The short version
How to measure AI visibility in five steps
The method behind every number in mentionedin.ai, written so you can follow it by hand.
- 1Write a fixed set of prompts across the buying stages: questions about your brand, your pricing, comparisons with competitors, discovery questions where a buyer does not know you yet, and the few prompts you most want to win.
- 2Ask every prompt more than once on every engine, because AI answers vary from run to run. We ask each one 3 times per engine per test.
- 3For every answer, record whether the engine searched the web, which sources it cited and which model version answered.
- 4Count a mention only when the answer names you as the company you are, not when it repeats the question or means someone else with a similar name.
- 5Keep two control prompts you never work on, and freeze the prompt set, so a change in the numbers can be traced to your work and not to the engine.
What counts as a mention
Your name appearing in the answer is not enough on its own.
A run counts as a mention only when all four of these hold. It is the same rule for you and for every competitor, which is what makes share of voice on the same prompts comparable.
- The run completed without an error.
- The answer names your brand or one of its aliases.
- Echoes are stripped first. When the prompt contains your name, an answer that only repeats the question does not count.
- The entity is confirmed by your domain, a fact from your truth file or a judge step. An answer that uses your name for a different company is not a mention; it is shown to you as a mix up with someone else, because that is its own problem to fix.
Answers to branded and pricing prompts are also checked against your truth file. An answer that names you but contradicts one of your facts, such as a wrong price or a feature you do not have, still counts as a mention and is marked as not matching, so each prompt shows how many of its checked answers matched your facts. Sentiment is recorded as well.
A mention and a citation are two different numbers
Being named is one thing. Being the page the answer points to is another.
| Mention | Citation | |
|---|---|---|
| What it is | The answer names your brand | The answer lists a page as one of its sources. For Google AI Overviews this includes the top 10 organic results, not only the overview's own sources |
| Where it shows | Mention rate, consistency labels, share of voice | Cited sources, ranked by domain, yours and third party, per engine |
| What it tells you | Whether buyers hear your name | Which pages the engines trust in your category |
Every answer is stored with the URLs it cited. From those, mentionedin.ai builds source gaps: third party pages the engines cite that name your competitors but not you, ranked by how often they are cited, with the prompts and answers behind each one. Google AI Overviews is counted differently: a domain counts as cited there when it is anywhere on the result surface, meaning the overview's own sources plus the top 10 organic results, so read that engine's citations as result surface presence rather than overview sources alone.
How the four engines are asked
The mode in which a change to your site can actually change the answer.
ChatGPT, Claude and Gemini are asked through their official APIs with web search or grounding switched on, so an answer can draw on live pages and list them as sources. An API answer can differ from what a signed in user sees in the app, for example through personalisation and chat history, so treat the numbers as the engine's default answer to a fresh question. Google AI Overviews is read from a live Google results page, in the country set on your brand. A results page with no overview is a valid run, recorded as overview missing, not an error.
Every run stores the full answer, the cited URLs, whether the engine searched the web and the model version that answered. The four engines tracked are ChatGPT, Claude, Google AI Overviews and Gemini, all included in the plan. Perplexity and other engines are not tracked.
Engine runs, spot checks and consistency tests
Three units, used the same way in the product, the reports and the pricing.
| Unit | What it is | What it is used for |
|---|---|---|
| Engine run | One prompt asked once on one engine | The only unit the plan counts. There are no credits. |
| Spot check | Every prompt in the set, the 2 controls included, asked once per engine, weekly once you switch it on | Noticing change. Shown as one sample, never as a percentage or a label. |
| Consistency test | One prompt asked 3 times on each of the four engines, 12 engine runs in total | Every headline number: mention rate, labels, share of voice, cited sources |
Why 3 runs per prompt
One answer is a yes or a no. Several answers are a rate.
Ask the same question twice and an engine can name different brands. A single answer cannot tell you whether you are named most of the time or got lucky once, so mentionedin.ai never turns one answer into a score. Each engine gets 3 runs per prompt in every consistency test.
best client reporting tool for a small agency?
Run 1Named, 2nd
AgencyAnalytics, Your brand, Whatagraph, DashThis
Run 2Not named
AgencyAnalytics, Whatagraph, DashThis, Swydo
Run 3Named, 3rd
Whatagraph, AgencyAnalytics, Your brand
The number 3 comes from a calibration, not a guess. In a 10 run calibration test before the 7 October 2026 switch, 3 runs gave the same named or not named verdict in 225 of 240 prompt and engine cells; every test since asks each engine 3 times. That lets one plan test more prompts each month.
The trackers we read run each prompt daily and chart the drift over time; their pages do not say how many times a prompt is asked in one reading. We ask each engine 3 times per test and label how often you are named.
The six labels
The share of completed runs that named you, and what happens when runs fail.
| Label | Rule |
|---|---|
| Strong | Named in 3 of 3 runs (70% or more of completed runs) |
| Moderate | Named in 2 of 3 runs (40% to 69%) |
| Weak | Named in 1 of 3 runs (at least once, under 40%) |
| Invisible | Named in 0 of 3 runs, every run completed, and the engine did not answer purely from memory (no search signal at all counts as unknown, not as no search) |
| Inconclusive | Zero mentions, but some runs failed or the engine never searched, so zero is not a verdict |
| Error | Every run failed, so there is nothing to score |
The rule behind Invisible is strict on purpose. A prompt the engine answered purely from memory cannot be moved by your site in the short term, so calling it Invisible would send you after a fix that cannot work. Each result also shows its search rate, in how many of the answers the engine searched the web, when the engine reports it.
