Measure what analytics can't see.

An assistant can recommend you and send no click, no referrer, nothing. Prompt-level tracking is the only way to know you were in the answer.

Book a callCheck your AI visibility

Prompts × engines

Illustrative, not client data

Prompt
“best CRM for a mid-size team”NamedAbsentMentionedNamed
“alternatives to [competitor]”MentionedMentionedAbsentCompetitor
“which telehealth provider to trust”AbsentCompetitorNamedAbsent
“best NDIS provider near me”CompetitorNamedCompetitorMentioned

Each cell is a repeated run, not a single check, answers vary between runs, so one screenshot proves nothing. The value is the trend, and the sources column behind each cell that names the pages the answer was built from.

What the reporting looks like

01 Why this exists

The channel is invisible by design

Your buyers’ questions down one axis, the engines across the other. Named, mentioned, or beaten, and behind every cell, the sources that decided it.

When a model names you in an answer, the buyer may act on it without ever visiting your site, and if they do visit, it can arrive as direct traffic with no attribution. Your analytics show nothing. Not a decline, not a channel: nothing.

Rank tracking doesn't cover it either. There is no position ten in an answer; there is being named or not being named, and it varies by how the question was asked, which engine answered it, and what that engine retrieved that day.

So the measurement has to be built the way the surface works: ask the questions your buyers ask, on every engine, repeatedly, and record what comes back.

02 Definition

What is AI Search Analytics?

AI Search Analytics
AI search analytics is the measurement of how often AI assistants name a brand in their answers. Because an assistant can recommend a brand without sending a click or a referrer, the visibility is invisible to conventional web analytics. It is measured instead by running a fixed set of buyer questions across ChatGPT, Google AI Overviews, Gemini, Perplexity and Claude on a schedule, and recording who is named, who is named instead, and which sources each engine cited to decide.

  • From sessions to mentions

    The unit of visibility is no longer a visit. It is whether your name appeared in a sentence someone read instead of visiting.

  • From rank to share of answer

    There is no position to track. There is how often you are named, against how often a competitor is, for the same question.

  • From one check to a schedule

    Answers vary between runs. A single screenshot is a sample, not a measurement, and treating it as evidence is how teams chase noise.

03 What’s included

Four things land. All of them are built, not filed.

  • 01

    A prompt set for your category

    The questions your buyers actually type, built from the source map rather than from keyword volume, including the comparison and alternatives phrasings where purchase decisions get made.

  • 02

    Coverage across the engines

    ChatGPT, Google AI Overviews and Gemini, Perplexity, and Claude, tracked separately. They retrieve from overlapping but different source sets, so a mention in one carries no guarantee in another.

  • 03

    Share of answer

    Not just whether you appear, but who else does, in what order, and how often, the competitive read that tells you whether you're gaining or the category is.

  • 04

    The sources behind each answer

    Which pages the engine cited to build its response. This is the actionable half: it converts a visibility problem into a specific list of pages to earn a place on.

04 How it runs

In order, and in the open.

  1. Step 01

    Build the prompt set

    We draft the questions, you cut the ones that aren't how your buyers talk. Getting this wrong makes every number afterwards describe someone else's market.

  2. Step 02

    Baseline

    A first full run across all engines, so later movement is measured against something real rather than against the month you started paying attention.

  3. Step 03

    Track and report

    Scheduled re-runs with monthly reporting: what changed, which competitor moved, and which sources are doing the work.

05 The difference

Web analytics vs AI search analytics

 Web analyticsAI search analytics
Unit measuredSessions and referrersMentions in answers
CoverageTraffic that arrivedAnswers whether or not anyone visited
Competitive readAbsentWho was named instead of you
AttributionDirect, unattributedThe source the engine cited

06 Who it’s for

Four markets, four different questions.

  • B2B software

    Buyers research through assistants long before a demo. The questions are comparative, alternatives, integrations, fit for team size, and the answer names two or three vendors.

  • DTC and ecommerce

    Category questions now start with an assistant. “Best X for Y” returns a shortlist, and being absent from it costs the consideration set, not just the click.

  • Healthcare

    Trust and provenance dominate. Assistants lean hard on sources they can attribute, which makes structured, citable content and clean entity signals disproportionately valuable.

  • NDIS and disability services

    Participants and coordinators ask assistants for local providers. The answers lean on directories, reviews and clearly structured service pages, which most providers have never marked up.

07 Questions

What buyers ask about this one.

Something not covered? Ask us directly and we’ll answer straight.

How do you track AI visibility?
By running a fixed set of your buyers' questions across ChatGPT, Google AI Overviews, Gemini, Perplexity and Claude on a schedule, and recording who is named, in what order, and which sources the engine cited. It is measurement by sampling, because there is no index to query and no rank to read.
Why can't we just use Google Analytics?
Because an assistant can recommend you and send nothing. No referrer, often no click at all, the buyer acts on the recommendation directly, or arrives later as direct traffic. Analytics shows neither a channel nor a decline, which is why the visibility has to be measured at the source.
How often should prompts be re-run?
Monthly for reporting, more often when something is in flight. The reason is variance: answers differ between runs, so any single check is a sample. Trends across repeated runs are the signal; one screenshot is not.
Can we see the prompts you use?
Yes, and you should. You cut the ones that aren't how your buyers talk. We report the prompts alongside the numbers so you can run them yourself and get roughly what we got. A measure you can't reproduce isn't a measure.
What is share of answer?
How often you are named for a set of questions, relative to how often each competitor is. It replaces rank as the headline number, because in an answer there is no position. Only who got mentioned and who didn't.
Isn't this just running the prompts ourselves?
You can, and for a handful of questions you probably should. It's the fastest way to see the problem. What doesn't survive the manual version is consistency: the same phrasing, on every engine, on a schedule, recorded so that a change three months out is measurable rather than remembered.
How much do answers vary between runs?
Enough that a single check proves very little. Models are non-deterministic and retrieval shifts, so we treat any one answer as a sample and report on trends across repeated runs. Anyone showing you a single screenshot as evidence of a ranking is showing you noise.
Can we buy this without the rest?
Yes. Some clients run measurement with us and the remediation in-house, and that's a reasonable split, the reporting names the pages to work on, and someone has to do the work either way.

Next step

Start with what your site already says.

The free checker fetches your site the way the assistants do and scores what comes back. It takes a minute, and it makes the first call a conversation about your findings rather than about our process.

Book a callRun the free check

08 The rest of the work