Measuring AI visibility: mention rate, rank, sentiment, citations, and their limits

Start with the unit: one answer to one question

Every metric here is built from the same unit: one answer, from one platform, to one question, at one point in time. A question is a prompt a customer might type, such as “best physiotherapy clinic in Abu Dhabi”. Measuring means asking a fixed set of questions on each platform, on a schedule, and recording what each answer says.

Because answers vary from one ask to the next, a metric is only as meaningful as the number of answers behind it. Keep that in mind for everything below.

Mention rate

What it is. The share of answers that name your business. For example, if 40 answers were collected and you were named in 10 of them, your mention rate is 25%.

Why it matters. It is the closest thing AEO has to “are we in the conversation at all?”, and it is easy to compare across platforms, languages, locations and competitors.

Its limits. It depends entirely on which questions you ask. A set of questions that includes your own name will produce a high rate that means little; a set of very broad questions may produce a low rate that no business could beat. Choose questions customers really ask, keep the set stable, and compare the rate over time rather than against an abstract target. Matching matters too: a mention written with a different spelling, an abbreviation or a possessive (“Joe’s” for “Joe”), has to be recognised as you, and a different business with a similar name must not be.

Rank

What it is. Your position among the businesses named in an answer: first, second, third. It exists only for answers that name you, and is usually reported as an average.

Why it matters. Answers often list options in order, and a reader may only consider the first few.

Its limits. An average rank says nothing about how often you appear: a business named once, in first place, has a better average than one named in every answer in second place. Always read rank together with mention rate. Rank also means less in answers written as prose, where the order of names may not signal any preference.

Sentiment

What it is. Whether an answer describes your business positively, neutrally or negatively, and how it frames you: premium or affordable, reliable or risky, suited to whom.

Why it matters. Being named with a warning (“some customers mention long waiting times”) is different from being named as the obvious choice. Framing also shows which qualities the platforms associate with you, which tells you what to reinforce or correct.

Its limits. Sentiment is itself a judgement, usually made by a language model reading the answer, so it carries some error, especially for mild or mixed wording. Treat it as a direction across many answers rather than a verdict on any single one, and read the answers behind a surprising number.

Citations

What it is. The sources an answer links to or lists. Citation metrics include how often your own domain is cited, and which other domains are cited for your questions, for you and for your competitors.

Why it matters. Citations show where answers come from, and that is something you can act on. A domain that is cited for your competitors but not for you is a concrete gap: a directory to join, a publication to approach, a page to write.

Its limits. Not every answer shows sources, and the platforms display them differently, so citation counts are not directly comparable between platforms. A cited source is not necessarily the only thing the answer relied on. And being cited is not the same as being recommended: your page can be cited for a fact while a competitor is named as the choice.

Sampling: how many answers are enough?

Every figure above is an estimate from a sample of answers, and with few answers, chance moves the number a lot. With 10 answers, one extra mention moves your rate by 10 points; with 100 answers, by 1 point. Before reacting to a change, ask how many answers the number is based on, and whether it has held for more than one run.

• Keep the question set stable, or compare only the questions that stayed the same.
• Look at trends across several runs rather than at single runs.
• Break results down by platform, language and location only when each slice still has enough answers behind it.
• Be suspicious of large swings after small changes, and measure again before drawing conclusions.

Location: how the platform was told where the customer is

For local businesses, the same question can get different answers in different cities. Measuring “in Denver” means telling the platform that the customer is in Denver, and platforms accept this in different ways: some take a structured location with the request, some only see a place name written into the question, and some infer an approximate location from the connection.

These methods are not equivalent. A place name in the question tells the platform what the customer is asking about, but not necessarily where they are, so the answer can differ from what a person physically in that city would see. When you compare locations, check how each platform received the location, and compare like with like.

Other limits worth stating

• Personalisation. Real users may be logged in, with conversation history and saved memory. Measured answers usually come from a clean session, so they represent a typical answer, not any individual customer’s.
• Platform changes. Platforms update their models and search systems without notice. A shift in your numbers can reflect a platform change rather than anything you did; check whether competitors moved at the same time.
• Answers are not outcomes. Being named does not mean being chosen. Where you can, connect mentions and citations to what you measure further down the line, such as referral visits from AI platforms, enquiries and bookings.
• Blended numbers hide differences. A single figure across languages, locations and platforms is convenient, but it can hide a gap in exactly the market you care about. Keep the breakdowns.

How Pharos helps

Pharos measures all four metrics on ChatGPT, Gemini, Perplexity, Claude and Google AI Mode. Each scheduled run asks every tracked question on all five platforms; your plan sets how many runs happen each week, and measured answers per month are questions × 5 platforms × runs per week × 4.3. Every answer is kept in full so you can read what sits behind a number, spelling variants of your name are matched with guards against false matches, and per-location results come with a note on how each platform received the location.