You can check your ChatGPT visibility by hand with a fixed list of buyer questions, a clean chat session for each one, and a spreadsheet where you log whether your brand was mentioned, where it sat in any recommendation list, and which of your pages were cited. This guide walks you through that from the first question to your first repeat check. By the end you will have a prompt set, a written test protocol, a baseline you can defend, and a schedule for seeing what changes. If you also care about Google, there is a separate guide to the best tools that track Google AI Overviews.
You will need access to ChatGPT, a spreadsheet, a list of the names your brand goes by, the websites you own, and a place to save screenshots or copied answers. This works as a basic ChatGPT visibility checker you build yourself, so it stays manual on purpose. It does not compare tools, and it does not explain how automated tracking works behind the scenes.
1. Build and freeze your buyer prompt set
Start by writing the questions a buyer might ask before they know which vendor to pick. You will ask ChatGPT these same questions every round, so the wording matters and it should not change once you begin.
About 12 prompts is a good starting size, covering category discovery, specific problems, use cases, requirements, and shortlisting. A smaller team can start with fewer, and twelve is not a statistically sufficient sample of the market.
Customer call notes, sales objections, and support questions are good sources for the wording. If you want help with finding the prompts to track, there is a full discovery playbook.
Here are the templates to adapt. Replace every placeholder in brackets before you test, because these are not measured popular queries.
| ID | Buyer intent | Prompt template |
|---|---|---|
| P01 | Category discovery | Which [category] tools should a [persona] at a [company size] consider? |
| P02 | Outcome | What software helps [persona] achieve [outcome] without [constraint]? |
| P03 | Feature requirement | Which [category] tools support [must-have capability]? |
| P04 | Budget | Which [category] options fit a budget of [amount] for [team size]? |
Carry on the same pattern for problem, integration, shortlisting, comparison, migration and replacement questions until you reach about 12. Use real budgets and real integrations, not constraints chosen to favor your brand.
Keep branded questions in their own group
Some questions name your brand, such as "What does [brand] do?" or "Compare [brand] with [competitor] for [use case]." They are useful, but they answer a different question. If the prompt contains your brand name, a mention can't tell you whether ChatGPT would have found you on its own. Label these as diagnostics and keep them out of your main results. There is a longer look at branded, unbranded and competitor prompts if you want to decide how many of each to run.
Give every prompt a stable ID and a version number. Once the baseline starts, freeze the text. If you rewrite a question, it becomes a new version and should not quietly replace the old one in a trend line.
DeepSmith's Discover Prompts can suggest candidate questions from your product, persona and buyer-stage context. You can use those as raw material, pick the ones that fit, and still run every check by hand. They are suggestions and not verified search-volume data.
Done when: every prompt has an ID, its exact text, an intent label, a group (core or diagnostic), and the market it is aimed at.
Common mistake: asking ChatGPT whether it recommends your brand, then reporting the resulting mention as if it were unprompted visibility.
2. Define your test conditions and build the sheet
Before you collect a single answer, write down how you will test. If you skip this, you won't be able to tell later whether a change came from your work or from a different account, mode or location.
Create a tab called Protocols and record these conditions:
- A protocol ID and the date you created it.
- Web, desktop or mobile, signed in or signed out, and the account plan.
- The model or mode label if ChatGPT shows one. If it doesn't, write "not shown" instead of guessing.
- Your search policy: Search selected by hand, or not selected.
- Your personalization choice, or the account settings you used instead.
- The language and market, and any location you put in the prompts.
- Where you tested, whether you shared device location, and any VPN change.
- Your brand aliases, product aliases, and the exact websites and subdomains that count as yours for citation credit.
- Your scoring rules, repeat count, test window, and who owns the work.
