Skip to content
Built for AI agentsConnect yours

Methodology

How to measure AI visibility, step by step, so you know what got you recommended

Getting recommended by AI is work, and work needs a fair before and after. Here is how a run works, step by step, and every rule that constrains what we are able to report, so a lucky week never passes for a change that worked, including several rules that would make the product look better if we dropped them.

Start 7-day free trialHow a run works

Card required, nothing charged for 7 days

A run, end to end

How a run works: from prompt to answer in four steps

You choose what to ask. We ask every engine, keep every answer in full, show who was recommended and what the engine read, and keep asking, so every change you make has a before and an after.

  1. Step 1

    Pick the prompts your buyers type

    Whole prompts, not keywords. We suggest a set from your category and competitors, sized to your plan, and you edit it. Then keep the wording fixed.

  2. Step 2

    We ask every engine, on a schedule

    Every prompt goes to every engine on a fixed schedule, from your market, so a change in the answer is a change in the world. A failed probe is retried, then left out of the denominator.

  3. Step 3

    We find who was recommended instead

    Which brands were named and in what order, which pages the engine read, and the full text, with your competitors tracked beside you and the sites that name them and not you flagged.

  4. Step 4

    You act, and the re-checks show the result

    Each engine gets its own rate, trend and sample size, so you know which engine and which sources to work on. The scheduled runs keep going, so the trend shows whether what you changed worked.

Example dataThe app’s screens for each step, shown with an example brand.

Prompt tracking keeps the set on its schedule, AI citation tracking keeps the pages each answer cited, and the rules below govern every step, from the minimum sample to what a failed probe does to your numbers.

The rules

Six rules for measuring AI search visibility

These are enforced in the product, not aspirations written on a marketing page, and each one is visible in it.

  1. Rule 1

    Counts always carry their denominator

    Every rate is shown with the answers behind it, because a percentage with no sample size is a claim you cannot check, and at low volume usually a wrong one. Below five answers we show the fraction only: one run is a draw, not a rate.

  2. Rule 2

    Engines are never averaged together

    Each assistant retrieves differently, and a blended score hides the one you are losing. Each engine gets its own rate, trend and sample size; where the Overview totals across engines, it pools the counts and says so.

  3. Rule 3

    Mentions and citations are different things

    Being named in the answer and being cited as a source are separate states, and a brand can have one without the other. We report them apart because the fixes are unrelated.

  4. Rule 4

    Our failures are excluded, not counted against you

    A probe that fails leaves the denominator rather than counting as an answer that did not name you, so our outage never reads as your brand disappearing. A Google query with no AI Overview is left out the same way.

  5. Rule 5

    The same prompt, every time

    Wording changes the answer, so the prompt is held constant. A reworded prompt is a new prompt with its own series, because a trend across two different prompts is not a trend.

  6. Rule 6

    Answers are stored verbatim

    Every response is kept in full, with its citations and how it was collected, so any number can be traced to the sentence behind it. A metric you cannot audit is one you should not trust.

Example dataThe screens each rule governs, shown with an example brand.

On a phone

The same numbers on a phone, sample sizes included

The small screen keeps what makes a number checkable: every rate still carries the answers behind it, and every stored answer opens in full, with the brands marked and the sources listed.

The same sample sizes.
Each engine keeps its denominator, and a fraction stands in for the percentage wherever there are too few answers for one.
The same stored answers.
The text the engine wrote, with the cited, not named note where it applies.

Example dataThe Overview and an answer, on a phone.

Alerts

When a change in AI visibility is real, and worth an email

Answers vary between runs, so a tool that emails you about every movement is reporting noise. We alert only when both weeks carry enough answers to tell a change from ordinary variance, and when the change is big enough to act on. The same test is how you should judge whether something you did worked.

Both weeks need a sample.
This week and the week before must each carry enough answers, ten by default. A drop from thirty answers to three is a gap in collection, not a change in visibility.
Your threshold or the noise floor, whichever is higher.
You set how big a move counts, fifteen points to start. The noise floor is what two samples that size could show by chance, at 95% confidence, and on a small plan it can be the higher.
One finding needs no sample.
Cited, not named is one answer that read your page and named someone else. It is true or it is not, so it is the one alert with no minimum.

Prompt selection

Which prompts to measure

A measurement is only as good as what it measures, so prompt selection is a methodological choice, not a setup step. Track the wrong prompts and every figure downstream describes a conversation nobody is having.

