Reference
How the measurement works, and where it stops.
If you are going to put an AI visibility number in front of a client, you should be able to explain how it was produced and what it does not mean. This page is that explanation, including the parts that are unflattering.
Vocabulary
Five words that are not interchangeable.
Most confusion about AI search measurement is really a disagreement about which of these someone means.
- Mention
- The client’s brand name appears anywhere in a sampled answer, whether or not it is linked or recommended.
- Position
- Where the client falls in the order of brands named. First, second, third or later. A client named last in a list of six is not in the same situation as one named first.
- Citation
- A page on the client’s own domain was linked as a source for the answer. Being cited is materially stronger than being mentioned, and the two are always reported separately.
- Visibility
- The share of sampled answers, over a window, in which the client was mentioned. A percentage, always reported with the observation count behind it.
- Share of voice
- The client’s mentions as a proportion of all mentions of the named competitor set, within the same sample.
Sampling
There is no index of AI answers. So we run the questions.
Nobody can crawl an answer engine’s results the way a rank tracker crawls a results page. The only honest way to measure is to ask, repeatedly, and count.
What is sampled
A defined prompt set per client, selected with your team from the questions that actually decide the category, not an automatically expanded keyword list.
How often
Every prompt is checked daily on ChatGPT, Google AI Overviews and Gemini, and weekly on Perplexity, Claude and Copilot, the engines that answer less of the market today. Technical crawls run weekly by default and can be triggered on demand.
Where from
Prompts are run against a stated geography, because answers differ by region. The region is part of the record, not a setting that silently changes the numbers.
What is stored
The response text, the brands named, the order, the cited sources, the engine, the model version where exposed, and the timestamp. Historical records are kept intact.
- ChatGPT+14.2%72%
- Google AI Overviews+8.6%61%
- Gemini+11.4%49%
- Perplexity−3.2%43%
- Claude+5.9%37%
- CopilotLimited coverage+2.4%29%
The limits
What we will not claim.
Every one of these is a place a competitor could quote a more impressive number than we will. We would rather lose that comparison than the client conversation six months later.
01A single observation proves nothing
Generative answers vary between users, sessions, regions and model versions. One run showing a client absent is not evidence of a problem; the same prompt absent across a fortnight of runs is. Everything on this platform is reported over a window for that reason.
02Accurate as a trend, approximate as an absolute
A visibility figure of 68% should be read as ‘consistently named in most answers’, not as a precise measurement of reality. The direction of travel over consistent sampling is the reliable part, and it is the part worth putting in front of a client.
03Coverage differs by engine, and changes
Each provider exposes a different amount. Some link every source, some link none, some restrict automated access entirely. Where an engine limits what can be observed, the interface says so rather than filling the gap with an estimate.
04Correlation is not attribution
When visibility moves after a piece of work ships, we show both on the same timeline. We do not claim the work caused the movement, because a competitor launch or a model update in the same fortnight is equally consistent with the data.
05Nothing here is a ranking guarantee
Being cited is a decision an engine makes. Opportunities are phrased as eligibility, meaning the conditions under which a client can be cited, because that is the part anyone can actually influence.
When an engine changes
The ground moves. The history should not.
A model update can shift a client’s visibility overnight without anything about the client changing. Distinguishing that from real movement is most of what a measurement layer is for.
- The measurement window is versioned, so figures from before and after a change are never silently averaged together.
- Timelines are annotated where we detect a step change in an engine’s behaviour across the whole client base rather than in one account.
- Historical records are kept intact and are not retroactively recalculated, so a report issued in March still says in September what it said in March.
Find out which of your clients AI is skipping.
Book 30 minutes. Name one client, and we will show you their real results on the call.
Name one client
Tell us which client you want to see.
We check it first
We sample their AI answers before the call, so the numbers are real.
30 minutes, live
See where they stand and what to fix first.
No slide deck and no contract on the call. If it fits, we run a guided pilot on one of your real clients before you pay for a full plan.