To check if ChatGPT mentions your brand, don't ask about your brand: ask the questions your customers ask ("what's the best X for Y?") and see whether you appear in the answer. Do it in a fresh session, repeat it several times across several AI models, and record not just whether you're mentioned but how. One-off checks are close to meaningless: AI answers change between runs, so what you need is a mention rate, tracked over time.
Here is the method, step by step, and where its limits are.
Key takeaways
- Test category questions ("best X for Y"), never your brand name: recognition is not visibility.
- Use clean sessions and repeat every prompt: AI answers vary between runs, so the metric is a mention rate, not a single check.
- Log competitors, position, and how you are described, because a rate only means something in comparison.
- Done manually, one measurement cycle is ~400 conversations across four models; that is the job monitoring tools automate.
Why not just ask "do you know [my brand]?"
Because the answer is flattering and useless. Asked directly, a model will retrieve or recall something about almost any brand with a website. That tells you nothing about the moment that matters commercially: when a potential customer asks a category question ("best SEO tool for a small agency", "alternatives to X") and the model composes a shortlist. You want to know if you're on that shortlist. That is a completely different test, and it's the one this guide covers.
Step 1: write the prompts your customers actually use
Build a list of 10–20 realistic prompts across four intent types:
- Category: "What's the best [category] tool for [audience]?"
- Problem: "How do I [job your product does]?", where a good answer would name tools.
- Comparison: "[Competitor] vs alternatives: what should I pick?"
- Local/niche variants: the same questions with your market's qualifiers ("...for an Italian e-commerce", "...on a small budget").
Write them the way a real person would type them: hurried, specific, imperfect. Skip prompts that contain your brand name: those measure recognition, not visibility.
Step 2: run them in clean conditions
Your logged-in ChatGPT knows you: memory, custom instructions, and chat history all contaminate answers. For each test:
- use a fresh chat with memory off (or an incognito/logged-out session),
- don't "warm up" the conversation: the test prompt goes first,
- note whether the model browsed the web or answered from memory, when visible: the distinction maps to the two doors described in How AI models choose which brands to cite.
Step 3: record more than yes/no
For each prompt and run, log:
| Field | Why it matters |
|---|---|
| Mentioned? (you / competitors) | The core visibility signal, theirs as much as yours |
| Position in the answer | First name carries more weight than "also worth a look" |
| How you're described | Wrong category or outdated facts is its own problem to fix |
| Sentiment / caveats | "Popular but pricey" shapes decisions |
| Sources cited | Which pages earned the citation: yours or someone else's review of you |
The competitor columns are not optional. "Mentioned in 3 runs out of 10" is uninterpretable alone; next to "competitor A: 9 out of 10" it becomes a roadmap.
Position and phrasing are worth the extra columns. The research paper that introduced GEO measures visibility with metrics that weight where a citation appears in the answer and how much of the answer it carries, not merely whether it is present, and adds a subjective score for relevance, influence and uniqueness (Aggarwal et al., KDD 2024).
Step 4: repeat, because one answer is an anecdote
AI answers are non-deterministic: the same prompt, same model, same day can produce different shortlists. A single run tells you what one dice roll looked like. Meaningful measurement needs:
- several runs per prompt (5–10 gives a usable rate),
- several models: ChatGPT, Gemini, Claude, and Perplexity retrieve differently and know different things; being visible on one says little about the others,
- regular re-runs (weekly or monthly) so you see whether your GEO work moves the number.
This is the same shift described in GEO vs SEO: from a deterministic rank to a sampled mention rate.
The honest math on doing this manually
Twenty prompts × 5 runs × 4 models = 400 conversations per measurement cycle, each needing a clean session and structured note-taking, then aggregation into rates and trends. As a one-off audit, it is a half-day well spent, and worth doing once just to feel the problem. As a monthly practice, it's a part-time job, and the least automatable part (consistent scoring of position, sentiment, and description) is exactly where manual work gets sloppy.
This is the job LLM visibility monitoring exists for. On paid plans, Huntair checks the prompts you choose across ChatGPT, Gemini, Claude, and Perplexity on a schedule, tracks your mention rate against competitors over time, and puts it next to your traditional SEO data, so one dashboard answers both "do I rank?" and "am I cited?".
What if the answer is "you're invisible"?
Then you've learned something most of your competitors don't know about themselves, and the fix has a defined path:
- Check retrievability first: readable without JavaScript and the AI search crawlers allowed in robots.txt: OAI-SearchBot for ChatGPT, Claude-SearchBot for Claude, PerplexityBot for Perplexity, and Google-Extended if you want Gemini to ground its answers on your pages. This is the most common and fastest-fixed cause.
- Make your content quotable: answer-first pages, question-shaped headings, real data. The patterns are in What is GEO? A practical guide.
- Sharpen your entity: one name, one category, one description, everywhere.
- Re-measure: same prompts, same method, four weeks later. The rate is your scoreboard.
FAQ
How often should I re-check my AI visibility?
Monthly is the minimum for a trend; weekly if you're actively publishing GEO content and want feedback on what moves the number. What matters most is keeping the prompt set stable, so changes reflect the models, not your questions.
My brand appears but is described wrong. Does that matter?
Yes, arguably more than absence. A wrong category or outdated pricing follows you into every answer. Fix the source: your own pages, then the third-party pages models cite about you (directories, reviews, comparison articles).
Do I need to test all four models?
If your customers use them, yes. Retrieval differs, because each provider runs and documents its own search crawler and index, so visibility on one model genuinely doesn't transfer. At minimum, test the two your audience uses most.
Can ChatGPT's answers about my brand change even if I change nothing?
Yes: model updates, index refreshes, and plain sampling variance all move answers. That's exactly why a mention rate over repeated runs, not a single check, is the metric.