," "premium," "hard to onboard." Repeated negative descriptors matter most, because that is what a buyer reads right before choosing.
Use cases and audience
Who does the model recommend you to? Startups, enterprises, beginners? Models often hold fixed ideas here that lag your commercial reality. If you moved upmarket a year ago and the model still calls you a "tool for freelancers," you are losing enterprise shortlists without ever seeing it.
Competitive framing
When you are compared head to head, do you win, tie or lose, and on which criteria? Losses on traits your customers care about go to the top of the work list.
Promptwatch's Brand Narrative takes a slightly different route: it pulls individual claims out of every tracked response ("affordable for small teams," "mobile app lacks features"), groups repeats, and scores themes like pricing, features, support and reputation against competitors. Claim-level detail is more actionable than a single sentiment number.
How to audit the gap, step by step
1. Write the questions your buyers actually ask
Skip "what is [brand]?" Use the ones that precede a purchase:
- "[Brand] pros and cons"
- "Is [brand] good for [your best-fit segment]?"
- "[Brand] vs [top competitor]"
- "Best [category] for [use case]"
- "[Brand] pricing"
- "Is [brand] worth it?"
A Query Fan-Out Generator can help you see the sub-searches AI runs behind these, and Promptwatch has a free ChatGPT query fan-out generator for that.
2. Run each question many times, on several engines
Answers vary run to run. Asking once is anecdote. Refractia's own guidance says the same: ask the questions your customers ask, many times, and read the answers together. Cover ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews at minimum.
3. Record the claims, not the vibe
Copy the specific statements into a sheet. "No free plan." "Limited integrations." "Best for enterprise." Then mark each one accurate, inaccurate or outdated.
4. Trace each wrong claim to a source
This is the step most teams skip. A wrong claim lives somewhere. It might be your own old pricing page, a comparison article from 2024, a stale G2 profile or a forum thread. Check the citations the answer shows. Fixing the model's wording is impossible. Fixing the page it learned from is not.
5. Compare against what customers actually say
Now the other half of the gap. Pull your support tickets, NPS verbatims, win/loss notes and recent reviews. Tools like [tool:qualtrics] and [tool:trustpilot] data are useful here. Where customers praise something AI never mentions, you have an invisibility problem. Where AI praises something customers complain about, you have an accuracy problem, and that one is more dangerous.
Tools for the monitoring side
You can do a manual audit with a spreadsheet for one afternoon. Doing it weekly across five engines is where tools earn their keep.
Promptwatch is the one I would start with if you want to go from diagnosis to fixes. It monitors the real interfaces of ChatGPT, Gemini, Perplexity, Google AI Overviews, Copilot and others, scores visibility and sentiment, extracts the claims AI makes about you through Brand Narrative, and on Business plans lets you mark claims accurate or inaccurate with Fact Check so you can see which of your pages state the wrong thing. Its crawler logs show whether AI bots can even read the page you just fixed.

Others worth knowing, depending on your needs:
| Tool | Angle | Best for |
|---|---|---|
| Promptwatch | Monitoring plus content fixes, crawler logs, Brand Narrative, Fact Check | Teams that want to close the gap, not just see it |
| Profound | Enterprise AI visibility tracking | Large brands with big monitoring budgets |
| Otterly.AI | Lightweight prompt and mention tracking | Small teams starting out |
| Peec AI | Visibility tracking for marketing teams | Simple multi-engine reporting |
| LLMClicks | Tracking with hallucination detection | Spotting factually wrong claims |
| Brand24 | Social and web mention monitoring | The traditional-perception side of the comparison |
Profound

Otterly.AI

If you want to compare more options, the GEO software directory at bestgeosoftware.com lists a broader set. Traditional listening tools like [tool:brandwatch] cover the human-perception half, and AI visibility tools cover the model half. You need both views side by side to see the gap.
Fixing the gap: what to change first
Make your own pages answer the buying questions
Given that product and landing pages are close to half of branded citations, your homepage, product pages, pricing page and docs should state plainly what you cost, what you do, and who you are for. If a model cannot find "who it's for" on your site, it will guess from someone else's page.
Keep long-lived pages accurate
Landing, docs and pricing pages stay cited for a long time, so an error there is expensive. Treat them like long-term assets with an owner and a review date. Use press releases and news posts for short bursts only, since news loses half its pages within 36 days.
Collect reviews where your category looks
Branded prompts cite reviews at more than three times the usual rate. Get fresh, honest reviews on the sites buyers in your category trust. Old, thin review profiles produce old, thin descriptions.
Add a plain-language FAQ about your brand
FAQ pages are 4.8% of branded citations. A page that answers "how much does it cost," "who is it for," "what does it not do" in plain language is cheap to write and easy for a model to quote.
Check that AI crawlers can actually read you
A perfect page that GPTBot cannot fetch fixes nothing. Run a crawler access check. Promptwatch's Agentic Readiness Check and its AI crawler logs will tell you which bots reach which pages and where they fail.
Fix the third-party sources you traced in the audit
Email the author of the stale comparison. Update your G2 profile. Correct the Wikipedia-adjacent summary pages you can legitimately edit. This is slow, unglamorous work, and it is where most of the wrong claims actually get removed.
A note on the other kind of perception gap
There is a second gap worth mentioning, and it runs in the opposite direction. Research published by the NIM on consumer attitudes toward AI-generated marketing found that people reacted less favorably to ads labeled as AI-made, even when the content was identical to human-made versions. Only about 28% of participants understood how AI uses their personal data for personalization.
That is about AI-made content, not AI describing brands, so do not over-read it. But it is a reminder that customer trust in AI-shaped material is shaky, and a buyer who sees a vague or oddly negative AI summary of you may treat it with some suspicion too. Do not count on that to save you. The safer plan is to make the summary accurate.
What to do this week
- Pick ten buyer-intent questions and run each five times across ChatGPT, Gemini, Perplexity and AI Overviews.
- List every claim made about you and mark it accurate, wrong or outdated.
- Trace the three most damaging wrong claims to their source pages.
- Update your own pricing, product and FAQ pages first, since they are the fastest to change.
- Set up monitoring so you see changes weekly, not when a customer mentions it on a sales call.
For a no-commitment starting point, Promptwatch's free AI visibility report scans how ChatGPT, Gemini and Perplexity describe your brand without a login, and there is a 7-day free trial if you want continuous tracking. Check promptwatch.com/pricing for current plans.
The gap between AI perception and brand perception is not mysterious. It comes from what the model read versus what your customers lived. Close it by fixing what the model reads.


