Decide what “visible” means before you count it
An AI answer can mention a company, recommend it for a particular need, link to one of its pages, or describe it accurately. Those are related signals, but they answer different questions. Record each outcome in its own column so a citation does not quietly become a recommendation metric.
Start with a decision you want the measurement to inform. A marketing lead may want to know whether the brand enters a shortlist for a service category. A content lead may want to see which factual questions produce an accurate answer. The measurement should reflect that job, rather than a single score with an unclear meaning.
- Mention: the brand name appears in the answer.
- Recommendation: the answer suggests the brand as an option for the stated need.
- Citation: the answer links to or attributes information to a source you can inspect.
- Accuracy: the answer describes the brand, its location, and its offer correctly.
Create a question set that represents real buying journeys
Write questions the way a buyer would ask them, not as a list of keywords. Include discovery questions (“What should I look for?”), comparisons (“How do these options differ?”), and selection questions (“Which provider serves this need in this location?”). Keep branded and unbranded questions separate. A brand name in the prompt changes the task and can make presence look stronger than it is.
Define the geography and language in each question set. A Brazilian Portuguese sample, a United States metro-area sample, and a Spanish sample for Mexico represent different markets. “Spanish-speaking Latin America” is not one uniform location: where the buyer is, what they call the service, and which providers operate there can change the answer.
- Choose one audience, need, and market for each question group.
- Write a balanced set of natural questions across discovery, comparison, and selection intent.
- Freeze the wording and label every question by topic, intent, language, and location.
Run a controlled sample and keep the evidence
Record the product or interface, date, locale, location setting, exact prompt, complete answer, and every cited URL. If a result came from an API, record the model and endpoint too. An API response is not a replica of a consumer interface: product settings, search retrieval, personalization, and model versions can differ. Label the collection method instead of blending unlike observations.
Repeat the same questions over time and, when practical, more than once in a measurement window. Save the raw answer and source list before calculating a rate. If a consumer interface has personalization or a location setting you cannot control, write that down as a limitation. The goal is a consistent sample, not a claim that every buyer sees the same answer.
Report separate rates with their denominators
For each engine and question group, report how many sampled answers mentioned the brand, how many recommended it, and how many cited an owned page. Show the count beside the percentage. “The brand appeared in 8 of 20 sampled answers” is easier to interpret than “40% visibility” on its own.
Also label citations by source type: your site, a directory, a review publisher, a public institution, or another source. A citation is evidence that a source appeared in that answer. It does not prove that the answer endorsed the brand, or that every user would receive the same answer. Recommendation and citation rates may move independently.
- Keep the prompt set and inclusion rules fixed between comparisons.
- Compare like with like: the same intent, language, location, product surface, and time window.
- Keep “unknown,” no-answer, and ambiguous brand matches visible in the report.
Turn an observation into a next step, not a ranking promise
Inspect the answer and its sources before editing a page. If the answer gets a service area wrong, correct the business facts on authoritative pages you control. If it gives no useful explanation of a product category, consider publishing a clear comparison or selection guide that answers the missing question and cites its evidence. Then repeat the same sample and note what changed.
A small prompt sample can reveal useful patterns, but it cannot establish universal visibility or prove that a page change caused a change in answers. AI products and retrieval systems change, and no measurement protocol guarantees a recommendation or position. Google’s guidance for AI features emphasizes the existing fundamentals of helpful, crawlable content; it does not offer a special markup that guarantees inclusion.
A practical first measurement
For a first pass, use one market, two or three buyer intents, and a fixed set of questions that your team can review manually. Capture answers and sources in a sheet, then mark mention, recommendation, citation, and factual accuracy independently. Repeat the sample on a later date before making a trend claim. Expand to another location or interface only when the first set produces a question your team needs to answer.