The workbook
Five tabs are enough: Prompts (the fixed question set), Protocols (conditions and scoring rules), Runs (one row each time you ask a prompt), Sources (page-level citation evidence), and Summary (comparable results). The Runs tab is where the work happens, so here is the layout. The formulas in Step 7 depend on these column letters.
| Column | Field | How to fill it in |
|---|---|---|
| A to D | Run_ID, Wave, Date_time, Prompt_ID | A unique ID, a consistent wave label, the time with its time zone, and one fixed prompt version |
| E to G | Cohort, Protocol_ID, Repetition | "core" for the unbranded set, the conditions you used, and 1, 2, 3 and so on |
| H | Status | valid, error, quota or incomplete |
| I | Brand_mention | 1 or 0 for valid runs, blank if unscored |
| J | First_recommendation_position | The number if there is a list, otherwise blank with a note |
| K | Owned_citation | 1 or 0 for valid runs, blank if unscored |
| L to O | Distinct_owned_pages, Competitors, Evidence, Notes | Pages cited, competitor names and positions, where the saved answer lives, and the exact mention excerpt |
Add data validation to the 1 or 0 columns and to Status. Mark a run valid only after you have scored both the mention and the citations, and keep failed attempts in the sheet instead of deleting them.
If you want to go further on the setup, there is a guide to a repeatable prompt set for monitoring AI visibility.
Done when: a second person could reproduce your conditions and read every column without asking you what you meant.
Common mistake: building a sheet with only a prompt, a yes or no for your brand, and a date, which leaves no way to separate accounts, modes or evidence standards.
3. Start a fresh, unpersonalized session
For each run, open a new conversation and control personalization on purpose. A new conversation alone isn't enough, because ChatGPT can still carry your memories and custom instructions into it.
The preferred route
According to OpenAI's help page, an Unpersonalized Temporary Chat does not use memory, custom instructions or plugins, while a personalized one can. So "temporary" by itself is not a clean test. On the web version:
- Open a new chat.
- Select the Temporary control, which OpenAI describes as a pill-shaped button in the top-right corner.
- Before you start, choose Unpersonalized when it is offered.
- Check that the right state is selected before you send the prompt.
You can't change the choice once the conversation starts, so if you picked wrong, drop that run and begin a new one.
If you don't see those controls
Controls vary by plan, region and workspace settings. For an account-based fallback, turn off custom instructions (Settings, then Personalization, then Custom Instructions, then switch Enable customization off) and check each memory control under Settings, then Personalization, then Memory, such as Memory, Reference saved memories and Reference chat history. Write down which fallback you used. For this audit, use plain ChatGPT, with no custom GPT, project instructions or uploaded brand files.
Location needs its own control
OpenAI says ChatGPT may estimate a general location from your IP address, and that sharing your device location is a separate option that is off by default. So turning device location off doesn't make the test location-free. Write the state down, and put any target market in the frozen prompt.
If the visible model, account, region, prompt text or a scoring rule changes between rounds, mark a protocol break instead of crediting your marketing.
Done when: the fresh session, your personalization state, your location conditions and the protocol ID are recorded before you send the first prompt.
Common mistake: assuming an empty chat history means a clean test. Check the settings, because that is where personalization comes from.
4. Run each prompt exactly as written
Send the frozen prompt without hints, background material or follow-up nudges. The point is to see what ChatGPT does with the question as a buyer would ask it.
Because citations are part of what you want to see, select Search deliberately. OpenAI's help page on searching the web documents this path: open the conversation, select View all tools, select Search, then type and send the prompt. ChatGPT can decide to search on its own, so not selecting Search doesn't prove that no search happened. If you also want to look at ordinary conversations, run that as a separate protocol with its own ID, and describe each set of results as visibility in the answers you sampled under that setup.
For each run, confirm the prompt ID and protocol, start the clean session from Step 3, select the mode your protocol calls for, paste the exact prompt once and let the answer finish, then capture and score the response before you start a new session for the next run.
A practical repeat count is three runs per prompt in each round, which makes 36 answers with 12 prompts. If that is too much, pick a smaller fixed set and document it. An audit of product recommendations reports that the products ChatGPT recommends can change across repeated requests. That supports checking more than one answer, but it doesn't give you a universal variability figure or a required number of repeats. If you want to know how automated tools do this differently, read how automated tracking collects ChatGPT answers.
Search is available to free users and to people who aren't signed in, with usage limits, and that is what lets you check ChatGPT visibility for free. If you hit a limit partway through, record the quota status, continue later under comparable conditions, and don't score the missing answer as zero visibility. If you would rather try a ready-made option first, there is a list of free ChatGPT visibility checkers.
Don't use Regenerate as a quiet replacement for an answer you didn't like. Log an accidental one as an extra run and keep the original.
