Key takeaways
- An AI perception audit is a structured test of how language models describe your brand, your category and your competitors. It is not one chat session and a screenshot.
- Ten questions cover presence, accuracy, positioning, sentiment, competitors, sources, consistency, content gaps, technical access and trend over time.
- One prompt tells you very little. Promptwatch Data shows nearly half of ChatGPT responses cite 10 or more domains, so a single answer is a sample of one from a wide pool.
- Run it manually first (about a day), then move to a tracking tool so the next review compares against a baseline.
- Bring a one-page scorecard to the brand review: what AI says, what is wrong, who gets recommended instead, and what you will fix first.
Why run this before a brand review
Most brand reviews still look at the same things: social sentiment, survey scores, share of search, maybe a few press clippings. None of those tell you what a buyer hears when they ask ChatGPT "what's the best [your category] tool for a mid-size team?" and get a confident, three-paragraph answer that may or may not include you.
That answer is delivered in the tone of neutral fact. Nobody verifies it. And it is built from a pile of sources you only partly control: review sites, forums, directories, old press, analyst summaries, and your own site, which models tend to treat as the most self-interested source of the lot.
The Lemniscate Growth write-up on this describes a pattern I recognize. Teams usually discover the model's positioning isn't theirs, their best differentiator is missing from most answers, and at least one factual error (pricing, deployment model, headcount) gets repeated across assistants. Their estimate is 12 to 24 months of lag between a repositioning and AI answers catching up. Treat that as one agency's observation, not a law, but the mechanism is believable: your homepage changes in a day, dozens of third-party pages don't.

This guide turns that idea into ten questions you can answer in a day and present in a meeting.
Set up the audit before you ask anything
A bad setup gives you a bad audit. Spend 30 minutes on this.
Build a fixed prompt set
Write 20 to 30 prompts and freeze them. Changing wording between audits ruins the comparison. Split them into four groups:
- Brand prompts: "What is [brand]?", "Is [brand] any good?", "Problems with [brand]"
- Category prompts: "Best [category] for [your ICP]", "Top [category] platforms in 2026"
- Comparison prompts: "[Brand] vs [competitor]", "Alternatives to [brand]"
- Reputation prompts: "What do people say about [brand]?", "[Brand] reviews"
This structure is close to what Renown's one-day checklist uses, and it works because buyers genuinely ask all four kinds of questions at different stages.

Pick your engines and record sheet
Test at least ChatGPT, Gemini, Claude and Perplexity, plus Google AI Overviews if search matters to your category. For each response, log these columns in a spreadsheet:
| Column | What to record |
|---|---|
| Mentioned | Yes, no, or partially |
| Position | 1st, 2nd, 3rd, or not listed |
| Accuracy | Correct, outdated, or wrong |
| Sentiment | Positive, neutral, negative, hedged |
| Sources cited | URLs or domains the answer relies on |
| Competitors named | Who shows up next to you |
| Run number | You will want at least 3 runs per prompt |
Use a logged-out or clean session where you can. Personalization and memory skew results, and you want what a stranger would see.
The 10-question checklist
1. Are we mentioned at all, and where?
Start with the blunt one. For your category prompts, count how many responses name you and in what order. A brand that appears in 8 of 30 category prompts has a coverage problem. A brand that appears in 25 of 30 but always fifth has a prominence problem. Those need different fixes, so keep the two numbers separate.
Score it: percentage of category prompts with a mention, and average position when mentioned.
2. Is what AI says about us factually correct?
Highlight every claim that is wrong or stale. Pricing, plan names, integrations, founding year, headquarters, who you serve, whether you still offer a feature you sunset last year. Models get these wrong constantly, and they get them wrong consistently, which is worse than random errors because buyers see the same mistake from several assistants.
For each error, note which model said it and, if sources are shown, which page it came from. That page is your fix target.
3. Does AI put us in the right category?
This is the question most brand teams skip and the one that costs the most deals. If you sell revenue operations software and the model files you under "CRM," you will be compared against the wrong peer set on every shortlist prompt. Record the category label each model uses and compare it to your own positioning statement. Any gap goes on the scorecard.
