Measuring AI visibility

AI visibility is measured as rates counted over a fixed set of recorded AI answers. Every figure a tracker reports belongs to that set, and changing the questions, the engines, or the number of runs changes the figure.

Most of the measurement work happens before the first run, in deciding what goes into the set.

SEO practice carries over in one respect: a fixed list of queries, tracked over time. It misleads in another, because a ranking holds still between checks and an AI answer does not.

An AI answer is written fresh every time it is asked, a property the glossary calls non-determinism. One answer is one draw, and a brand's visibility is a rate across many draws.

Which questions belong in the prompt set?

A prompt set is the fixed list of buyer questions the measurement runs. It does the job a keyword list does in rank tracking, with one difference.

Buyers ask AI engines in full sentences, and they often include the constraint that decides the purchase: a team size, a budget, a system the product has to work with. The set should read like those questions.

Three sources keep the set close to what buyers actually ask:

• Search demand. Questions built on terms with real search volume keep the set from drifting toward questions nobody asks. Search volume shows that buyers ask a question. It says nothing about what engines answer, so the two are kept as separate signals.
• Sales and support. The questions buyers raise on calls, and the objections that stall deals, are the ones that decide shortlists.
• The buying path. Questions at each stage: the problem before the buyer knows the category, the category itself ("best tool for a given job"), and the comparison between finalists.

Once the set is chosen, it stays fixed. Prompt tracking runs the same set on a schedule, and a comparison over time holds only if the questions did not change.

When the set has to change, the change is recorded in the method note, the short statement that travels with every report and says how its numbers were produced. The new set then gets its own baseline.

The count that decides precision is answers, covered below. The prompt count follows coverage: each question that could change a shortlist appears once, with deliberate variants where the phrasing is likely to change the answer.

Which questions come out before anything is counted?

This step is the one most measurements skip, and it decides whether the headline rate means anything.

Before any rate is computed, including the baseline, the brand's own website gets read first. Its navigation is the most reliable statement of what the brand sells and to whom. Three kinds of question then come out of the set:

• Categories the brand does not sell. A skincare brand with no fragrance line is not losing fragrance questions.
• Audiences it does not serve. A brand with no men's range has no gap on a men's grooming question. That question sits outside its market.
• Filters it cannot meet. A question that asks for a brand from a particular country, or for a feature the product does not have, rules the brand out by its own wording.

Left in, these questions drag every rate down with answers the brand could never win, and they send the team after buyers it does not serve. Taking them out does not flatter the numbers, provided the method note states what was removed, what share of the set it was, and why.

The rule has a second half. A category the set never covered is reported as unknown, never as zero.

A blank where a category should be means the tracker never asked. It is not evidence that the brand is absent, and a reader who sees a zero will treat it as a finding.

What is every rate counted over?

Every rate is counted over the tracked answer: one recorded answer, from one engine, to one prompt, on one run. A mention rate of 30 in 100 means the brand appeared in 30 of 100 tracked answers, and the answer count sets how much weight that figure can carry.

Answers multiply quickly: prompts, times engines, times repetitions, times days. That helps precision up to a point.

Ten answers to the same prompt on the same engine are repetitions of one question, not ten separate observations. They show how much that one answer varies.

They say nothing about the questions the set left out. A large answer count drawn from a few prompts is less precise than it looks.

Reports therefore lead with answers, and the method note discloses how the answers were produced: the number of unique prompts, the repetitions per prompt on each engine, the engines run, and the dates. Without that note a reader cannot tell whether a large answer count came from as many different questions, or from a tenth as many questions asked ten times each.

Engine coverage belongs in the same note. Engines differ in how they find sources and how readily they mention brands, so each engine is reported on its own first. A brand can be strong on one engine and missing from another, and a blended figure hides both facts.

A cross-engine figure is useful only when it states how the engines were weighted, for example by each engine's share of traffic in the category. It is read next to the per-engine rates, never instead of them.

What does each of the four signals show?

Four signals describe a brand's place in a set of answers. Each one answers a different question.

