How to check if ChatGPT mentions your brand: a manual audit

To check if ChatGPT mentions your brand, run a fixed set of real buyer questions several times each, in clean sessions, and log the results. One question asked once proves nothing, because AI answers change from run to run. Record whether you were mentioned, whether the facts were right, which sources were cited, and the location and language of each run. Then repeat the same test monthly.

This is a practical how-to. For definitions of the metrics you will calculate from the log, see our guide to AI share of voice.

01Why can't I just ask ChatGPT about my brand once?

Because AI answers are probabilistic. In a January 2026 study, SparkToro and Gumshoe had 600 volunteers run 12 identical prompts across ChatGPT, Claude and Google's AI, collecting nearly 3,000 responses. The chance of getting the same list of brands twice was under 1 in 100, and the same list in the same order about 1 in 1,000.^[1]

The researchers concluded that tracking a brand's position in a list is close to meaningless, and that how often a brand appears across many runs is the more reliable measure; some brands appeared in 60–90% of responses for a given intent.^[1] Your manual audit should be built around that idea.

02How do I build the prompt set?

Use the questions buyers actually ask, not the ones that flatter you. A prompt that contains your brand name only tests whether the engine knows you; a category prompt tests whether it recommends you. You need both.

Aim for 20–30 prompts for a first audit of one product line in one market. Source them from sales call notes, support tickets, site search logs and the questions in your FAQ, and write them as a customer would type them, including typos and informal phrasing where that is realistic.

03Which AI engines should the audit cover?

ChatGPT is the obvious start, but buyers spread across tools. For most enterprises, a first audit should include ChatGPT, Google AI Overviews or AI Mode, Gemini, Perplexity and Microsoft Copilot, plus any engine that matters in a specific market. Run the identical prompt set on each so results are comparable.

Expect differences. Engines use different indexes, different sources and different models, so one can recommend you while another ignores you. Those differences are findings, not noise: they usually point to a specific source or access problem on one platform, such as a blocked crawler covered in our guide to Perplexity not citing your website.

04How do I set up a clean test session?

Personalization can skew results. ChatGPT's memory and custom instructions can shape answers, and OpenAI says a temporary chat does not create or update memories, though it can still use existing memories and custom instructions if personalization is on.^[2] For an audit, turn off memory and custom instructions as well, or use an account that has neither.

1. 1
   Use a clean account or temporary chat

   Disable memory and custom instructions, or use a dedicated audit account. Do not use the account of someone who works on the brand.
2. 2
   Fix the engine and mode

   Record the model and whether web search was on. Answers with and without search can differ a lot.
3. 3
   Control location

   AI search features can localize answers. Run from the market you care about, or record the location you ran from.
4. 4
   Control language

   Run each market's prompts in its buyers' language. An English audit says little about Arabic or German answers.
5. 5
   Open a new chat per run

   Do not ask follow-ups in the same thread; earlier turns change later answers.

For Google AI Overviews and AI Mode, use a signed-out browser window and note the country. Treat each engine separately; a brand can be strong in one and absent in another, as we explain in why Google rank does not equal AI answer.

05How many times should I run each prompt?

There is no official number. In practice, three to five runs per prompt per engine is a reasonable floor for a manual audit, enough to separate "never mentioned" from "sometimes mentioned" without consuming a week. Spread runs over different days if you can, since sourcing and models change.

The SparkToro data is a reminder of the limits: with a handful of runs you will get a rough mention rate, not a precise one.^[1] Report it as a range or a band (rarely, sometimes, usually) rather than a figure with false precision.

06What should I record for each run?

Log every run in one row. If it is not logged, it did not happen, and you cannot compare next month with this month.

Judge accuracy against an approved fact sheet, not from memory. Before the audit, ask product, pricing and legal owners to confirm the current facts for each prompt group, and mark an answer wrong only when it contradicts that sheet. It saves arguments later and turns each error into a ticket with a clear owner.

07How do I turn the log into findings?

Group the log by prompt type. For each group, calculate how often you were mentioned across runs, how often the facts were right, and which competitors and sources appeared most. Three outputs usually matter most:

1. 1Gaps: high-value prompts where you are rarely or never mentioned.
2. 2Errors: answers that mention you with wrong facts, each traced to a cited source.
3. 3Source map: the domains that appear most often in answers for your category.

Share the findings with the owners who can act on them: errors to the content or product owner, source gaps to PR and partnerships, access problems to web and security. A finding without an owner tends to reappear unchanged in next month's audit.

If you are absent for most category prompts, work through our diagnosis of why ChatGPT doesn't mention your brand before changing content.

08When should I move from a manual audit to a tool?

A spreadsheet is fine for one market, one language and a few dozen prompts. It breaks down when you need daily runs, many engines, several markets and trend lines. That is where tracking platforms earn their cost. Typical options, as of October 2026:

Other established options include Peec AI, Conductor, BrightEdge and AthenaHQ; check current pricing on each vendor's site, as plans change often. Whichever you pick, a tool measures; it does not fix. Someone still has to trace sources and ship changes through your CMS, PR and legal teams.

That gap is where OwnlyReal works: tool-agnostic, using the measurement platforms that fit, and acting on the data through your existing approval chains.