To check whether ChatGPT mentions your brand, run a fixed set of buyer-style questions in fresh chats, repeat each question on another day, and record three separate things: whether the name appears, whether your domain is cited, and whether the description is correct. Asking only “Do you know [Brand]?” is a weak test. One answer can change on the next run, so build a baseline before drawing conclusions.
We run LaunchAF, which offers a free GEO checker, so weigh that accordingly. This is a manual measurement method, not a fix for every missing mention. If you need a short first diagnosis, read why a startup may be absent from ChatGPT; this guide goes further on prompt design, logging, and interpretation.
Key takeaways
- Separate a brand lookup from an unbranded discovery question. The latter shows whether a buyer could encounter you without already knowing your name.
- Keep the wording, model, search setting, and location context as consistent as you can across runs.
- Record brand mention, first-party citation, and factual accuracy in separate columns.
- Repeat prompts. An isolated appearance is an observation, not a stable position in AI answers.
- Follow source links back to real pages before deciding whether the problem is crawl access, a stale profile, or simply an answer that chose other evidence.
ChatGPT's search and temporary-chat documentation was checked on October 14, 2026. Interfaces, model names, and settings can change by plan or workspace. Write down what your own screen shows rather than forcing your results into an outdated screenshot of someone else's setup.
What counts as a "mention"?
A mention is a brand name appearing in the answer. That does not tell you whether the answer linked to your website, described your product correctly, or recommended it to the right buyer. Use the following four labels as a practical rubric invented for this guide, not an industry standard or an official ChatGPT score.
| Outcome label | What you observed | What it does not prove |
|---|---|---|
| Not mentioned | Neither the brand nor an unambiguous product reference appears | That ChatGPT has never seen your site |
| Mentioned without a source | The answer names the brand but does not link your domain | Which page, if any, supplied the claim |
| Cited | Your own domain appears as a visible source link | That the answer is accurate or that every run will cite it |
| Recommended and described accurately | The answer presents the product as a relevant option and gets its core facts right | That it will recommend you to other users or for other prompts |
These labels are not a perfect ladder. An accurate recommendation may be uncited, and a citation may accompany a wrong claim. Keep citation and accuracy in separate columns. A directory link is a third-party source, not a citation of your domain.
Step 1: Decide what you are testing
Write down your canonical brand spelling and domain, the category buyers use, three to five real competitors or substitutes, and three customer problems. Test buyers' words, not only your homepage headline. A broad “best project management software” prompt may be unfair to a niche design tool; a relevant buyer question is better, provided it does not suggest your brand.
Choose the outcome you need to observe. An unbranded prompt tests discovery; a branded pricing prompt tests accuracy. Include a question where your product is not a fit. The GEO guide gives a shorter version of this exercise.
Step 2: Build a prompt set
Use eight to ten fixed prompts. The nine below cover recognition, discovery, comparison, and trust. Replace bracketed text once, save the final wording in your sheet, and do not keep polishing it between runs. A discovery prompt must not name your brand: otherwise you are measuring how ChatGPT responds to a suggestion you supplied.
| Prompt type | Template prompt | Main thing to observe |
|---|---|---|
| Brand lookup | “What does [Brand] at [domain] do, and who is it for?” | Basic identity and current description |
| Category discovery | “What are useful [category] tools for [audience]?” | Unprompted appearance among relevant options |
| Best for a use case | “Which tools can help [audience] [specific task] without [constraint]?” | Fit for a real buying job |
| Comparison | “Compare [Brand] with [Competitor] for [use case]. What are the trade-offs?” | Accuracy and balanced differentiation |
| Alternative to a competitor | “What are alternatives to [Competitor] for [audience]?” | Discovery through a familiar option |
| Problem-led | “How can [audience] solve [problem]? Name software only if it would help.” | Whether the need maps to the product at all |
| Pricing or access | “What does [Brand] cost, and is there a free way to try it?” | Stale plan and availability facts |
| Trust or legitimacy | “Who runs [Brand], and what can I verify before signing up?” | Operator and public evidence |
| Integration | “Does [Brand] work with [integration]? Where is that documented?” | Feature accuracy and source quality |
The category, alternative-to, and problem-led prompts deliberately omit your brand. Replace an irrelevant integration row with a real setup question. Keep the wording after a disappointing answer; a rewrite is a new test. Give any added prompt its own identifier, and use separate sets for different languages or countries.
Step 3: Run them cleanly
Start a fresh chat per prompt. Keep the same model where possible, and note whether web search was on, off, automatic, or unavailable. OpenAI says ChatGPT may search automatically and search responses may include citations. Record what you can observe rather than inferring an unseen retrieval step.
If available, choose an unpersonalized Temporary chat to reduce account-context effects. OpenAI's Temporary chat help says a personalized temporary chat can still use existing memories and instructions. Note any workspace setting you cannot control.
Run each prompt three times and repeat on another day. This is a suggested routine, not a validated minimum. Record date, model, search setting, location or language, and sources. If time is short, start with the brand, category, and problem-led prompts and label the baseline partial.
Step 4: Record results
Copy this header into a spreadsheet, one row per response. Describe position_in_answer as “first paragraph” or “third option,” not a stable rank. Save source URLs, and use unknown when a search setting is unclear.
| date | tool | model_shown | web_search_on | prompt_type | prompt | run | brand_mentioned | domain_cited | position_in_answer | sources_listed | accurate | competitors_named | notes |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2026-10-14 | ChatGPT | [model shown] | [yes/no/auto/unknown] | category discovery | [saved prompt] | 1 | [yes/no] | [yes/no] | [where or none] | [URLs or none] | [yes/no/unclear] | [names] | [wrong or missing facts] |
Here is the same header as CSV. Paste it into a new file or the first row of a sheet, then fill in observed results. Do not copy the example row as though it were a real test.
date,tool,model_shown,web_search_on,prompt_type,prompt,run,brand_mentioned,domain_cited,position_in_answer,sources_listed,accurate,competitors_named,notes
Save an important answer or screenshot. If it cites a launch profile rather than your site, record and check that URL too.
