The AI Brand Audit Framework: What ChatGPT, Claude, Gemini & Perplexity Tell Your Buyers
TLDR
An AI brand audit is a structured, repeatable test of how AI models describe your brand to buyers. The framework has five steps: (1) build a fixed set of 12 prompts across discovery, knowledge, comparison, and purchase-intent categories; (2) run them through ChatGPT, Claude, Gemini, and Perplexity with web search on; (3) score every answer 0–8 on a four-check rubric — mentioned, categorized, accurate, cited; (4) diagnose failures with the crawled→retrieved→cited funnel; (5) fix the highest-impact gap and re-audit with identical prompts. Optimly's AI Brand Index automates the whole loop in 30 seconds. The audit is step one of the larger find → fix → prove cycle — see Closing the Loop for what to do with the findings.
What an AI Brand Audit Actually Reveals
Most brands have never checked what happens when a buyer asks ChatGPT "What's the best [your category] tool?" When they do, the audit surfaces one of four problems:
You're Invisible
18.5%AI doesn't mention you at all. You have zero meaningful presence across any major AI model.
You're Miscategorized
59.8%AI puts you in the wrong category, so you never show up for the right buyer queries.
You're Misrepresented
CommonAI gets your story wrong — outdated positioning, confused differentiators, or a narrative that no longer matches who you are.
Competitors Show Up Instead
GrowingWhen buyers ask about your category, AI recommends everyone but you.
Each of the four problems has a different root cause and a different fix. The point of a structured audit is knowing which one you actually have — before spending a quarter fixing the wrong thing.
Step 1: Build Your Question Set — The Prompt Library
An audit is only as good as its questions, and only comparable if you ask the same questions every time. Use these 12 prompts — swap in your category, company, and top competitor. Discovery prompts matter most: they test whether AI recommends you to buyers who don't know your name yet.
Category discovery
Are you on the list when buyers don't know your name?
What are the best [your category] tools in 2026?I'm a [buyer role] evaluating [your category] platforms. Which should I shortlist and why?Compare the top [your category] vendors for a mid-market B2B company.
Brand knowledge
Does AI know who you are and what you do?
Tell me about [your company]. What does it do and who is it for?What category of software is [your company]? Who are its main competitors?What does [your company] cost? What's included in each pricing tier?
Competitive comparison
How do you come out when buyers compare you head-to-head?
[Your company] vs [top competitor] — which is better and for whom?What are the main criticisms or limitations of [your company]?Why would a company choose [competitor] over [your company]?
Purchase intent
What does AI say at the moment of decision?
Is [your company] worth it? Summarize what real users say.Write a short recommendation memo for my CFO on buying [your company].What questions should I ask [your company] in a sales call?
Step 2: Run the Set Through Each Model
The four major assistants retrieve and cite differently, which is exactly why you audit all of them. For every prompt, record the full answer and every cited source URL.
ChatGPT (GPT-5)
How: Ask with web search enabled (default in Plus). Sources appear as inline citation chips — expand them.
Watch for: Without search, answers come from training data and can be years stale. Run each prompt twice: once with search, once with "don't search the web" to see the parametric baseline.
Claude
How: Enable web search in settings. Citations appear as numbered footnotes under the answer.
Watch for: Claude is the most likely to say "I don't have reliable information" — that's an invisibility finding, not a null result. Record it.
Gemini
How: Grounding with Google Search is on by default. Click the chevron under answers to reveal the source list.
Watch for: Gemini leans heavily on Google's index — if you rank poorly in Google, expect it to surface third-party review sites instead of you.
Perplexity
How: Every answer is retrieval-first with numbered citations by design. This is the fastest way to see which URLs actually get retrieved for your category.
Watch for: Note whether your own domain ever appears in the citation list. If competitors' comparison pages show up and yours don't, that's your content gap, mapped for you.
Step 3: Score Every Answer — The Rubric
Score each answer 0–2 on four checks, for a maximum of 8 points per answer. Average across models per prompt, and across prompts per category. The score itself matters less than the trend — the whole point of a fixed prompt set is that next month's audit is comparable.
| Check | Question | 0 | 1 | 2 |
|---|---|---|---|---|
| Mentioned | Does the answer name your brand at all? | Absent | Named in passing | Featured or recommended |
| Categorized | Is your brand placed in the right category? | Wrong category | Adjacent / vague | Exactly right |
| Accurate | Are the facts (pricing, features, positioning) correct? | Wrong or invented | Partially right / stale | Matches your ground truth |
| Cited | Does the answer cite YOUR domain as a source? | Only third parties | Mixed | Your pages cited |
The Cited check is the one most teams skip — and the one that predicts everything else. If AI never cites your domain, your brand story is being told entirely by third parties: review sites, competitor blogs, and stale press.