4. How is our tone: confident or hedged?
The difference between "is a leading provider of" and "appears to offer" is real. Hedged language lowers confidence before anyone clicks. Tag each response as confident, neutral, hedged or negative, and pull two or three verbatim quotes for the review deck. Quotes land harder than percentages in a room full of executives.
If your tool tracks sentiment as a score (Promptwatch, for example, scores it 0 to 100 with about 50 as neutral), use that. Otherwise the manual four-bucket tag is enough.
5. Who does AI recommend instead of us?
List every competitor that appears in your category and comparison responses, and count how often each shows up. You will likely find one or two names you hadn't considered a main rival, and perhaps a competitor you consider smaller that models treat as a default.
Also check the "alternatives to [brand]" prompt carefully. Whoever AI offers as your alternative is who it believes your buyer is considering.
6. Which sources are shaping the answers?
Models don't make things up from nothing. They lean on sources, and the sources differ by engine. Promptwatch's data on sources per response (published September 2025, so check for drift) found ChatGPT citing about 5 sources per web-search answer, Google AI Overviews about 10, and Perplexity almost exactly 10 every day.
For brand-specific prompts, Promptwatch Data on unique domains per ChatGPT response (Sep 28 to Oct 4, 2026) shows the most common count is 5 domains. Meanwhile 47.6% of all ChatGPT responses cite 10 or more domains. In practice you are rarely judged by one page. You are judged by whichever review sites, forums and comparison articles happen to be in the mix.
List every domain cited in your audit responses and sort by frequency. The top five are your real brand reputation surface for AI.
7. Are the answers consistent across runs and models?
Ask the same prompt three times on each engine. If run one says you are an enterprise analytics platform, run two says SMB dashboard and run three says startup tool, your positioning is noise to the model. Veritas Links pitches this variance as a core audit issue; their "25% of runs" figure is a marketing illustration rather than a study, but the principle holds: one run is anecdote.
Perplexity is useful here. Promptwatch's data shows it cites almost exactly 10 sources daily, which makes it a steadier test bench. If your citations change there, it is probably your content, not engine noise. Copilot, by contrast, swung from under 2 to nearly 17 sources within weeks, so judge it on monthly trends.
8. Which content gaps are letting competitors win?
For each prompt where you are missing, ask what page on your site would answer it. Often the answer is none. Cross-reference with the content types AI actually cites. Promptwatch Data on ChatGPT citation share by content type (Sep 28 to Oct 4, 2026) found:
| Content type | Share of ChatGPT citations |
|---|---|
| Product pages | 22.4% |
| Landing pages | 16.8% |
| Listicles | 13.6% |
| How-tos | 8.7% |
| Documentation | 7.5% |
| Comparisons | 5.9% |
| Reviews | 1.7% |
For branded prompts specifically, reviews climb to 5.6% (more than triple their overall share) and social posts to 3.0%. So when someone asks about you by name, third-party opinion carries disproportionate weight. If your product and landing pages are thin, or you have no comparison pages against your two main rivals, you have found your first content tasks.
Also consider shelf life. According to Promptwatch's content half-life study, landing pages and docs stay cited longest while news loses half its pages within about 36 days. Evergreen pages are the better investment for perception work.
9. Can AI crawlers actually read our site?
Sometimes the audit finds a perception problem that is really an access problem. If your robots.txt blocks GPTBot, or key pages return errors to bots, the model is working from secondhand descriptions of you. Check:
- robots.txt rules for AI crawlers such as GPTBot, ClaudeBot and PerplexityBot
- Server logs for actual AI crawler hits on your key pages
- Whether important pages depend on client-side rendering that bots may not execute
- Whether you have an llms.txt file (optional, but cheap)
Promptwatch has a free Agentic Readiness Check and a robots.txt generator for AI if you want a quick read without setting up log analysis. Cross-check which crawlers are even active with Promptwatch Data on AI crawler traffic.
10. How does this compare with last quarter?
A perception audit with no baseline is a snapshot. A snapshot is still useful, but the real value shows up on the second run, when you can say "mention rate rose from 27% to 41% after we published the comparison pages, and the pricing error is gone." Date your spreadsheet, save raw responses, and schedule the next audit.