Step 5: Score and read the pattern
Apply the four labels, then review citation and accuracy separately. Report counts as “mentioned in two of six saved runs,” not as a market-wide visibility score or a prediction for other users.
| Pattern in your sheet | Plausible interpretation, not a diagnosis | Sensible next step |
|---|---|---|
| Brand lookup repeatedly fails | Access or identity may be unclear | Check the ChatGPT diagnosis, public page, and crawler rules |
| Brand lookup works, discovery does not | The product may be recognizable but not supported as a relevant option | Clarify use case and inspect independent, relevant references |
| Brand appears with wrong price or capability | A first-party page or cited profile may be stale | Correct the source and retest the same prompt |
| Competitors are cited, your site is not | Their pages may answer the question more directly or be more accessible | Read the cited pages and use the citation playbook for content and access checks |
| Brand appears only sometimes | Answers vary by run or search context | Keep the prompt and settings fixed, then collect more observations |
The table suggests what to inspect, not how a hidden recommendation system works. Do not delete a useful page after one omission.
Step 6: Follow the sources
Open the visible sources. OpenAI's search help says a Sources view may be available and warns that citations can be incomplete or outdated. Check whether each linked page supports the answer's claim.
Correct wrong facts on your own page or request a correction to a cited profile. If competing pages answer a question yours does not, consider writing one useful answer page. Give buyers evidence, not a citation-chasing rewrite.
Log the page, correction, and date; retest without expecting an immediate refresh. The Claude, Gemini and Perplexity guide explains why their results may differ.
Step 7: Check that AI crawlers can reach you
Run the free LaunchAF GEO checker on a public page. It checks named bot rules, index and snippet signals, sample firewall responses, basic entity and answer cues, optional llms.txt, and manual test prompts. Its suggested fixes are a site-audit heuristic. It does not read private ChatGPT answers or calculate a citation probability.
Use the robots.txt tester for a specific URL and agent, then check CDN logs. An allowed rule does not prove a successful fetch. OpenAI documents OAI-SearchBot search access separately from GPTBot training access; you can control them independently.
If access is clean, return to the answer evidence. A reviewed LaunchAF submission can add an accurate public profile after badge verification and review, but it cannot make ChatGPT name your company.
How often should you retest?
As a practical suggestion, retest monthly and after a domain, price, documentation, or major listing change. Keep old rows. Track referrals and signups separately: the sheet measures answer appearances, not business value.
Mistakes that make the results meaningless
- Asking only “Do you know [Brand]?” and treating a yes as proof of unbranded discovery.
- Counting one favorable run, then changing the wording until it appears again.
- Using a personalized chat without noting memory or workspace instructions; a temporary chat can still use prior context unless unpersonalized.
- Mixing ChatGPT, Claude, Gemini, and Perplexity in a single trend line, or comparing different models and countries as though they were the same conditions.
- Treating a directory citation as a first-party-domain citation, or one visible citation as a guaranteed future recommendation.
- Putting your brand in every category prompt, then calling the answers organic discovery.
- Hiding instructions aimed at AI models, cloaking content, inventing reviews, spamming directories, or paying a seller who guarantees citations. These tactics make the underlying evidence worse, not the measurement better.
Do paid AI-visibility trackers replace this?
Some paid services automate prompt runs and collect mentions, but automation does not remove sampling choices. Before buying, ask which models, search settings, languages, locations, and prompts they use; whether the same prompts are repeated; and how they distinguish a brand mention from a source citation. This manual sheet is a free baseline against which to judge a vendor's methodology, not a claim that manual work scales forever.
Frequently asked questions
Can I just ask ChatGPT if it knows my brand?
You can, but it tests recognition after you supplied the name. It does not tell you whether a buyer asking an unbranded category question would see the product. Include both kinds of prompt and keep their results separate.
How many prompts do I need?
Begin with the nine types in this guide, adapting only those relevant to your product. Repeat each saved prompt across fresh chats and another day. Nine is a workable starting set, not a statistically validated minimum.
Why does ChatGPT describe my product wrongly?
The answer may rely on a stale source, confuse a similar brand, or generate an unsupported claim. Open any cited pages and compare them with your current product facts. Correct pages you control, request important third-party corrections, and retest without expecting instant propagation.
Do Claude, Gemini and Perplexity give the same results?
No. Their documented crawler roles, search behavior, and source displays differ, and answers vary by prompt and run. Use separate sheets or a clearly recorded tool field. Our three-assistant diagnosis explains what to check for each.
Is there a free ChatGPT brand mention checker?
The prompt-and-sheet method here is free. LaunchAF's GEO checker audits public pages; it cannot see private ChatGPT conversations or count all user mentions. Be cautious of tools implying otherwise.
Will a LaunchAF listing help?
An approved listing can add a public, consistent product description and a relevant backlink under its terms. The free route needs a verified badge and review, and neither a listing nor a followed link guarantees an AI mention. Choose it for its founder audience and useful profile, then measure outcomes separately.
The bottom line
Make a saved set of nine buyer questions today and run the baseline in fresh chats. Log mentions, citations, and accuracy as distinct observations, then repeat the same prompts on another day. If access is one of the problems you uncover, run the free GEO checker and fix a concrete failing item. Keep the sheet honest: a pattern is useful, but one answer is not a verdict.