TL;DR
"AI search visibility" is usually reported as one number. It is three different measurements, with three different denominators, controlled by three different parties. Entity resolution asks whether an AI engine knows who you are. Prompt-set citation asks how often you are named inside a set of questions somebody chose. Editorial inclusion asks whether the people who compile buyer-facing recommendation lists name you. They are won separately, in that order, and the commercial value rises as your control falls. A report that does not say which layer it is measuring cannot tell you whether you are winning.
The problem with "visibility" as a single number
A GEO report that says "your AI visibility is up 40%" has not told you anything yet. Up against what? Across which questions? Compared to whom?
Every visibility figure is a fraction. The number on top is how often you appeared. The number underneath — the denominator — is what you appeared out of. Most reporting in this category is careful about the top number and silent about the bottom one, and the bottom one decides what the figure means.
Change the denominator and the same business can be "cited in half of all answers" and "absent from every list a buyer reads" in the same week, with both statements true. That is not a contradiction. It is two measurements of two different things wearing the same name.
Layer 1 — Entity resolution: does the machine know who you are?
The question: when someone asks an AI engine about your business by name, does it describe you — or a similarly named company, or a blend of several?
The denominator: your own name.
Who controls it: you, almost entirely. Consistent legal name, schema markup, a canonical URL, a Wikidata entry, the same address and description everywhere you are listed.
What it is worth: it is necessary and nowhere near sufficient. Nobody buys from you because an engine spelled your name correctly. But every other layer fails if this one does, because an engine that confuses you with someone else will credit your work to them.
How to check it yourself: ask two or three AI engines "What does [your business name] do?" and "How much does [your business name] charge?" Check the founding year, the services and the prices against your own site. If a same-name company's details appear, you have a Layer 1 problem and nothing further up is safe to measure.
Where we are: this is the layer we have worked hardest on, because a business called "The Metavision" shares its name with several unrelated companies. As of mid-September 2026, a branded query returns our own site first, with our legal name, founding year, services and published prices attributed correctly and the other companies attributed to themselves. You can check that today in under a minute.
Layer 2 — Prompt-set citation: how often are you named, inside questions you chose?
The question: across a fixed list of prompts, run repeatedly, how often does an engine name you — and how often does it name the competitors you are tracking?
The denominator: the prompt set and the competitor set. Both are chosen by whoever configured the tool.
Who controls it: you, substantially. You pick the questions. You pick who you are compared against.
What it is worth: it is a genuinely useful diagnostic. Tracked over time with the same prompts, it shows whether changes to your pages move how engines answer. What it is not is a market position. "We held half the share of voice" means half of the mentions, across the questions we wrote, against the companies we picked, in the window we measured. Written without those bounds, it reads as a claim about the whole market that the instrument never measured.
The three bounds every Layer 2 figure needs. If a figure is missing any of these, treat it as unreadable, whoever it came from:
- The prompt set. How many prompts, and are they buyer language ("best GEO agency UK") or brand language ("what does [agency] do")? Brand-language prompts inflate the result for the brand being measured, because the question already contains the answer.
- The competitor set, named. Share of voice is share of that set. It is never share of a country or a category.
- The window, and the date it closed. Answers change as engines re-crawl. A figure without a date is a figure without an expiry.
Where we are: we do not publish a Layer 2 figure for ourselves. The prompts and the comparison set would be ours, and a figure we configured is exactly the kind of number this article is warning you about.
Layer 3 — Editorial inclusion: do the list-makers name you?
The question: when a buyer searches "best GEO agencies UK" and reads the lists that come back, are you on them?
The denominator: what other people publish, without your involvement.
Who controls it: they do, entirely.
What it is worth: the most of the three. This is where buyers shortlist. AI engines frequently answer category questions by summarising exactly these lists, so a business absent from them tends to be absent from the answer as well.
What makes it hard: before you judge whether you should be on a list, check who publishes it. Some recommendation lists are independent editorial work. Others are published by a business that sells in the same category, and some inclusion is sold rather than earned. Neither is illegitimate. They are different things, and a report that counts them together cannot tell you which kind of visibility you have.
Where we are: not there yet. We are not named on the "best GEO agency" lists we track. We are saying so because this article would be worthless if we measured everyone except ourselves.
Why the order matters
The three layers are ordered by two things moving in opposite directions: your control falls and the commercial value rises.
Question
Layer 1 — Entity resolution: Does the machine know who you are?
Layer 2 — Prompt-set citation: Within chosen questions, how often are you named?
Layer 3 — Editorial inclusion: Do list-makers name you?
Denominator
Layer 1 — Entity resolution: Your own name
Layer 2 — Prompt-set citation: A prompt set and competitor set someone chose
Layer 3 — Editorial inclusion: What others publish independently
Who controls it
Layer 1 — Entity resolution: You, almost entirely
Layer 2 — Prompt-set citation: You, substantially
Layer 3 — Editorial inclusion: Them, entirely
Commercial value
Layer 1 — Entity resolution: Low: necessary, not sufficient
Layer 2 — Prompt-set citation: Low to moderate: a diagnostic
Layer 3 — Editorial inclusion: High: where buyers shortlist
It is entirely normal for a business with careful technical work and no track record in the press to score well on the first two and not at all on the third. If that is you, the answer is not more schema. It is the slower work of becoming worth citing: publishing things other people can check, and being findable when they go looking.
Five questions to ask of any GEO report
Whether you are reading an agency's report, a tool's dashboard, or your own spreadsheet:
- Which layer is this figure measuring? If the report cannot say, the figure is not usable.
- Who chose the prompts, and are they in buyer language or brand language?
- Who is in the comparison set, and who chose them?
- What window does this cover, and when did it close?
- Is any Layer 2 number being described in Layer 3 language? "Most-cited", "number one" and "leading" are market claims. A prompt-set tool cannot support them, however good the score.
The fifth question is the one that catches most over-claiming in this category, including, if you apply it honestly, the kind a business is most tempted to make about itself.
FAQ
What is GEO visibility?
It is how often, and how accurately, generative AI engines such as ChatGPT, Perplexity and Google's AI Overviews name and describe a business when answering questions. It is best measured as three separate layers — entity resolution, prompt-set citation and editorial inclusion — because each has a different denominator and a different owner.
Is AI share of voice a reliable metric?
It is reliable as a diagnostic, and only within its bounds. Share of voice is always share of a specific prompt set against a specific competitor set over a specific window. Without those three bounds stated, the figure cannot be interpreted and should not be compared with anyone else's.
Why might a business rank well in an AI visibility tool but not appear in "best agency" lists?
Because the tool and the lists measure different things. A tracking tool counts mentions inside questions somebody configured. A list reflects what an independent publisher chose to include. Scoring well on the first and being absent from the second is common, and it usually means the technical foundations are sound but the business has not yet earned external citation.
Which layer should a small business work on first?
Entity resolution. It is the cheapest layer to fix, almost entirely within your control, and every other measurement is unreliable until an engine can tell you apart from similarly named businesses.