How to Check What AI Says About Your Brand (The Actual Methodology)
The Manual Method — Do It Yourself
Walk through these 5 query categories on at least 3 models (ChatGPT, Claude, Gemini). Each one probes a different dimension of AI's understanding.
1. Direct identity queries
"What is [your brand]?"
"Describe [your brand]'s main products and services."
What you're testing: Does AI have basic parametric knowledge of you?
Look for: Correct industry classification, accurate product descriptions, current information.
Red flags: Wrong category entirely, outdated descriptions, confusion with similarly-named companies.
2. Category placement queries
"What are the best [your category] companies?"
"Compare [your brand] to competitors in [your space]."
What you're testing: Whether AI includes you in the competitive set where buyers are looking.
Look for: Your brand appearing in the list. Position relative to competitors.
Red flags: AI lists 10 competitors and you're not among them. You're invisible in the exact moment a buyer is deciding.
3. Buyer intent queries
"I need a [your solution type] for [common use case]. What should I consider?"
What you're testing: This is the money query. It simulates an actual buyer asking AI for recommendations.
Look for: Your brand appearing as a recommendation with accurate positioning.
Red flags: If you don't appear, you're losing pipeline to whoever does.
4. Sentiment probes
"What are the strengths and weaknesses of [your brand]?"
What you're testing: What narrative AI has internalized about you.
Look for: Strengths that match your actual positioning. Weaknesses that are fair.
Red flags: AI thinks your strength is something you deprioritized two years ago — your parametric knowledge is stale.
5. Competitor displacement queries
"Is [your brand] better than [top competitor] for [your core use case]?"
What you're testing: Whether AI positions you correctly relative to alternatives.
Look for: Accurate comparison of differentiators.
Red flags: AI recommending the competitor for your strongest use case, or describing your differentiators as shared features.
Important: run every query on at least 3 models (ChatGPT, Claude, Gemini). They each have different crawling behavior and different "versions" of your brand.
For buyer intent queries specifically, a good response names you in the top 3 recommendations with your actual differentiator. A bad response lists 5 competitors and doesn't mention you, or mentions you but describes your old product. For competitor displacement queries, pay attention to how AI frames the comparison. Does it say "[competitor] is the market leader and [you] is an alternative" or does it present you as equals? The framing reveals AI's confidence hierarchy — and that hierarchy is what determines who gets the recommendation when a buyer asks a neutral question.
Here's What a Full Manual Audit Actually Looks Like
We audited a mid-market cybersecurity company (anonymized). Here's what we asked and what each model said:
| Query Category | ChatGPT | Claude | Gemini |
|---|---|---|---|
| Direct identity | ✓ Correct — "cybersecurity platform" | ✗ Called them "IT staffing agency" | ✓ Correct but outdated product list |
| Category placement | ✗ Not listed in "best cybersecurity tools" | ✗ Not listed | ✗ Not listed |
| Buyer intent | ✗ Recommended 3 competitors | ✗ Recommended 2 different competitors | ✗ Didn't mention category |
| Sentiment probe | Mixed — cited outdated weaknesses | ✗ Described IT staffing strengths | Neutral, generic description |
| Competitor displacement | Competitor framed as "leader" | Couldn't compare — wrong category | Fair comparison, lacked depth |
Diagnosis: Misread on Claude (wrong category entirely), Phantom on category/intent queries across all models (present when asked directly, invisible when buyers are searching). Total time for manual audit: 45 minutes. This brand had no idea AI was misrepresenting them.
The root cause was clear: their Crunchbase profile still listed "IT staffing" from their pre-pivot days. Claude weighted this heavily. ChatGPT's retrieval pulled from their updated website and got it right, but the buyer intent and category queries relied on parametric knowledge — which was wrong everywhere.
How to Interpret What You Find
Use the archetype framework as the diagnostic tool. After running the queries, the self-diagnosis takes 60 seconds:
Phantom
(111 of 5,829 brands)Signal: AI doesn't mention you at all across most queries
Urgency: High — every AI-assisted buying decision happens without you.
Misread
(47 of 5,829 brands)Signal: AI mentions you but gets significant facts wrong
Urgency: Highest — AI is actively steering buyers away with incorrect information.
Challenger
(463 of 5,829 brands)Signal: AI mentions you correctly but names competitors first
Urgency: Medium — you're present but not preferred.
Incumbent
(379 of 5,829 brands)Signal: AI describes you accurately and recommends you
Urgency: Monitor — maintain your position.
Why Models Disagree With Each Other
Each AI model has two types of knowledge: parametric (baked in during training) and retrieved (fetched in real-time). The balance differs by model.
ChatGPT relies more on retrieved data — OpenAI sent 10,816 crawler requests to our directory this week.
Claude leans more on parametric knowledge — Anthropic sent 4,669 requests.
Perplexity is retrieval-first — 1,699 requests, each answering a live user query.
A brand might score well on ChatGPT (which just retrieved your updated page) but poorly on Claude (working from older training data). This is why single-model testing gives a false sense of security.
| Model | Crawl Volume | Knowledge Type | Update Speed | Implication |
|---|---|---|---|---|
| ChatGPT | 10,816 req/wk | Heavy parametric + retrieval | Fast (active crawling) | Most likely to have current data, but parametric memory can conflict with retrieved data |
| Claude | 4,669 req/wk | Parametric-heavy | Moderate | May lag behind on recent changes; strong on established brands |
| Perplexity | 1,699 req/wk | Retrieval-first | Real-time | Most accurate for current state, least affected by stale training data |
| Gemini | Varies | Mixed, Google-indexed | Google-indexed | Leverages Google's index; strong where Google is strong |
The practical takeaway: if you only test one model and it looks good, you'll miss that another model has you completely wrong. ChatGPT's high crawl volume means it often has the most current data — but that also means a recent error on your site propagates to ChatGPT faster than to other models. Test at least 3 models for every query category. Read more about crawler behavior in our crawler data report.
What to Do With the Results
The remediation framework, based on our data from 5,829 brands:
- 1.
Fix your structured data
Add Organization schema with correct industry, product descriptions, and founding date. Guide →
- 2.
Align your authoritative sources
Make sure your website, Crunchbase, LinkedIn, Wikipedia, and G2 all tell the same story. Guide →
- 3.
Create an llms.txt file
The AI-specific equivalent of robots.txt — tells AI models what your brand is and does. Guide →
- 4.
Configure your robots.txt
Ensure AI crawlers (GPTBot, ClaudeBot, PerplexityBot) are welcomed, not blocked. Guide →
- 5.
Create content that corrects misclassifications
If AI thinks you're in the wrong category, publish a definitive page that corrects the record. Guide →
When to Automate It
If you have 1 brand, the manual method works. Run the 5 query categories across 3 models quarterly, track the changes in a spreadsheet, and you'll have a reasonable picture.
If you're managing multiple brands, need to track changes over time, or need to audit across 4+ models simultaneously — that's where tooling helps. Our free AI Visibility Checker runs the full methodology and gives you a BAI score in seconds.