Done when: every planned run has either a saved answer or a clear status such as quota or error.
Common mistake: asking "Why didn't you include us?" in the same chat and counting that follow-up answer in your visibility score.
5. Log mentions and first recommendation position
Now you code each answer with a short written rule, so that anyone looking at the saved answer would give it the same score.
The brand mention rule
Enter 1 when the body of the answer names your brand or an alias you approved before testing, and 0 when it doesn't.
- Several mentions in one answer still count as one mention-positive run.
- Save a short word-for-word excerpt so a reviewer can check it.
A mention doesn't mean ChatGPT endorsed you, so write incorrect descriptions and passing mentions in the notes column.
The position rule
When your brand appears as a recommendation in a list of products, record the first position it holds. Treat this as list position, and avoid calling it your rank in ChatGPT, since list order doesn't necessarily mean preference.
- Second product in a numbered shortlist: position 2.
- Named only in an opening paragraph, or named as an unsuitable choice: mention 1, position not applicable, with a note.
Done when: another reviewer can reproduce your mention score and position from the saved answer.
Common mistake: recording position 0 when your brand is absent. Use the mention score for that and leave the position blank.
6. Inspect citations and save the evidence
A website showing up on screen is not the same as ChatGPT citing it, so you need to look at the actual source links. If you want a wider definition, see what counts as an AI citation.
OpenAI says Search responses may include citations, which you can select to open the source. When a Sources panel is present, it holds the cited sources and also other relevant links. That difference matters, because a page sitting among the other links isn't evidence that ChatGPT used it to support the answer.
To inspect them, look at the citation markers and the passages they are attached to, open Sources if it is there, and work out which entries are cited sources and which are only extra links. Then open each source that might be yours, check where it goes, and record its title, its exact address, who owns it, and the passage it seems to support. If a citation doesn't support the statement beside it, record that separately from the fact that it was shown.
Set Owned_citation to 1 if at least one of your pages is identified as a source, and to 0 only after you have finished the inspection. If the answer or the panel fails to load, mark the run incomplete instead of guessing a zero. A few edge cases come up often:
- A third-party review that discusses your brand is a third-party source, and it is not an owned-page citation.
- The same owned page cited three times makes the run citation-positive once, with a distinct-owned-page count of one.
- A citation can appear while your brand is missing from the answer body, so mention 0 and citation 1 can both be correct.
Save the evidence before you leave
Save the full answer text, plus screenshots showing the prompt, the response, the recommendation list and the sources. Use a consistent file name such as Wave_PromptID_Repeat. Unsaved Temporary Chats can't be reopened later, and saving one turns it into a regular chat that follows your account's personalization settings, so keep your own copy of the evidence.
DeepSmith's AI Visibility Pages view shows which of your pages receive citations and which prompts drive them, which is useful for page-level review later. A citation appearing and a citation being correct are two separate observations, so log them separately.
Pro tip: save the answer before you close the temporary session. A score with no response behind it can't be reviewed by anyone else.
Done when: every valid run has saved evidence, and every citation-positive run points to an identifiable page you own.
7. Calculate your baseline without overstating it
Summarize only runs that can be compared, meaning the same wave and the same protocol. These are the four measures to use:
- Mention rate: mention-positive valid core runs divided by valid core runs.
- Citation rate: owned-citation-positive valid core runs divided by valid core runs.
- Mention prompt coverage: distinct core prompts with at least one mention divided by distinct core prompts tested.
- Citation prompt coverage: distinct core prompts with at least one owned citation divided by distinct core prompts tested.
Spreadsheet formulas
These use the Runs columns from Step 2. In the Summary tab, put the wave in A2 and the Protocol_ID in B2, then use C2 for valid runs, D2 for mention rate and E2 for citation rate, with D and E formatted as percentages.
C2: =COUNTIFS(Runs!B:B,$A2,Runs!F:F,$B2,Runs!E:E,"core",Runs!H:H,"valid")
D2: =IFERROR(COUNTIFS(Runs!B:B,$A2,Runs!F:F,$B2,Runs!E:E,"core",Runs!H:H,"valid",Runs!I:I,1)/$C2,"")
E2: =IFERROR(COUNTIFS(Runs!B:B,$A2,Runs!F:F,$B2,Runs!E:E,"core",Runs!H:H,"valid",Runs!K:K,1)/$C2,"")
A worked example with made-up numbers
These numbers are hypothetical and only show how the scoring works. They are not DeepSmith results, customer results or a market benchmark. All ten rows are valid core runs under one protocol.