The signals move independently, and the disagreements between them carry the information. A high mention rate with a late average position describes a brand that engines know but rarely recommend first.

A strong mention rate with a low citation share describes a brand discussed on the strength of other sites: its own pages are not the evidence engines use. Share of voice can rise while mention rate stays flat, because competitors dropped out of the answers rather than because the brand gained ground.

For that reason the four are reported side by side and never weighted into one score. A composite score moves when any input moves, so it cannot say which one did.

Its weights are a judgment presented as a measurement. And two brands with the same composite can need opposite work. AI Search Intelligence metrics that matter takes the signals one at a time.

How do citation rate and citation share differ?

Two citation measures sound alike and answer different questions. Citation rate counts answers: the share of answers that cite at least one of the brand's pages. Citation share counts citations: the brand's portion of every source cited across the same answers.

An illustration with round numbers. A set of 100 answers cites 500 sources in total. A brand cited once in each of 20 answers has a citation rate of 20 in 100 and a citation share of 20 in 500.

If the engines begin citing fewer sources per answer, the brand's citation share rises with no change in how often it is cited. If they cite more, citation share falls while citation rate holds.

Citation rate shows whether the brand's pages are used as evidence at all. Citation share shows how much of the evidence pool the brand holds against competitors and third-party sites.

Either one reported alone, or without the count it was divided by, invites the wrong conclusion. Engines also differ sharply in how many sources they cite and which kinds, as Gemini, AI Overviews, and AI Mode cite differently shows, which is one more reason citation figures are read per engine.

When is a change a real change?

A baseline is the first full measurement, taken before any work meant to change it. Later runs compare with it only if they repeat its conditions: the same prompt set, engines, repetitions, and collection rules. How to baseline AI visibility before buying a tool sets one up by hand.

The baseline also has to capture how much the answers move on their own. Answer volatility is that movement: the same prompt, asked again, comes back with a different set of brands and sources.

A change between two runs counts as a finding only when it is larger than the movement seen with nothing done. Two working rules follow from that:

• Each prompt is repeated inside the baseline period, so the spread is measured rather than assumed.
• Periods are compared, not single runs. A week of answers carries more than a day, and a single answer is a spot check.

The answers a claim needs depend on the size of the change it claims. A move from 20 in 100 to 40 in 100 shows up in a modest sample. A move from 20 in 100 to 22 in 100 needs far more answers than most tracking runs collect.

When the volume is not there, the honest report is that no change was detected.

Which competitors belong in the benchmark?

Competitive benchmarking runs competitors through the same prompts and engines, so that the results compare. The competitor list comes from the answers as well as from the sales team.

Engines often mention brands a team does not think of as rivals, and those brands are taking shortlist places. After the baseline, any brand that appears often joins the list.

Competitors follow the same scoping rules as the brand. A rival that sells a category the brand does not should not be scored against the brand on that category's questions. How to audit competitor AI visibility maps which questions each rival holds.

What can AI referral traffic show, and what can it not?

AI referral traffic is the visits that arrive from a link inside an AI answer, and analytics tools can separate it from other sources. It is real and worth watching. It also sits on a different layer from everything above.

The four signals measure the answer: what the engine said and which sources it used. Referral traffic measures what some readers did next.

Many readers of an answer never click, so a brand can be mentioned often and receive little traffic. A cited page can draw visits from answers that never mention the brand. And some visits from AI tools arrive with no referrer at all and are counted as direct traffic.

The two layers can point the same way, and when they do, they are consistent with each other. Neither one proves the other. They are reported side by side, and no figure is computed across them: a cost per AI mention built from traffic data mixes two measurements that were never taken on the same thing.

What does measurement not show?

Measurement shows where a brand stands and whether that changed. It does not show why. The causes sit in what engines can read, which the next chapter on technical access covers, and in what they choose to cite, which the chapters on owned content and earned authority cover.

Zumi tracks a brand's prompt set across AI engines and reports mention rate, share of voice, average position, and citation share side by side. Book a demo