For twenty years, brand visibility in search had a comfortable ground truth. There was an index, a ranking, and query volume, so the question "did we show up for this term?" had a real answer you could pull from a report. AI-driven answers dismantle that comfort. The question is no longer "did we rank?" but "are we included, cited, summarised, or recommended in the answer the customer actually reads?" And that shift matters commercially, because a customer can now assemble a shortlist inside an AI response, complete with reasons and caveats, before they ever land on a website.
The uncomfortable truth is that this new question does not have the same kind of answer the old one did. There is no denominator. In search you were measuring against a known index and known demand. In AI answers you are estimating a distribution from samples, over a set of prompts you invented yourself. Every "AI visibility score" on the market, however slick the dashboard, is really a share of voice across the questions you chose to ask, on the days you asked them, under the sampling conditions you used. That is not worthless. It is simply a different kind of object to a search rank, and most programmes go wrong the moment they forget that and start treating a soft, sampled estimate as if it were a hard position on a page.
Within that constraint, though, you can build something genuinely defensible. It just requires being honest about what you are measuring.
The prompt set is the whole game
If AI visibility is share of voice across the questions you ask, then the questions you ask are not a detail, they are the entire foundation. This is keyword research without the volume data, which means the prompts cannot be brainstormed in a workshop. They have to be sourced from real demand: Search Console queries, sales call transcripts, support tickets, the questions people genuinely ask about your category on Reddit and in forums, and the landing pages that AI referrals are already hitting.
Then the set has to be segmented, because different kinds of questions behave completely differently and averaging them into a single number hides everything worth knowing. At minimum, separate category-entry questions ("best tool for small business payroll"), comparison prompts ("X versus Y"), brand-specific prompts, and objection or risk prompts ("is X actually reliable?"). A brand can dominate its own branded prompts and be invisible in the category-entry questions that decide who gets onto the shortlist in the first place. One blended score will never surface that.
Sample repeatedly, per platform, and log the version
AI outputs are stochastic. Ask the same question twice and you can get two different shortlists, so a single run tells you almost nothing. What you want is an inclusion rate expressed as a percentage across a decent number of runs, with a confidence interval attached, rather than a yes or no. Presence on one run is an anecdote. Presence on sixty percent of thirty runs is a measurement.
Platforms also have to be kept apart. ChatGPT, Gemini, Perplexity, Copilot, and Google's AI Overviews retrieve and reason differently enough that a blended cross-platform figure is mostly noise dressed up as insight. You are better off with five honest per-platform numbers than one tidy average that means nothing.
The quiet killer here is model versions. When a platform ships a new model, inclusion rates can move in a step change that looks exactly like your optimisation working, or exactly like it failing. Without logging the model version and keeping a control set of prompts you are not actively optimising, you will end up narrating coincidence as causation, taking credit for a vendor's release or blaming your own team for one. The version log and the control set are what let you tell the difference.
Presence is the least useful thing you will collect
Simply being mentioned is the weakest signal in the whole exercise. The metrics that actually carry information are the ones underneath it.
The first is citation versus mention. Are you the linked source the model is drawing from, or is a review aggregator the source and you are merely the subject being discussed? Those are completely different problems with completely different fixes.
The second is position and framing. Being named as the recommendation is not the same as being listed third out of eight, which is not the same as being the option that gets mentioned with a caveat attached. The rank-style thinking wants to collapse all of these into "present", and in doing so throws away the most important part.
The third is attribute association, meaning the qualities the model consistently attaches to your name. This is where a brand discovers that the model describes it as the cheap and cheerful option after three years and a fair amount of money spent positioning itself as premium. That gap is invisible on a presence dashboard and glaring the moment you read the actual language.
The fourth, and the most actionable, is source provenance: which URLs get cited when the model answers your category questions, whether or not you appear at all. You cannot optimise the model directly. You can influence the corpus it retrieves from, and that starts with knowing which sources it trusts for your category. Most programmes under-invest here precisely because it does not look like a scoreboard, but it is the closest thing to a lever you actually have.
Two validity problems worth naming out loud
There are two structural weaknesses it is better to state plainly than to paper over.
The first is that clean-room measurement measures a customer who does not exist. To get repeatable numbers you run fresh sessions with no memory and no personalisation. That is methodologically sensible, but it describes a synthetic user, and as assistants lean harder on memory, history, and context, the gap between your measured answer and your customer's real answer widens. Worse, you have almost no visibility into the size of that gap.
The second is that attribution is broken by design. A large share of AI answers produce no click at all, so the commercial effect does not appear in your referrer data. It surfaces instead as lift in branded search, direct traffic, and what people tell you when you ask them. The only practical instrument for the no-click path is a crude one: a "how did you hear about us?" field on your forms, paired with the visibility panel. It is not elegant, but it is the only thing that sees the part of the journey your analytics cannot.
What is achievable, and what is not
So, to the central debate about whether brands can build a reliable, scalable view of their AI presence, the honest answer splits in two.
Reliable relative and directional measurement is achievable. Share of voice against a named competitor set, trend over time, drift in the attributes the model associates with you, concentration in the sources it cites: all of this can be measured with a stated methodology and known error bars, and all of it is useful.
Reliable absolute visibility is not achievable, and probably never will be, because the thing being measured is personalised and non-deterministic by construction. There is no single true number for "how visible are we in AI", any more than there is a single true number for "what do people think of us".
The programmes that work treat this as panel measurement, much closer in spirit to brand tracking than to rank tracking. They publish their methodology, they show their confidence intervals, and they talk in terms of direction and relative standing. The programmes that fail are the ones that put a single, tidy number on a dashboard and let an executive read it as a rank. The measurement is real and worth doing. It just needs to be understood for what it is, and reported with the honesty that its foundations demand.