Before they know your name.
We read your site, name your category the way a buyer would, and propose the prompts a buyer types before they know your name: someone asking for you by name has already found you.
Search volume, not prompt volume.
Ask, and we look up each suggestion's monthly Google searches, live from Google Ads data, cached for a week, in your workspace's market. A proxy: it assumes the person who types a question into Google is much the same person who asks an assistant.
No figure rather than an invented one.
Nobody can measure how often a prompt is sent: the assistants do not publish it and no panel sees it at scale, so a prompt count presented as measured is a model. If the lookup fails you get suggestions without figures. A missing figure is recoverable; an invented one is not.

What we do not measure

Stated plainly, because the gaps in a measurement matter as much as its coverage.

We cover six engines today. Microsoft Copilot, Grok, Meta AI and DeepSeek are not covered yet, so we report no number for them rather than an estimate.

We also do not estimate how many people saw a given answer. Nobody outside the assistant vendors knows that, and a confident number would be invented. See the engines we do track and how a run works.

Measured, on every plan

Not measured

  • Engines we do not cover yet. Microsoft Copilot, Grok, Meta AI and DeepSeek.
  • How many people saw an answer. Nobody outside the assistant vendors knows.
  • How often a prompt is sent. The assistants do not publish it, so we show search volume and call it a proxy.

Questions about how to measure AI visibility

How each engine is collected is set out on platforms. Anything else, ask us before you sign up rather than after.

Start 7-day free trial

Card required, nothing charged for 7 days

Why does RankHound refuse to show a single AI visibility score?
Because a blended score would be the most saleable and least honest thing on the screen. The engines disagree with each other, often sharply, and averaging that disagreement produces a number that describes no real surface. The one total we do show, the Overview's headline, pools the counts instead: the answers that named you over every answer collected, both numbers printed, the words all engines pooled beside it, and the six engines on their own cards directly below.
What is the minimum sample before you show a percentage?
Five answers. Below that we show the fraction, such as 1 of 3, rather than 33 percent. The rounding is not the problem; the implied precision is.
How do you decide a brand was mentioned?
We match the brand's name, its other spellings and its domain against the answer text, by fixed rules rather than a model's judgement, then record where it was first named and its place among the brands named. If your brand is named after an ordinary word, it is matched only as written, capital letter included, so the word in an ordinary sentence never counts. The full answer is stored, so you can check any call we made.
Do answers vary if you run the same prompt twice?
Yes, substantially. That variance is the reason this product runs on a schedule rather than on demand. A single check of a generated answer is close to meaningless, and any tool reporting one as a status is reporting a coin flip.
How do I tell whether a change got me recommended more often?
Keep the prompts fixed and compare the weeks before and after the change, on the engines and prompts it was aimed at. The rules on this page decide what that comparison can say: each side needs enough answers, and the move has to clear the noise floor that two samples that size could show by chance. On a small prompt set that often means several weeks of too early to say before the re-checks can show whether it worked, and we would rather print that than call a lucky week a win.
How long before RankHound shows anything useful?
The first answers land within minutes of your first run, and you will see who is recommended in your place the same day. A trend takes about two weeks, because answers vary between runs and it takes that long for a direction to separate from noise. The same applies to anything you change: on a small prompt set, expect the re-checks to say it is too early to tell for a few weeks before they can show whether it worked.
How many prompts should I track?
Enough to cover the ways your category gets asked about, and few enough that you keep them stable. Plans run from 2 prompts on Solo to 15 on Scale, and the suggestions are sized to whichever you choose. Consistency matters more than breadth: a prompt you rewrite each month produces no trend at all.
Do I need to add my competitors?
Yes, at least one. Onboarding asks for it, because share of voice needs someone to compare against. Being named in two answers out of twenty means one thing if nobody else is named either, and something very different if one competitor is named in eighteen.
What happens when a run fails?
It is retried, and if it still fails it is excluded from your denominator rather than recorded as an answer that did not name you. You are emailed if a whole run fails or if most of one does, because a silent failure is indistinguishable from a quiet week.

Related reading

7-day trial

See whether your next change works

The trial runs on the same rules described above. Card required, nothing charged for 7 days.

Start 7-day free trialSee pricing

Card required, nothing charged for 7 days.

On every plan

  • Every answer stored in full, with its citations.
  • Every rate shown with the number of checks behind it.
  • Cited-but-not-named reported as its own state.