For reference, AI engines themselves shift. ChatGPT started using the site: operator at scale on August 8, 2026, and Reddit's share of ChatGPT citations dropped from around 4% to 0.5% on August 14, per Promptwatch's data on Reddit citations. Changes like that can move your numbers with no action on your part, so log the date of every run.
Turning answers into a one-page scorecard
Brand reviews run on summaries. Collapse the ten answers into a single page.
| Dimension | Metric | Last run | This run | Owner |
|---|---|---|---|---|
| Presence | % category prompts with mention | |||
| Prominence | Average position when named | |||
| Accuracy | Count of wrong or outdated claims | |||
| Category fit | % responses with correct category | |||
| Tone | % confident vs hedged | |||
| Competition | Top 3 competitors named | |||
| Sources | Top 5 cited domains | |||
| Consistency | % prompts with stable answer across 3 runs | |||
| Access | AI crawlers able to read key pages |
Then add three lines at the bottom: the worst error, the biggest competitor surprise, and the first fix. That is what people will remember.
If your numbers point in a direction, read them like a diagnosis. Low presence is a coverage problem. Late position is a prominence problem. Mentions without citations to your own domain suggest a content or crawlability issue. Weak sentiment is a reputation problem and usually lives in third-party sources, not your site.
Doing it manually versus with a tool
The manual version is fine for a first pass. It is also tedious, and it breaks down quickly: 25 prompts times 4 engines times 3 runs is 300 responses to read and log. Responses also differ by country and language, and API outputs can differ from what users see in the real interface.
For ongoing tracking, a few categories of tools help:
| Approach | Good for | Limits |
|---|---|---|
| Spreadsheet plus manual prompts | First audit, small prompt sets | Slow, no trend history, easy to introduce bias |
| Free one-off graders | A quick directional read | Usually one run, limited prompts |
| Monitoring-focused trackers | Dashboards and alerts | Often stop at reporting |
| End-to-end AI visibility platforms | Tracking plus fixing and measuring | Higher cost than basic trackers |
Promptwatch sits in the last row. It monitors the real interfaces of ChatGPT, Gemini, Perplexity, Google AI Overviews, Copilot and others across countries and languages, scores visibility and sentiment, and its Brand Narrative feature pulls out the individual claims AI makes about you ("affordable for small teams", "mobile app lacks features") and groups them by theme. That maps directly onto questions 2, 3 and 4 above. Fact Check, on Business plans and up, lets you mark a claim inaccurate and find the pages on your site teaching the wrong thing. The platform then helps close the loop with content tools and crawler logs for question 9. Pricing starts at $95/month for Essential; confirm current plans at promptwatch.com/pricing.

If you want to start without paying, Promptwatch's free AI visibility report gives a no-login scan of how ChatGPT, Gemini and Perplexity talk about your brand.
Other options worth a look in our catalog, depending on budget and need:

Otterly.AI and Peec AI are lighter monitoring tools. Profound is a heavier enterprise option. For a wider survey of the category, the GEO software directory at bestgeosoftware.com lists many more.
Common mistakes in AI perception audits
- Testing one engine. ChatGPT, Gemini and Perplexity retrieve and weigh sources differently, so results differ.
- Asking once. Run each prompt at least three times and report the spread.
- Only asking brand prompts. Buyers ask category and comparison questions first, and that is where you may be invisible.
- Fixing your website only. If the wrong claim comes from a review site or old directory entry, editing your homepage won't help.
- Treating third-party stats as gospel. The "47% of B2B buyers use AI for purchase research" figure in Renown's checklist and the Gartner 2028 prediction it cites are unverified here. Use them as context, not as proof in front of your CFO.
- Skipping the dates. Models and retrieval behavior change. An undated audit is hard to trust six months later.
What to do in the review itself
Open with the scorecard, not the methodology. Show two verbatim answers, one good and one bad. Name the three biggest fixes, each with an owner and a re-test date. Then book the next audit.
If you would rather hand the whole thing to specialists, agencies that cover GEO can run the audit and the follow-up work. 1001 SEO Media does this for brands that want to be cited in ChatGPT, Perplexity and Google AI Overviews, and uses Promptwatch for tracking.
Otherwise the sequence is simple: freeze the prompts, run them across engines three times, score the ten questions, and fix the errors that sit in sources you can actually edit. Then run it again in 90 days and see what moved.

