How AI visibility is measured

What does an AI visibility audit measure?

A Referido AI visibility audit measures how often an AI assistant names a business when it is asked the buying questions that business's real customers ask. The unit of measurement is the answer: 81 answers are collected per audit, and the report counts how many of them name the business, how many name its competitors, and in what order.

The measurement is taken from the outside, the way a customer would take it. Nothing is installed on the business website, no accounts are requested and no internal data is used. Everything in a Referido report is information any person could obtain by asking the same platforms the same questions.

The audit does not predict sales and does not promise rankings. It describes a state on a date, with the questions and the platforms listed in the report itself.

Which questions are asked, and who writes them?

Referido writes the questions, not the business being audited, and none of them mention the business by name. Each audit uses 14 questions built from the services, the service area and the customer type of the business, and five of them are marked as primary because they carry the highest purchase intent.

Leaving the name out is the condition that makes the measurement valid. A question that asks "how good is Clinic X?" guarantees an answer about Clinic X, which measures nothing. Audit questions look like "what are the best implant clinics in this area?", which is what someone types when they do not yet know who to call.

The 14 questions follow a fixed mix:

About 60%

search for the service in the area, in the words a customer would use ("best X in this town").

About 30%

carry high intent: price, a comparison between options, or urgency.

About 10%

are natural questions with personal context, the kind someone gives an assistant before asking for a recommendation.

Which AI platforms are queried?

Five platforms are queried: ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews. The first four receive all 14 questions and answer conversationally, with web search enabled. Google AI Overviews is queried on the five primary questions only, because it is not an assistant you write to but the AI summary Google places above its results.

Each platform is queried with a model of the class a person on a free plan receives, not with the most expensive model available. The measurement has to resemble what the business customer sees, and the vast majority of people ask the free version.

Questions are asked without a signed-in session and without history. An assistant that has already talked to someone tailors its answers to that person, and that personalization is neither measurable from the outside nor comparable between two businesses.

What counts as the business appearing in an answer?

An appearance means the business name, or one of its declared alternate names, is written literally in the text of the answer. The comparison is an exact text match, not an interpretation: if the name is not written, there was no appearance, even when the answer describes a similar business.

This rule exists because Referido uses a language model to read the answers and pull out competitors, position and tone, and that model does not get a vote on the one thing the report headline claims. When it was allowed to weigh in, it counted answers naming the parent brand as appearances of the business, and the headline number moved between two readings of the same material. Its opinion is now stored separately and kept out of the count.

Everything else does come from the reading model: which other businesses the answer recommends, the order they appear in, and the verbatim sentence that serves as evidence. Every claim in the report ships with that quote so it can be checked against the original answer.

How is it decided which businesses count as competitors?

Competitors are businesses in the same category and area that the answer recommends alongside the audited business or in its place. They are taken in the order the answer names them, up to eight per answer, then summed across all 81 answers to build the report scoreboard.

AI platforms, directories and aggregators are excluded. ChatGPT, Google, Yelp, TripAdvisor and the like are cited in many answers and compete with no clinic and no contractor, so counting them would inflate the scoreboard with names that cannot take a customer away.

When a business operates under a parent brand, the parent brand is listed separately and labelled as such. It does not compete with its own affiliate, but it does absorb mentions the business is not getting, and hiding it would erase the number that explains the rest of the report.

What is checked on the website itself?

Five things are checked on the website, each deterministic and each producing a quotable piece of evidence. These are not design opinions: they verify whether the information about the business is available and legible to the programs that feed the AI platforms.

Structured data

whether the site declares what kind of business it is using schema.org, or carries no machine-readable block at all.

Answer-shaped content

whether there is an FAQ section or question-shaped headings a model can quote directly.

English version

only when the business serves international customers, where the absence makes it invisible to anyone searching in English.

Google profile

whether the site links a Google Business or Maps profile, the primary source behind AI summaries on local questions.

AI crawler access

whether robots.txt blocks the crawlers of the AI platforms, which is the most expensive and most silent of the five failures.

Why are the five primary questions asked twice?

The five primary questions are asked twice because an assistant does not give the same answer to the same question every time. Models generate text with a random component and they search the web in the moment, so two consecutive queries can return different lists of businesses with nothing having changed in the world.

The practical consequence is that a single query is not a measurement. Asking once and appearing does not prove a business always appears, and asking once and not appearing does not prove the opposite. Asking twice does not solve the problem, but it does avoid the crudest error, which is reporting a dice roll as a fact.

Referido is currently measuring how large that instability is across its own archive of audits, and will publish the figure with its method and its margin of error once the study is finished. It is not published earlier because the business names compared between two answers are extracted by a model, and the share of the change that is real rotation has not yet been separated from the share that is the model reading the same text differently. A number without that separation would look firmer and be less true.

What does this audit not measure?

The audit does not measure how many people ask an AI those questions, nor how many of them go on to buy. It measures whether the business appears when asked, which is a precondition for the rest and is not the same as the rest. The full limits are these:

It is a single day

results correspond to the date printed on the report. Platforms change models and indexes without notice.

It does not measure volume

no platform publishes how often it is asked anything. Anyone claiming to have that number does not have it.

It does not measure sales

appearing in an answer is not a visit and not a customer. Traffic arriving from assistants can be measured, but that is a separate measurement.

It is one location

local answers depend on where the person asking is. The audit fixes one location, states it in the report, and does not cover the rest.

It is the free tier

a customer on a paid plan may get a different answer to the same question, with more searches and more sources.

It is not causal

the audit shows the current state and the gaps. It does not prove that closing a given gap will produce a given appearance.

How long does a result stay useful?

A report describes the day it was run and is best treated as expired after three months. The platforms change their models several times a year, and a model change can move the list of recommended businesses without any website having changed.

That is why every Referido report carries the date, the questions and the platforms on its face. A visibility number without those three things cannot be compared to anything, not even to the same business three months earlier.

How can this method be verified?

Anyone can repeat the measurement: the questions used in an audit are listed in the report, and the platforms are public and reachable. Asking the same questions and comparing is the most direct way to audit Referido, and the report is built so that this can be done.

Every claim in the report ships with the verbatim sentence from the answer that supports it. When a figure comes from an outside source, the source and its date are named. When a figure is our own, the sample and the method are published, which is the same rule that governs the Referido blog.

If repeating a query produces a different result, some difference is expected: that is the point of the section on why the primary questions are asked twice. What should not differ is the criterion, and the criterion is written out here in full.