To measure brand visibility in ChatGPT, save a fixed set of buyer questions and record whether each answer names your brand, which pages it cites and whether visitors arrive at your site. Keep those three measures separate. A mention can refer to your company while linking to somebody else’s page.
This guide gives you a small, repeatable measurement sheet. It covers ChatGPT search, with Bing and Google reports as separate sources of evidence. For the wider content and distribution work, use our guide to improving AI search visibility.
Decide which question you are measuring
Start with a business decision. A payments provider might ask whether it appears when buyers look for cross-border payment software. Its question differs from whether ChatGPT can find a release when someone types the company’s name.
Create separate groups for category discovery, comparisons and brand facts. Use wording a prospective customer would plausibly use. For example, “Which payment platforms support small online retailers in the UK?” is a category question. “What does [brand] offer?” tests an answer about a named business.
These are illustrative prompts, not results from a campaign. Replace them with your actual product, buyer and supported market. Keep brand names out of the category group so that prompting the answer with your name does not inflate the discovery result.
Write down the prompt set before collecting answers. If you change it halfway through, save the changed set as a new version and compare like with like.
Record the conditions as well as the answer
OpenAI’s search documentation explains that ChatGPT can rewrite a prompt into targeted searches. General location and, when enabled, memory can affect the search. A measurement sheet therefore needs more than a copied answer.
Record the engine and surface, whether search was used, language, stated market, account context and the time of the run. Note whether memory or a continuing conversation supplied relevant information. Use a consistent setup across your baseline and later checks. Avoid putting customer data or confidential launch details into a test prompt.
Save the answer and its visible sources together. If no sources appear inline, inspect the Sources panel where available. A response without a displayed link cannot be recorded as a citation to your website merely because it sounds like your copy.
| Field | What to enter |
|---|---|
| Prompt ID and version | The unchanged question and its category |
| Run context | Engine, search surface, language, market and time |
| Brand mention | Yes or no, plus the exact passage |
| Description accuracy | Accurate, incomplete or wrong, with the disputed detail |
| Sources | The displayed URLs, including third-party pages |
| Owned-domain citation | Yes or no; record the actual URL |
| Follow-up | A factual correction or source question to investigate |
Keep a raw record behind any summary. It lets someone check what you counted.
Calculate presence using a stated denominator
For your own sheet, define mention rate as answers that name the brand divided by recorded answers. Define owned-domain citation rate separately, using answers that link to your domain. State the number of prompts, repeats and answers each time you report either rate.
A worked arithmetic example: if a deliberately fictional sample contains 20 answers and five name the brand, the sample mention rate is 25%. If two link to its site, the owned-domain citation rate is 10%. Those figures demonstrate the calculation; they say nothing about a real brand.
Repeat the same questions across more than one collection window. Treat a small change as an observation to investigate. A single favourable answer gives you one answer, not a stable position in a recommendation list. Keep comparison prompts and branded prompts out of the same headline rate unless you explain the mixture.
Keep publisher dashboards in their own columns
Microsoft’s AI Performance announcement describes total citations, average cited pages, grounding queries and page-level citation activity across supported Microsoft AI experiences. It explicitly says page citation counts do not indicate ranking or placement. Grounding queries are a sample of citation activity.
Use that evidence to see which of your pages are being referenced on those surfaces. It is not a count of every answer that mentions your brand, nor a ChatGPT-wide visibility report.
Google’s AI features guide says appearances in AI Overviews and AI Mode are included in Search Console’s overall Web performance data. That documentation does not describe a separate ChatGPT measure. Do not add Google clicks, Bing citation counts and manual answer mentions into one “AI visibility” total.
Connect the answer record to visits and enquiries
Referral data answers another question: did somebody follow a link to your site? Read your existing analytics using its actual source classification and attribution rules. A citation without a visit still counts as a citation, while an enquiry may arrive through a different route after someone saw the brand.
Keep visits and completed enquiries alongside the answer observations. Explain the attribution window and any missing data. Comparing before and after a press release also requires a note of other changes: product launches, site edits, coverage and shifts in the prompt set.
A change after distribution is not proof that distribution caused it. Our distribution report checklist helps separate delivered placements from subsequent audience activity.
Questions to ask before buying a visibility report
Ask for the exact prompts, engine surfaces, markets, collection windows and raw answers. Ask how mentions and citations are defined, how repeats are counted and how disputed descriptions are reviewed. A useful answer lets you reproduce the calculation and see its limits.
Start with the free AI Shortlist Check if you want a first reading, or send a brief with the product and buyer questions you need to understand.