AI visibility audit: do AI engines name and cite your brand?

What are you actually counting: a mention, a citation, or a position?

We use the same definitions as our published method. Mentioned (named): the answer names your brand, including spelling variants agreed in advance. Cited: the answer links to, or lists as a source, a page on your domain. Ads and sponsored placements do not count. Citation share is the share of answers that name you or cite your domain.

These are different events. An engine can name you without linking to you, link to you without naming you in the text, do both, or neither. Record the two separately for every answer, then combine them for the headline. Position (where you sit in a list) is a third thing, and it is the least stable; see below.

Why can't you trust a single AI answer?

Ask an AI assistant the same question twice and you can get different companies and different links. Our method runs each prompt three times on each engine for that reason, and we say that three runs reduce the noise without removing it.

A published study points the same way. As reported by Search Engine Journal (we could not open the original SparkToro post, which returned an access error), SparkToro and Gumshoe.ai had volunteers run twelve brand-recommendation prompts, each 60 to 100 times per platform, on ChatGPT, Claude and Google's AI features, over November and December, ahead of a 30 January 2026 report. They reported that ChatGPT and Google returned the same brand list under 1% of the time and the same list in the same order under 0.1% of the time, that Claude repeated lists slightly more often but order less often, and that tighter categories were more stable than broad ones. Despite varied wording of human prompts, a small set of brands appeared in a large share of headphone responses.

Read that study with its own caveats. As summarised, it is not peer reviewed, it was done with Gumshoe.ai, a company that sells AI tracking, the authors called for larger follow-up work, and open questions include how many runs are needed and whether API results match manual prompts. It supports three working rules and no more: do not judge a prompt on one answer, report visibility as a rate over many answers rather than as a rank, and treat list position as unreliable.

How do you design the prompt panel?

The panel is the fixed list of questions you will ask. Our method uses 50, drafted with you, approved in writing and frozen, so a later measurement is a fair comparison. Prompts are phrased as a buyer would type them, and only a small share name your brand; a panel of questions that contain your name measures nothing about discovery. The prompt panel builder covers where to find the questions, the intent buckets, controls, freezing and versioning.

Fix the settings before the first run and write them down: country, language, device type, whether web search is on where the engine offers it, signed out with no memory or saved instructions, and the model or mode each engine shows on screen. The workbook has a Settings sheet and a column for the model or mode shown, because an engine can change under you during a measurement.

How many runs, on which engines, and how do you log them?

Results fetched through an API can differ from what the consumer app shows, so numbers from a tool and numbers from the app are not interchangeable. If you use a tracking tool, keep its export next to your saved answers and say which one a number came from.

1. Choose engines. Our method names ChatGPT, Perplexity, Claude and Google AI Overviews, and notes where Gemini is included. Pick the engines your buyers use and keep the list fixed.
2. Run each prompt 3 times on each engine, each in a fresh conversation. With 50 prompts and 4 engines that is 600 answers. We have not timed this for a manual run; budget for it.
3. For each answer, save the evidence the same day: a screenshot or share link, and paste its reference into the log. If the answer cannot be re-read later, do not count it.
4. Fill the row: engine, model or mode shown, prompt ID, run number, Mentioned Y/N, Cited Y/N, position if the answer is a list, competitors named, domains cited.
5. If Google shows no AI Overview for a run, our method counts that run as not cited and reports how often an Overview appeared at all. Log it so you can report that.
6. Do not re-run an answer you dislike. Re-runs chosen after seeing the result bias the rate.

How do you choose a competitor set and compute share of voice?

Pick 3 to 5 competitors that a buyer would plausibly see in the same answer. Include at least one you consider a rival and one a buyer might see as an alternative, and write down why each is in. The set is a choice, and share of voice depends entirely on it: a different set gives a different number. Freeze it with the panel.

The workbook computes share of voice as follows, and states it in the README so a reader can check it.

Share of voice (%) =
  answers naming your brand
  / ( answers naming your brand
      + sum over each tracked competitor of answers naming that competitor )
  x 100

Counts answers, not mentions inside an answer: a brand named twice in one
answer counts once. Matching uses text search on the competitor column, so
use distinct names (Acme and Acme Pro would be double counted).

What is a source-gap analysis and how do you run one?

A source gap is a domain that engines cite in answers that leave you out. The method comes from our own checklist (check 31: ask each engine your category questions and list the cited domains). The aim is a short list of third-party places to investigate, not a promise that appearing there raises your citations. We have no source showing that, so treat each action as a hypothesis to re-measure.

1. From the log, list the domains cited, one per row, in the Source gaps sheet.
2. The sheet counts, for each domain, how many answers cited it, how many of those answers did not name you and how many did. The first of these, cited where you were absent, is the gap score.
3. Classify each domain: your own site, review site, forum, news, competitor, directory, other.
4. Check whether you have a presence there, and whether it is complete and consistent with your site (checks 32 and 38).
5. Choose honest actions only: a real profile, real reviews from real customers, an accurate listing, original data others can cite. Never buy or fabricate reviews or create promotional entries on Wikipedia-style sites.
6. Re-measure on the same panel before claiming anything worked.

What should the report contain?

Link the report to the underlying log so any row can be opened and read. That is the standard we hold our own method to, and a number you cannot open is a claim, not a measurement.

Do not infer traffic from citation share. Google says traffic from AI Overviews and AI Mode is counted within overall Search traffic in the Performance report under the Web search type; the page we read describes no separate AI filter, and we did not check the live interface. Search Console shows clicks and impressions; it does not tell you what an AI answer said about you.

What can you do in 20 minutes?

A rehearsal, not a measurement. Pick 10 prompts from your buyer questions, none naming your brand. Open one engine in a fresh signed-out session. Run each prompt 3 times and log 30 answers in the Raw answers log of the workbook, with a screenshot reference. Read the Summary and the per-prompt Pattern column. You will learn how long logging takes, how often the answer varies for the same prompt, and which domains appear. Do not publish the numbers; 30 answers on one engine is a practice run.