How we measure AI visibility

A Professional report and ongoing monitoring both put your buyers’ questions to five AI assistants: ChatGPT, Gemini, Claude, Perplexity and Google AI. The free report is deliberately smaller: it uses Gemini only, across six questions, so it can stay genuinely free. It is a sample, and we label it as one rather than presenting it as full coverage.

We do not ask an assistant about you. We ask it the questions your customers ask, in your category and your area, phrased the way a buyer phrases them, then read who gets named. Asking “what do you think of X” measures whether a model can describe a company it was just told about. Asking “who is the best X near me” measures whether it recommends you unprompted, which is the thing that actually costs you customers.

The question set is fixed per category so this week can be compared with last week.

You see every question used in your own report. Each finding is printed next to the question that produced it and the answer that came back, so you can audit the evidence yourself rather than take a score on trust. What is not published publicly is the method for generating new sets and the questions reserved for future monitoring runs. That protects the integrity of the measurement, because a set published in advance becomes a target to optimise against, and the number stops describing anything.

The categories are not a secret. Questions cover buying intent (“who is the best X near me”), comparison (“X or Y”), problem-led searches, trust and credential questions, and price or value questions, selected for what buyers in your category actually ask.

This is the part we care most about getting right. A report contains three different grades of statement and labels which is which.

1. Measured evidence. We asked a question, we received an answer, we recorded what it said and which sources it cited. This is a fact about what happened. Counts of who was named, which competitors appeared, and which sources were cited are all in this category, and they are reproducible.

2. The assistant’s own explanation. We can show a model its answer and ask why it chose what it chose. We print that reply word for word, and we tell you plainly that it is testimony, not proof. When an assistant explains an answer, it generates that account after the answer was produced. The explanation can be genuinely useful and often points at something real, but it is not a verifiable record of every internal process behind the original response, and we have no access to those processes. We print it as an account, never as measured evidence.

3. Our analysis. Our reading of the first two, and what we would do about it. Every conclusion points back at the evidence underneath it. Where we could not measure something, we mark it not measured instead of estimating.

When an assistant searches before answering, it exposes sources. We record them. A citation is solid evidence that a source was retrieved or attributed for that answer.

It is not proof of everything the model weighed internally. Some of what shapes an answer comes from training rather than retrieval, and none of that is visible to us or to anyone outside the model provider. So we report citations as what they are: the observable raw material an answer was built from, which is both genuinely useful and genuinely incomplete.

We can measure which businesses appear in the answers we sampled, record the sources exposed alongside them, compare you against named competitors over a fixed question set, track how all of that moves week to week, and recommend actions based on the patterns we measured.

We cannot guarantee that an assistant will recommend you. Nobody can. Model providers change their systems without notice, retrieval varies between identical questions, and no external party controls what a model outputs. What we influence is the material an answer is built from. That is a real lever and it is not a switch, and any product telling you otherwise is selling you something it cannot deliver.

The AI Trust Score is a GetRanked Ai measurement framework. It is derived from what we observed inside a defined scan sample, and nothing else.

It is not produced, issued, endorsed by or affiliated with OpenAI, Google, Anthropic, Perplexity or any other AI provider. It is not a general measure of your company’s reputation or quality. It is not a prediction, and a high score is not a guarantee of future recommendations. It describes a sample of answers at a point in time.

Four components, each counted from your own answers:

Knowledge. Whether the assistants appear to recognise your business and describe it accurately within the answers we measured. We can only judge what an answer said, not what a model internally knows.

Authority. Whether the sources cited or retrieved in those answers support your presence and your claims.

Evidence. Whether the measured answers contain concrete, attributable detail about you, rather than a passing mention.

Coverage. How broadly you appear across the tested question set and the assistants included in your plan.

We publish what each component means and what moves it. We do not publish the weights, the formula or the thresholds, for the same reason we do not publish the full question set: a score with a published formula is a score people write pages against rather than improve against.

Every report is a timestamped snapshot. It records what we observed on the day we observed it, and it stays true as a record of that even after the answers change.

AI outputs change between scans. Providers update models without notice, retrieval varies, and the web underneath the answers moves. So a repeat measurement is only meaningful if it is comparable: we re-run the same or equivalent questions under the same conditions wherever we can, and where something material changes, the engine, the model, the question set or the scan configuration, that change is recorded against the run rather than quietly absorbed into the trend.

Monitoring shows you movement. It does not create certainty. A score that rises over six weeks is evidence that something is working; it is not proof of cause, and we do not present it as such.

Coverage differs by plan, and comparing across plans is not like for like: the free report is six questions on one assistant, Professional is sixty questions across five assistants producing up to 300 AI answers, and Monitoring re-runs twenty across five.

If something in a report is wrong, we want to know, and we will correct it. Write to hello@getrankedai.com about any of the following:

An incorrect fact about your business. A competitor we identified who is not actually a competitor. A citation or source that looks wrong or does not support what it is attached to. A concern about the methodology on this page. Anything else in a report that does not look right.

Please include the report link or scan reference so we can look at the same evidence you are looking at.

Ask the same assistant the same question twice and you can get two different answers. That is a property of the systems, not a defect in the measurement, and it is why a single answer is never treated as a result. Findings come from patterns across a question set and across assistants, and a one-off mention is reported as a one-off mention.

A free report samples six questions on one assistant. A Professional report is sixty questions across five assistants, producing up to 300 AI answers. It is a ceiling, not a promise: a provider can fail, an answer can come back empty, and the real total for a given run can be lower. The larger the sample, the more weight a finding carries, and we never present the small one as though it were the large one.

It complements search engine optimisation rather than replacing it. SEO is concerned with where your pages rank in a list of results. This measures whether an assistant mentions, cites or recommends you when someone asks a question and gets a single answer instead of a list. Strong technical SEO, genuinely useful content and credible third-party mentions feed both, because they are what an assistant reads. The category is starting to be called generative engine optimisation (GEO) or answer engine optimisation (AEO).