Step 4: Diagnose with the Crawled → Retrieved → Cited Funnel
When a page you expected to be cited isn't, find out where it fails. There are only three places to lose:
1. Not crawled
Check your server or CDN logs for AI crawler user agents — GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User, PerplexityBot, Google-Extended. If they never fetch the page, fix discoverability: robots.txt, sitemap, internal links, llms.txt.
2. Crawled, but not retrieved
Crawlers fetch the page, but answer-time retrieval never selects it for buyer questions. Usually the page doesn't match how buyers phrase questions — the heading is a slogan instead of the question, and the answer is buried under marketing copy.
3. Retrieved, but not cited
The model reads your page but quotes a third party instead. Your content lacks quotable, self-contained claims: direct one-paragraph answers, concrete numbers, structured data. This is the "answer-shaped content" problem.
Being crawled is not being cited. A page can be fetched by AI crawlers every week and still never appear as a source, because citation is won at the content level — not the infrastructure level.
Step 5: Fix What You Found — Then Re-Audit
Every finding maps to a specific fix. Pick the single highest-impact gap, ship the fix, then re-run the identical audit in 2–4 weeks and compare scores.
| Finding | Likely cause | Fix |
|---|---|---|
| Invisible on category prompts | No answer-shaped content matching buyer questions | Publish pages that literally answer the category questions buyers ask, with FAQ/HowTo schema |
| Wrong category | Inconsistent positioning across your site, reviews, and directories | Align category language everywhere; publish a ground-truth brand profile and llms.txt |
| Stale or wrong facts | Models leaning on outdated third-party sources | Publish authoritative pricing/feature pages; correct review-site listings; keep llms.txt current |
| Crawled but never cited | Content isn't structured for answer engines — no direct answers, no schema, buried claims | Restructure: question-led headings, one-paragraph direct answers, stats with sources, FAQ schema |
| Competitors cited instead | They own the comparison and definition pages for your category | Publish comparison pages, definitional guides, and original data competitors can't match |
Schema markup visibility boost
Content hub visibility boost
Brands with llms.txt
Frequently Asked Questions
How do I audit what ChatGPT, Claude, Gemini, and Perplexity say about my brand versus competitors?
Run the same question set through all four models with web search enabled: category discovery prompts, brand knowledge prompts, and head-to-head comparisons. Score every answer on four checks — mentioned, categorized correctly, factually accurate, and whether your own domain is cited. Repeat monthly with identical prompts so scores are comparable. The AI Brand Index automates this across all four models in one run.
Why does AI describe my company in the wrong category?
Miscategorization accounts for 59.8% of AI brand inaccuracies. It happens when your website, review-site listings, and press use inconsistent category language — AI synthesizes the loudest signal, which is often a stale third-party description. Fix it by aligning category wording everywhere, publishing a ground-truth brand profile, and maintaining an llms.txt file.
What prompts should I use for an AI brand audit?
Use four prompt types: category discovery ("What are the best [category] tools?"), brand knowledge ("Tell me about [company]"), competitive comparison ("[company] vs [competitor]"), and purchase intent ("Is [company] worth it?"). The 12-prompt library above covers all four.
How do I know if AI models are reading my website?
Check your server or CDN logs for AI crawler user agents: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User, PerplexityBot, and Google-Extended. But being crawled is only step one — audit the full crawled→retrieved→cited funnel above.
How often should I run an AI brand audit?
Monthly at minimum, with identical prompts each time so scores are comparable. Weekly is ideal while you're actively fixing findings — retrieval-based answers can change within days of publishing new content.
Is the AI Brand Index audit free?
Yes. The AI Brand Index is a free, open community resource tracking thousands of brands. Look up your brand, add it if it's missing, and help verify the data.
Related Resources
Page Summary: The AI brand audit framework is a five-step, repeatable process: (1) a fixed 12-prompt library across discovery, knowledge, comparison, and purchase-intent questions; (2) per-model runs on ChatGPT, Claude, Gemini, and Perplexity with web search enabled; (3) a 0–8 scoring rubric per answer (mentioned, categorized, accurate, cited); (4) diagnosis via the crawled→retrieved→cited funnel; (5) remediation mapped to each finding, followed by a re-audit with identical prompts. The most commonly skipped check — whether AI cites your own domain — is the strongest predictor of long-term AI brand accuracy.
Skip the Spreadsheet — Run the Audit in 30 Seconds
Optimly runs this exact framework automatically: multi-model queries, citation tracking, accuracy scoring against your ground truth, and run-over-run comparison.