AI search visibility: the metrics and KPIs that matter

Every metric of AI search visibility comes from one of three places, and each sees a different part of the picture. Knowing which is which keeps a report from adding up numbers that were never counted the same way.

What each engine’s report counts, and how to read them side by side, is set out in the guide to measuring results in AI answers. This guide is about what to compute from all three, and which figures to hold a team to.

Google Analytics 4 sorts visits into a default channel group, and one of its channels is AI Assistant: visitors who arrive from sources such as ChatGPT, Gemini, Deepseek, Copilot or Grok. A session lands there when its medium is ai-assistant, which GA4 sets, with the campaign (ai-assistant), when the referrer matches its list of AI assistants^[1].

Google’s own AI features are not in that channel. Clicks from AI Overviews and AI Mode are counted in Organic Search, with the rest of Google’s organic results^[1]. So the AI Assistant channel never holds a click from Google’s AI features, and Organic Search cannot separate them from classic results; count those in Search Console instead.

1. 1Report AI Assistant sessions by landing page, not only in total. How to tie them to revenue, and with which label, is set out in the guide to measuring results in AI answers.
2. 2Keep the channel’s figures apart from Organic Search in every table, and say beside each which AI features it includes and which it leaves out.
3. 3Compare the channel with a period of the same length: AI referrals are small for most sites, and a short window turns a handful of sessions into a swing.
4. 4Treat an assistant that GA4 does not recognise as missing, not absent. A referrer outside its list lands in another channel, usually Referral, the channel for links on other sites^[1], and is worth a look in the referrer report.

A metric is only useful when its denominator is fixed. Each one below is a count over a set you name in the report: the prompts tracked, the engines asked, the runs made. The unit of most of them is the mention or the AI citation.

Bing’s AI Performance report computes citation share the same way for each grounding query it reports^[2], so a share you compute from your own prompts can be set beside Bing’s for the same topic.

Placement has a research version. The paper that named generative engine optimization measured a source’s visibility as the share of an answer’s words that sit in sentences citing it, weighted down the later those sentences come, because readers read the start of an answer first. It called this the position-adjusted word count, and paired it with a subjective impression score that rates relevance, influence, uniqueness and the likelihood of a click, among other aspects^[3]. Both are hard to compute by hand, but the idea is sound: a citation beside the first sentence is worth more than one in a list at the end.

A mention rate is a share of a sample of answers, and a sample has a margin. Before a change goes into a report, check that it is larger than the margin of either month. The usual approximation for a share needs nothing but the rate and the number of answers:

margin = 2 × √(rate × (1 − rate) ÷ answers) rate 0.50, 20 answers   →  margin 0.22  →  0.28 to 0.72rate 0.50, 200 answers  →  margin 0.07  →  0.43 to 0.57rate 0.10, 200 answers  →  margin 0.04  →  0.06 to 0.14

With 20 answers, a brand named in half of them could, by chance alone, come out anywhere between about three and seven in ten next month without anything changing. With 200, the same rate is pinned to within about seven points either way. That is why a small prompt set can show a direction but not a trend, and why the margin belongs beside the rate in any report that compares months.

Count the margin on prompts, not on runs. Two or three runs of one prompt are not independent answers: they share the prompt, the day and the engine, and their answers tend to agree. A set of 30 prompts run three times gives 90 rows but evidence closer to 30 answers. Use the runs to settle each prompt’s result, such as named in two runs of three, and the prompts as the count.

Most of these metrics are diagnostics: they explain a change. A KPI is the one or two figures a team is asked to move, and it should follow from what the business wants from AI answers.

Whichever you choose, four rules keep a KPI honest:

1. 1Fix the prompt set, the engines and the number of runs for the period, and print them in the report.
2. 2Report the KPI by topic and by engine as well as in total: a total can hide a fall in one engine behind a gain in another.
3. 3Give every figure its base and its margin. Five answers of ten and 500 of 1,000 are both half, and the larger base is stronger evidence.
4. 4Keep scores a tool reports about itself, such as a visibility index, as diagnostics until you can check them against the answers they were computed from.

A KPI that lives only in a dashboard drifts: a prompt is added, an engine is dropped, the runs change, and the line on the chart no longer measures what it did. Write each one down as a short definition, agreed before the first month is counted, and change it only with a dated note.

name:         Mention rate, category promptsformula:      prompts whose answer names the brand ÷ prompts askedprompt set:   category-v3, 40 prompts, frozen 2026-10-01engines:      ChatGPT, AI Mode, Perplexity, Copilotruns:         3 per prompt; a prompt counts as named in 2 runs of 3source:       tracking sheet, answers kept with their sourcesbaseline:     October 2026margin:       shown beside every monthowner:        content leadreview:       every three months, with the prompt set

The definition is also the answer to the first question anyone asks about a number: what exactly does it count? A KPI that cannot be read back to its prompts and answers is a score, not a measurement.

How to read a change once the months come in, with its likely cause, and how to lay the figures out for a client or a board, are the subject of how to measure GEO results. How to collect the answers in the first place is covered engine by engine: how to track your brand in ChatGPT, and how to track AI Overviews for Google.