Spot checks: the weekly pulse
A cheap way to notice change, never a way to measure it.
Switch on a weekly spot check and every prompt in the set, the 2 controls included, is asked once on every engine each week. It is shown as "spot check, 1 sample" and is kept apart from the headline numbers. Spot checks do not use the consistency engine runs in your plan.
When a prompt that last tested Strong or Moderate is missed in two spot checks in a row on an engine, a 3 run consistency test on that engine is queued to confirm it, at most twice per prompt per month. The confirmation test decides whether the drop is real; the spot check alone never does. Confirmation runs count toward your monthly consistency engine runs, and the result shows in your list of checks.
Keeping the trend honest
Control prompts, frozen prompt sets and the model version on every answer.
- Two control prompts per brand are never targeted by a fix, are left out of every headline number and do not count toward your tracked prompts. They show whether the engine itself is drifting.
- Prompt sets are versioned and frozen. A frozen set never changes; an edit makes a new draft version, and the same wording keeps its history across versions. Checks always run the latest frozen set.
- Every answer stores the model version that answered, so a model update can be told apart from your own work.
- Score history plots one point per consistency test and engine, never a spot check.
What a month of measurement includes
The plan is counted in engine runs, so you can see exactly how many prompts get tested.
Founder plan
$99a month
or $79 a month billed yearly ($948 a year). Prices exclude VAT.
Start free trial14 days free. A card is required at checkout; nothing is charged before the trial ends. Cancel anytime.
- 1 brand with its competitors
- 15 tracked prompts plus 2 control prompts
- 4 engines: ChatGPT, Claude, Google AI Overviews and Gemini
- A weekly spot check of every prompt on every engine, switched on in one click
- 204 consistency engine runs a month: enough to test all 17 once on every engine
- Page grader and First 50 Words rewrites for your own pages
- Playbook of pages to create, pages to fix and source gaps
- Intervention log and a shareable web report
Consistency pack
$19per pack, only when you need it
$19 for 99 more consistency engine runs, about 8 more prompts tested 3 times on every engine. Buy one for a month you want to test more, such as right after a batch of fixes. Packs are the only add on.
One prompt, fully tested
The plan's 204 runs test all 17 prompts, tracked plus controls, once a month this way. A pack adds about 8 more.
The Founder plan covers 1 brand with 15 tracked prompts plus 2 control prompts, and 204 consistency engine runs a month: at 12 engine runs per fully tested prompt, enough to test all 17 once on every engine. A consistency pack adds 99 engine runs for $19, about 8 more prompts, for the month it is bought. A new account starts with a 14 day trial: a card is required at checkout, nothing is charged before the trial ends, and there is one trial per account.
The same method on our own brands
We publish what the method says about us.
We track four of our own brands publicly as case studies, mentionedin.ai included, and publish their current numbers on the case study pages; one of the four has not been measured yet. The numbers published so far were measured with the older method, before the switch to 3 runs per prompt on 7 October 2026, so do not read them as 3 run results. The word proven stays out of our copy until a case clears the success criteria written down for it before its first test (so far written for one of the four).
Questions people ask
How do you measure AI visibility?
Ask a fixed set of prompts on each engine several times, count the answers that name you as the right company, and record the sources each answer cites. mentionedin.ai asks ChatGPT, Claude, Google AI Overviews and Gemini each prompt 3 times per test and labels the result Strong, Moderate, Weak or Invisible.
Why does mentionedin.ai run each prompt 3 times?
AI answers vary from run to run, so one answer is a yes or no, not a measurement. In our calibration, 3 runs gave the same named or not named verdict as a test with more than three times the runs in 225 of 240 prompt and engine cells, so every engine gets 3 runs per consistency test.
What is the difference between a mention and a citation in AI answers?
A mention is the answer naming your brand. A citation is the answer listing a page as a source. mentionedin.ai reports both: mention rate and labels for mentions, and a ranked list of cited domains plus source gaps for citations.
Does mentionedin.ai count a mention when my name is in the question?
No. Echoes of the prompt are stripped before a mention is counted, and the answer must describe your company, not another one with a similar name. A name used for a different company is shown as a mix up instead.
What is a spot check in mentionedin.ai?
A single run of every prompt in the set, the 2 controls included, on every engine, weekly once you switch it on. It is shown as one sample and never as a score. Two misses in a row on a prompt that tested Strong or Moderate queue a 3 run consistency test on that engine.
Does mentionedin.ai check the real answer or an API response?
ChatGPT, Claude and Gemini are asked through their official APIs with web search switched on, and Google AI Overviews is read from a live Google results page. An API answer can differ from what a signed in user sees in the app, for example through personalisation and chat history, so treat the numbers as the engine's default answer to a fresh question. Every answer records whether the engine searched and which model version answered.
Does mentionedin.ai track Perplexity?
No. It tracks ChatGPT, Claude, Google AI Overviews and Gemini, all four included in the plan.