Method · Updated November 20, 2026

How RankRush measures AI visibility

Your score is one fraction: the answers that name you, divided by the answers we could score. This page shows what counts, what is left out, how each engine is asked and how we tell whether a fix worked.

See the formula How we compare tools

The formula

Answers that name you, over answers we could score

The same fraction runs for every engine. The dashboard works it out when you open it, from the answers stored for your keywords.

Per engine

Visibility = answers that cite or mention you ÷ answers we could score

Each keyword you track is asked on every engine you have. The score reads the newest answer for each keyword on each engine, in one country at a time.

Overall

Each engine counts by how many answers it scored

An engine with 30 scored answers counts for more than one with 10. An engine with no scored answers adds nothing, so it never drags the overall score toward zero.

A worked example

On ChatGPT, 18 of 30 scored answers name you: 60%. On Gemini, 6 of 20 do: 30%. The overall score is (18 + 6) ÷ (30 + 20) = 48%. Two more Gemini checks that failed are not in the 20, and they are not counted as misses.

What is left out

An answer we could not score is unknown, not a miss

Two things leave an answer out of both sides of the fraction. Neither is counted for you or against you.

A check that failed

An engine timed out or returned an error. There is no answer to read, so there is no verdict. Failed checks are asked again on the next scan, and the dashboard tells you how many are still missing.

A search where Google showed no AI Overview

Google does not show an AI Overview for every search. When none appears, there is no answer to name you in, so that check drops out instead of counting as a loss.

Why this differs from counting every response

A plain mention rate divides by every response sent, including the ones that never came back. That is simple, and it is how some tools publish their score. We divide by the answers that exist, so a score moves when what AI says moves, not when a vendor has a bad hour. A keyword or an engine with nothing scored shows as no data, never as 0%.

The engines

How each answer is fetched

Every question is asked from your country and in your language. Each keyword belongs to one country, and a website can track up to three.

EngineHow the answer is fetchedIncluded
Google AI Overviews The live Google results page for your search, in your country and language. When Google shows no AI Overview for that search, there is no answer to score and the check is left out. Every plan
ChatGPT The answer chatgpt.com shows for the question, with the sources it cites. If that route fails, we ask OpenAI with web search on instead. Every plan
Gemini Google Gemini through its official API, with Google Search grounding switched on. Every plan
Perplexity Perplexity through its official API, which searches the web for every answer. Add-on, €45/mo per website
Claude Anthropic Claude through its official API, with the web search tool on. Add-on, €45/mo per website

Google AI Overviews and ChatGPT are read as users see them. Gemini, Perplexity and Claude are asked through their official APIs with web search on. Scans run on the schedule you set, from once to seven times a week, with automatic scans switched on or off for each country, and whenever you press Analyse.

What counts as named

A citation or a mention, decided by code

Either one makes an answer visible. A link to you is a citation. Your name in the text, with no link, is a mention.

Citation

The answer lists a page on your domain among its sources.

Mention

The answer's text names your brand or your domain without linking to you. Names match as whole words, so a short brand is not found inside a longer word. A brand with several words counts when most of them appear.

No model decides whether you were named

Code checks your brand name, your domain and the links the answer cites. A small model reads each answer once to list the brands it names, which is how RankRush shows who AI recommends instead of you. Code then looks in that list for your link or your name. The model is never asked which brand is yours. Where you appear in the answer is recorded, but it does not change the score.

The split

One score, cut four ways

A blended number hides what to fix. The same fraction is also worked out for each kind of search, so you can see where you win and where you lose.

Category

Broad searches where a buyer asks for options, such as "best running shoes under 100".

Question

Searches that want information, such as "how do I clean suede".

Comparison

Searches that weigh options against each other, such as one brand against another.

Branded

Searches that already contain your own name. They show recall, not growth.

A keyword we cannot classify sits in a fifth group, so the four groups and that one always add back up to the whole score. A group with nothing scored shows as no data, not as 0%.

Did the fix work

Won, lost ground, no clear change, or too early

A fix is judged by what AI answered before and after, against keywords you did not touch. These are measurements, not proof that the change caused the move.

A change you made for a keyword

RankRush counts how many engines cite you for that keyword in each of the next four scans and compares the average with the average of the last three scans before the change. It then subtracts how much your untouched keywords moved over the same scans.

  • Won. The keyword gained at least three quarters of an engine more than the untouched keywords, by clearly more than their own ups and downs.
  • Lost ground. The same, in the other direction.
  • No clear change. Anything between, or fewer than three untouched keywords to compare with.
  • Too early. Fewer than four scans have run since the change. The row shows the count, such as scan 2 of 4.

Only a result that is large and consistent across the scans is shown in colour.

A page change from an Actions card

Once the change is live, RankRush compares the answers after it with the latest answers before it, keyword by keyword and engine by engine, using only pairs measured on both sides.

  • Success. The page is cited on more keyword and engine pairs than before, or the card's questions held the level they had.
  • Try again. The page is cited on fewer pairs, or the questions lost ground.
  • Pending. The change is not live yet, or no scan has run since it went live.

An answer we could not measure is left out of both sides, never read as a loss.

FAQ

How the score works: common questions

Short answers, taken from the sections above.

How is the AI visibility score calculated?

For each engine, visibility is the number of answers that cite or mention your brand divided by the number of answers we could score. The overall score weights each engine by how many answers it scored, so an engine with more scored answers counts for more. The score reads the newest answer for each keyword on each engine, one country at a time.

Why are some answers left out of the score?

An answer is left out when there is nothing to score: a check that failed, or a search where Google showed no AI Overview. We leave it out instead of counting it as a miss, so a failed check never looks like a drop in visibility. An engine with no scored answers adds nothing to the overall score rather than pulling it toward zero.

Which AI engines does RankRush ask, and how?

Google AI Overviews from the live Google results page, ChatGPT as chatgpt.com shows its answer, and Gemini with Google Search grounding on every plan. Perplexity and Claude are paid add-ons at €45 a month per website. Every question is asked from your country and in your language.

What is the difference between a citation and a mention?

A citation is an answer that links to a page on your domain among its sources. A mention is an answer whose text names your brand or your domain without linking to you. Both count as visible. Where you appear in the answer is recorded but does not change the score.

Does an AI model decide whether my brand counts as named?

No. Code decides, from your brand name, your domain and the links the answer cites. A small model reads each answer once to list the brands it names, which is how RankRush shows who AI recommends instead of you, and code checks that list for your link or your name. The model is never asked which brand is yours.

How do you know a fix worked?

For a keyword you changed something for, RankRush compares the next four scans with the scans before the change, against keywords you did not touch over the same scans. The result reads Won, Lost ground or No clear change, and Too early until four scans have run. For a page change from an Actions card it reads Success or Try again.

How often is the score updated?

Every time your keywords are asked again: on the schedule you set, from once to seven times a week (once a week by default on a paid plan, switched on or off for each country), and whenever you press Analyse. Rechecking a single keyword is free.

See your own AI visibility score

RankRush asks ChatGPT, Gemini and Google AI Overviews your questions in your market, shows who they name instead of you and turns every gap into a fix. 3-day free trial on every plan, from €59 a month.

Start free trial See pricing