How we measure AI visibility

Citation monitoring: which engines, and how they answer

We ask these engines through their developer APIs:

• ChatGPT, model gpt-5.4-mini: answers from what the model learned in training. Monitoring does not switch on a web-search tool for it.
• Perplexity, model sonar: answers with a live web search, which is built into this model.
• Claude, model claude-sonnet-4-6: answers from what the model learned in training. Monitoring does not switch on a web-search tool for it.
• Gemini, model gemini-3-flash-preview: answers from what the model learned in training. Monitoring does not switch on a web-search tool for it.

That difference matters. An answer from memory reflects the web as it was when the model was trained, so changes to your site show up slowly or not at all. A searching engine can pick up a change once it reads your page again. Monitoring tells no engine where the asker is, Perplexity included, so a local question needs the place in it.

What monitoring sends

• Your question, exactly as you wrote it, with no system prompt and no instructions around it.
• No location. A question like "best physiotherapist near me" reaches the engine without knowing where "me" is, so put the place in the question ("best physiotherapist in Leslieville").
• No account, cookies or history: each request starts fresh.

Three runs per monitoring check

The same question can get a different answer a minute later, so every check asks each engine three times at once. You count as named for that check when a strict majority of the answers that came back name you: 2 of 3, or 2 of 2 if one run failed. A tie counts as not named. If fewer than two runs come back, that engine records nothing for that check. An outage is left out of your numbers rather than counted as a miss.

Each query runs daily or weekly, as you choose. We keep the text of one answer from each check that agrees with the majority, so you can read what the engine actually said.

How "named" is decided

• It is text matching, not another AI model's opinion. We look for your brand name and the variants you give us as whole words, ignoring case, and for your domain anywhere in the answer.
• Only the answer text counts. Perplexity also returns a list of sources; a source link to your site that the answer itself never mentions does not count.
• Matching is literal. A misspelling of your name is missed, and a different business with the same name would be counted. Choosing distinctive variants is the best guard against both.

Citation rate: numerator and denominator

One check is one query, asked of one engine, on one day it was due. Citation rate is the number of checks that named you divided by the number of checks that produced a result, in a time window:

• the dashboard uses the last 30 days and compares it with the 30 days before;
• the daily email uses the last 24 hours and the weekly email the last 7 days, each against the period before;
• per-engine rates use the same formula, restricted to that engine.

A report flags a drop only when a statistical test says it is unlikely to be noise, and only over the queries present in both periods, so adding or removing queries does not register as a change in visibility.

The AI visibility check in an audit on a monthly plan

On a monthly plan (trial included), an audit asks up to 5 questions after the crawl. The one-off audit does not include this check. The questions are the ones you track or, if you track none, ones built from what your site names: its service and area ("Who is the best physiotherapist in Toronto?"), else its service alone ("Who is the best physiotherapist?"), else its area alone ("What are the best local businesses in Toronto?"), else your business by name ("What can you tell me about …?"). Each question goes to each engine in two modes, 3 times per mode:

• With live web search: ChatGPT, Perplexity and Claude. Each is given an approximate searcher location when we have one: your tracked location's city, region and country, or else the area your site names.
• From memory: ChatGPT, Claude and Gemini, with no search and no location. Perplexity always searches, so it has no from-memory mode.
• Gemini is asked from memory only: the software we use to reach it cannot switch on its web search yet.

For each question we record how many of the runs that came back named you, per engine and per mode (for example 2 of 3). The result is a dated baseline, not a rate over time, and it is compared only with a later check that asked the same way. The check runs at most once per domain every 7 days: an audit inside that window shows the last check with its own date. Each check also has a spending cap and a time limit, so a slow or failing engine can leave some runs without an answer.

The free scan is a sample of one

The free scan asks one engine (the first one configured, normally ChatGPT) one question, once, with no location. The question is built from your site's own structured data: the service and area ("Who is the best physiotherapist in Toronto?"), else the service alone ("Who is the best physiotherapist?"), else the area alone ("What are the best local businesses in Toronto?"), else your business by name ("What can you tell me about …?"). One answer shows what an engine can say; it is not a rate.

What these numbers cannot tell you

• API answers are not app answers. The ChatGPT, Claude and Gemini apps can search the web, know your location and remember past chats. Monitoring requests do none of that, and even the audit check, which searches from an approximate location, has no account or chat history, so a person using the app may see something different.
• Your questions are a sample. The rate covers the questions you track, not every question your customers ask.
• Engines change. A model update can move your rate overnight with no change on your side.
• A mention is not a recommendation. We label the sentence around a match as positive, negative or neutral with a simple word list; read the stored answer before drawing conclusions.

New to the subject? What answer engine optimization is. Plan limits are on the pricing page.