Imagine ten valid core runs, two for each of five prompts. The brand is mentioned in five of the ten answers, at list positions 2, 3, 4, 1 and 2. Four of the ten answers cite an owned page. One of the ten answers (P05, first run) cites an owned page without naming the brand in the answer body, and P02 gives a mention in one run and none in the other.
That gives a mention rate of 50% and a citation rate of 40%. Three of five prompts produce at least one mention (60% coverage), and four of five produce at least one owned citation (80% coverage). The mixed P02 result is why you keep the repeats.
A fair way to report this is that in ten recorded answers across five fixed prompts under this protocol, the brand was mentioned five times and cited four times. Avoid saying your brand has 50% visibility across ChatGPT, since you only measured a sample.
Done when: the summary shows the prompt group, the number of valid runs, the protocol and the test period beside each rate.
Common mistake: mixing branded questions, unbranded questions and failed runs into one unexplained percentage.
8. Repeat on a fixed schedule and choose what to do next
One answer is not a verdict. Re-run the same prompt versions under the same protocol, then look for patterns that persist.
A weekly schedule is a reasonable place to start. That is a workload suggestion and not a known ChatGPT update cycle, and a slower schedule works fine if you keep it consistent. If you want to track visibility in ChatGPT over time, a steady rhythm matters more than a fast one. Four rounds of 36 answers make 144, which describes your workload and is not a count that guarantees a reliable market estimate.
If competitors keep appearing and you don't, look at those prompts and the sources cited before drawing conclusions. If your brand is mentioned but your pages aren't cited, look at which other sources support the answer. If results vary across repeats, report the spread and keep every run.
If repeated Search tests show no owned citations at all, one bounded technical check is whether your site is eligible. OpenAI says inclusion requires allowing OAI-SearchBot to crawl your site, and letting its published searchbot traffic through your host or content delivery network.
When the sheet starts to feel heavy, that is the point to think about when a spreadsheet stops scaling, and about choosing a ChatGPT tracking solution that fits your budget and team time. If you want to compare AI visibility tools, a baseline like this one tells you what you need them to do.
DeepSmith is one option for that ongoing view. It lets you review mention and citation trends, prompt-level answer history and the pages behind your citations in one place. A manual sample won't exactly match a platform metric, and neither is a census of every ChatGPT answer.
Done when: the next test has an owner and a date, the baseline is saved, and you have a short action list tied to specific answers or cited pages.
Common mistake: changing several prompts and settings after a weak result, then claiming the next percentage shows improvement.
What to do next
Save your baseline, put the next testing date on the calendar, and let repeated absences or citation patterns tell you where to look first. Once you have a manual baseline, you can start a DeepSmith seven-day free trial to review the prompts, answers and cited pages as you track visibility in ChatGPT over time.
Related guides
- How to Track Your Brand Visibility in Gemini: Set up Gemini tracking: choose prompts, run repeat samples, record mentions, sentiment and cited sources, and read results by topic so you can see where Gemini describes your.
- How to Track Your Brand Visibility in Perplexity: Track your brand in Perplexity by building a prompt set, running each one repeatedly, logging mentions and the sources cited, and comparing citation share against competitors over time.
- How to Monitor Claude Mentions and Referral Traffic: Monitor Claude by running prompt checks for brand mentions, then find referral traffic in analytics by identifying Claude referrers, tagging landing pages and separating it from direct visits.
- Per-Engine Trackers vs Cross-Engine Trackers: Which Approach Fits?: Decide between single engine and cross engine trackers by weighing depth of data, coverage, cost, reporting effort and your team's priorities, and pick the approach that matches how.
- How to Track Visibility in AI Shopping Results Across ChatGPT and Copilot: Track product visibility in ChatGPT and Copilot shopping results: choose product prompts, record which items and merchants appear, note listed attributes and prices, and repeat on a schedule.


