A screenshot of one AI answer can make a program look like it’s working, or like it’s failing. Neither tells you much. Ask the same question again tomorrow, or in a different engine, and the answer may change.
That is why we measure AI visibility the way we do: a fixed set of buyer questions, the same five engines, a dated baseline, and the same questions run again every month. This article explains the five measures we report, how each one is counted, and where each one stops being reliable.
The short answer: We track Share of Model, Citation Velocity, Positioning Read, Search vs Model Signal and AI-Referral Traffic. Each answers a different buyer question: are we showing up more often, is AI describing us accurately, and is it bringing qualified traffic. None of them proves on its own that AI visibility caused revenue, and the limits of each measure are set out below.
Start With a Fixed Question Set
Every measure depends on the questions behind it. If the questions change from month to month, a rising number can mean the yardstick moved rather than the results.
So the question set is agreed at the start and frozen. It covers two kinds of question. Buyer questions are what a prospect asks when evaluating a category, such as which vendors to shortlist or how two options compare. Build-decision questions are what a developer or operator asks when an AI agent is helping them choose a tool, library or service.
The questions are run across the same five engines: ChatGPT, Perplexity, Google AI Mode, Claude and Gemini. The first full run is dated and kept as the baseline. Each monthly run after that uses the same questions on the same engines, so every comparison is like for like.
If the question set ever has to change, for example because you launch into a new category, the change is recorded and the new questions get their own baseline. Comparing a new set against an old baseline would produce a number that looks precise and means very little.
Why One Answer Is Never the Result
AI answers are generated, not looked up. The same question can produce a different answer on a different day, in a different session, with search on or off, or after a model update. A brand can appear in one response and be missing from the next without anything on its website changing.
We treat each answer as an observation, not a verdict. Each one is recorded with its date, the engine, the question, and whether the engine searched the web or showed sources for that response. Results are reported as counts across the whole question set, and month-over-month movement is read as a trend rather than a promise. A small change between two runs can be ordinary variation.
The Five Measures
1. Share of Model
What it answers: Are we showing up more often than before, and more or less often than competitors?
How it is counted: The unit is one answer: one question, on one engine, in one monthly run. The numerator is the number of answers in which your brand appears. The denominator is every answer collected for that question set, on that engine, in that run. Competitors are counted over exactly the same answers, so the comparison is fair.
We report Share of Model within each engine first, and only then as a combined figure. A combined number can hide the fact that you are strong in one engine and absent from another, and visibility in one engine does not carry over to the others. We cover this in more detail in our article on why AI visibility does not transfer between engines.
Its limits: Appearing in an answer is not the same as being recommended. Share of Model counts presence. The other measures explain what that presence is worth.
2. Citation Velocity
What it answers: Are we gaining or losing ground as a cited source over time?
How it is counted: Citation Velocity is the change between monthly runs, measured against the dated baseline and the previous month. It is calculated on the same question set, the same engines and the same denominator as the figures it compares, and reported as a change in percentage points per engine.
Its limits: A single month’s movement is a weak signal on its own, because answers vary between runs. Velocity becomes useful when the direction holds across several runs. It also shows timing, not cause. If visibility rises after a piece of work ships, that is worth reporting, but timing alone does not prove the work caused it.
3. Positioning Read
What it answers: Is AI describing us accurately?
How it is counted: The unit is each answer in which your brand appears. For each one we record how the engine describes you: the category it puts you in, the use cases and buyers it associates with you, and any limitations it repeats. That description is compared with how you actually position the business. We also capture the sources the engine cites or links alongside the description, because the source often explains the framing. An outdated review, a competitor’s comparison page or an old directory listing can shape the answer more than your own site does.
Its limits: Positioning Read is a structured judgement, not an automatic score. It is reported as findings with the answers that support them, so you can check them yourself. When an answer cites no sources, we record that it cited none. We do not guess which material the model drew on.
4. Search vs Model Signal
What it answers: Does your presence depend on the engine searching the web at the time of the question?
How it is counted: We compare answers produced under two observed test conditions. In the first, the engine visibly retrieves from the web for that response, usually shown through search activity or linked sources. In the second, no retrieval is observed. The measure is only reported where the condition can actually be verified for that engine and response. Where it cannot, the answer is left out of this comparison rather than assigned to a side.
Its limits: This measure describes test conditions, not what is inside a model. An answer that mentions you without searching does not prove your content is in the model’s training data, and an answer that leaves you out does not prove it is absent. What the comparison can show is whether your visibility holds when live retrieval is not part of the answer, or whether it relies on it.
5. AI-Referral Traffic
What it answers: Is it bringing qualified traffic?
How it is counted: The unit is a visit recorded in GA4 whose referral source is an AI engine, together with the conversions those visits complete, such as a demo request or a form submission you already track. It is reported separately from the answer-based measures, because it is measured on your site rather than in the engines.
Its limits: Some visits that start in an AI answer arrive without referral information, for example when someone copies a link or opens it from an app, and those visits show up as direct or unattributed. AI-Referral Traffic therefore undercounts the real total. It also counts visits and conversions, not incremental revenue. A visitor who came from an AI answer might have found you anyway, so referral numbers show contribution to activity, not proof of new pipeline.
Comparing Competitors Accurately
Competitor numbers are only useful when they are measured the same way as yours. That means the same questions, the same engines, the same run dates and the same denominator. A competitor figure taken from a different question set, or from a single screenshot, cannot be compared with your Share of Model.
We also report how competitors are described and which sources are cited for them, not only how often they appear. A competitor that appears less often but is consistently recommended for the requirements your buyers care about may be the bigger threat.
An Illustrative Example
This example is hypothetical. The numbers are invented to show the arithmetic and are not taken from any client.
Imagine a fixed set of 40 buyer questions run on five engines. That gives 200 answers in the baseline run. Your brand appears in 30 of them, so your combined Share of Model is 30 out of 200, or 15%. A competitor appears in 50 of the same 200 answers, or 25%.
Split by engine, the picture changes. Your 30 appearances are 12 of 40 answers in ChatGPT, 9 in Perplexity, 5 in Google AI Mode, 3 in Claude and 1 in Gemini. The combined 15% hides the fact that you barely appear in two engines. That is where the work should go first.
The next month, using the same 40 questions on the same five engines (200 answers again), suppose you appear in 34 answers, or 17%. That is a two-point change. On its own it could be ordinary variation. If the next runs keep moving in the same direction, Citation Velocity starts to mean something. If the Positioning Read also shows the engines describing you in the right category, and AI-Referral Traffic shows more tracked visits that convert, the three buyer questions are being answered together.
What a Monthly Report Should Let You Do
A good measurement report should let you check the evidence yourself. Every number should come with its denominator, the engines and dates it covers, and the answers behind it. Every claim of improvement should be compared against the dated baseline. And every limit, from answer variation to unattributed visits, should be stated rather than smoothed over.
That is the standard behind our AEO and GEO program. It is also why we look past the brand name to the substance of the answer, as we discuss in ChatGPT brand mentions vs. recommendations.
See Where You Stand Today
If you want a starting point, the Free AI Visibility Audit runs your brand across all five engines using real B2B buyer queries. It shows where you appear, how you are described, who is being cited instead of you, and the biggest gap to fix first, delivered in 48 hours.