Method

What is read, how often, and what a reading cannot tell you.

Every finding Gault reports carries its source and the date it was read. This page is the rest of it: which questions are asked, of whom, how many times, and where the method stops. The numbers below are read from the code that runs the survey, not written by hand — if the product changes, this page changes with it.

What is read

Three areas, per location, from public data only.

Gault is never given an account. It holds no password, no access token and no connection to your Google profile, your website or anything else of yours — a reading is taken the way a customer would see it, from the open web.

18checks on the Google Business Profile — the listing itself: whether it exists, what it says, and whether what it says matches the location. Worth 35% of the location score for an insurance network.
29checks on the website — what the page tells a search engine about who and where this location is, and how fast it answers. Worth 40%.
14checks on AI visibility — the signals an AI answer is assembled from, plus the answers themselves (below). Worth 25%.

The location score is the weighted mean of the areas that were actually measured. An area that could not be read drops out of the divisor rather than scoring zero: a location with no website is not a location with a bad website, and the score says so by leaving it out.

How an AI engine is asked

The real question, from the real town, once per reading.

Gault does not ask an AI engine about your brand. It asks the question a customer would ask, in the location's own city and state, and then reads the answer to see whether that location is in it. For an insurance network the questions are:

1“best insurance agent in Tomball, TX”Carries the score
2“auto insurance near me”
3“home insurance near me”
2–3engines per network, chosen from 4 (Perplexity, Claude, ChatGPT, Gemini). The default set is Claude, ChatGPT, Gemini. There is no all-engines mode: 3 is a hard cap.
1reading per engine, per question, per survey. Gault does not ask the same question several times and average the answers. An engine's answer to the same question can differ between readings, and a single reading cannot measure by how much — which is why a change of a few points between surveys is not a finding, and Gault does not report one as a trend.
Top 5is what “named” means. The engine is asked for the five businesses it would recommend; a location is named if it is one of them. Sixth is not named.
Cityis the finest unit. Every question carries the location's own city and state, passed to the engine as quarantined text that it is told to read as a value and never as an instruction. A national percentage for a brand is not a reading of a branch, and Gault does not publish one.
When Gault says it does not know

Inconclusive is a result. Zero is a different result.

If an engine does not answer — a refusal, a timeout, an outage — that question is counted as unanswered, not as an absence. The appearance rate is the share of the questions that wereanswered in which the location was named. When nothing was answered, the rate is not 0%; it is nothing at all, and the location page says “inconclusive” rather than showing a number that would read as a failure.

The same rule governs the rest of the survey. A profile that cannot be found, a website that will not load, a check that cannot be evaluated: each is marked as unmeasured and leaves the score alone. Gault does not invent a green tick and it does not invent a red one.

How fresh a reading is

30 days, and the date is on every screen.

Every location carries the date and time it was last read, on its own page, in the roster, on the map and in the weekly email. A reading counts as current for 30days — the same window the survey uses to decide what is due — so a location that has gone stale drops out of “read this month” rather than quietly standing in for a current one. Networks choose their own cadence, and any location can be re-read on demand.

What this cannot tell you

Four limits, stated.

OneA reading is a sample of a moving thing. AI answers change between readings, and a single reading per engine cannot say by how much. Treat a one-survey move of a few points as noise.
TwoBeing named in an AI answer is not the same as ranking in Google's local results, and neither predicts the other. Gault measures both because a location can be first on the map and absent from the answer for the same question.
ThreeAnswers vary by who is asking. Gault asks as an ordinary user in the location's city, with no history and no personalisation. A real customer with an account and a search history may see something different.
FourGault reports; it does not act. It never edits a profile, posts a review, or publishes on your behalf — which is what lets it grade work done by whoever did it, including work done by another vendor.

The research behind every check — what it measures and why it matters — is published in full at the research library, and each finding in the product links to the section it came from.

Questions are answered by Aidan Bunch, who built Gault, Monday to Friday, 9 to 6 Central. Questions about any of this, including the awkward ones, go to hello@gaultandco.com.