How we track AI mentions: our methodology in full

This page is the public version of how we measure AI visibility, in full: what gets checked, on which platforms, how findings are reported, and the conclusions we refuse to draw. It exists because almost nobody in this industry publishes their method, and unpublished methods are how junk numbers survive. Two instruments are documented here: the one-time review (a dated snapshot, which is what the free report and the AI Visibility Roadmap are built on) and continuous tracking (the ongoing measurement behind Managed AI Visibility).

We measure AI visibility by asking AI assistants the questions customers actually ask, recording exactly what each one answers, and reporting whether a business gets named. A one-time review does this once, as a snapshot across a designed question set. Continuous tracking does it on an ongoing basis across the questions we track for a business. Neither produces a made-up composite score, and neither claims to prove causation.

The one-time review: a dated snapshot

The review’s method, as confirmed in our methodology document (version 1.1; the method was confirmed at 1.0, and 1.1 is a vocabulary-only revision):

Prompt set. A designed set of questions per review, phrased the way buyers ask, covering the business’s primary offer and market. Phrasings vary by industry. Why a designed sample and not one question is a real design decision: one question is an anecdote, a designed sample is a measurement.

Platforms queried. The free report covers ChatGPT, Perplexity, and Google AI Overviews. The roadmap covers the wider set of assistants we can measure, narrowed to the ones that matter for the industry.

What gets recorded. Verbatim answer text plus a screenshot, dated, with location context noted. A finding you can’t trace to a recorded answer doesn’t go in the report.

What gets reviewed. Where the business shows up and where it doesn’t, who AI names instead, how AI describes the business, and the sources the answers draw on. For local businesses, the signals behind those answers have published definitions.

Findings, no composite score. Each finding is reported with the specific supporting observation under it. We don’t produce a 0-to-100 score because an invented composite implies measurement precision the underlying checks don’t have. Findings with visible support are more honest and, in practice, more convincing.

Continuous tracking: what the monthly service measures

Continuous tracking applies the same recording discipline to an ongoing panel instead of a one-time snapshot.

What gets checked. The questions we track for a business: the questions buyers ask AI assistants, tracked so you can see whether the business gets mentioned in the answers. Core questions are the ones we already know buyers ask in the industry; custom questions are the ones the client picks with us.

Platforms covered. The assistants listed on the method page, narrowed per client to the ones their buyers actually use.

How often. Every tracked question is checked on a fixed schedule. Repeated checking is what turns single answers into trends: one answer is a sample, a month of them is a pattern you can act on.

What counts as a mention. The business named in the answer text itself. A citation of a page about the business without the name in the answer is recorded as a citation, not a mention; the difference between the two outcomes is part of why both get tracked. The derived numbers (mention rate, citation rate, share of voice) are defined in their own reference, which this method feeds.

Trends, not single results. AI answers vary between askings, so tracking reports movement across many samples over time, and single-answer changes are treated as noise until they repeat.

The boundary we keep on purpose

Measurement is explainable, and this page explains it. The reasoning behind our recommendations is not published: here’s exactly what we measure, and the prioritization logic is ours. We’d rather state that boundary openly than pretend the whole system is public.

What we refuse to conclude

These clauses appear in every report, and they bind everything we publish:

• Answers vary by model, location, phrasing, and week. A report is a dated snapshot, not a permanent state.
• Where we state that a signal influences AI answers, the support is either observed citation behavior in recorded answers or a named source. Correlation is labeled as correlation.
• A review does not measure rankings, traffic, or anything requiring analytics access. It does not predict outcomes or promise timelines.

The same discipline applies to tracking: we report whether you’re named, we don’t promise that any action produces any mention, and anyone in this market who does promise that is promising someone else’s output.

Versioning

The review methodology is versioned (currently 1.1; the method was confirmed 2026-07-16 at 1.0); material changes bump the version, and this page updates in the same cycle with its changelog visible below. If a number of ours is quoted somewhere and doesn’t match this page, this page is current.