AI Visibility Partners

Case file · Local service businesses · US · 2026

How do we measure AI visibility?

By registering the measurement before taking it. The prompt panel, the engines, the session protocol, the pass count, the scoring rule, the three measures, the noise floor, and the publication rule are all written down and dated before the first prompt is run. What follows is that registration, in the form every client receives it.

By Austin Brewer, founder of AI Visibility Partners · Updated 3 September 2026

What is registered before measurement starts?

Nine items, fixed on a dated document that both parties hold before any engine is queried. The point is not ceremony. A measurement whose rules can be adjusted after the results are in cannot show movement, because any result can be made to look like movement. Everything below is the standard registration; the client-specific values (the actual prompt strings, the market, the competitor set) are filled in and signed off at the readout of the Baseline Audit.

Measurement registration · Standard form Version 2026.09 · Supersedes 2026.06
Prompt panel
25 to 50 buyer-intent prompts for the Baseline Audit; 50 to 75 for Growth Authority. Exact strings fixed and listed in the appendix. Phrased as a customer would ask, not as a marketer would search.
Engines
ChatGPT, Perplexity, Gemini, Google AI Overviews. Default model and default settings on the date of the run, recorded per run. An engine added later starts its own baseline.
Session protocol
Signed out, fresh browser profile, no prior conversation, location set to the client's market by the engine's own location control where one exists. Screenshot and full text captured per pass.
Passes per cell
Three. A cell is one prompt on one engine. Each pass is a new session. Passes are run on the same calendar day where possible and the dates are logged where not.
Scoring rule
Each cell scores "named in k of 3" for each measure, k from 0 to 3. Panel-level coverage is the share of cells with k of at least 2. Single-pass hits (k of 1) are reported but never counted as coverage.
Measures
Three, reported separately and never blended: mention coverage, owned-citation coverage, third-party-source presence. Defined below.
Noise floor
Set from the baseline run: the baseline panel is run twice, one week apart, before any work starts, and the difference between those two runs is the noise floor. Later movement smaller than that floor is reported as "within noise".
Re-measure cadence
Monthly, same panel, same protocol, within five days of the baseline's calendar date. A missed month is reported as missed, not interpolated.
Publication rule
The monthly report goes to the client whatever it shows. Where a client has agreed to a public case study, the public write-up follows the same rule: results publish whether they favor us or not, with the noise floor stated beside them.

A copy of this form, with the client-specific appendix, is part of every Baseline Audit deliverable and stays with the client at exit.

Why three passes instead of one?

Because the same prompt to the same engine returns different answers on different runs, and a single run can name your firm by chance. Three passes turn a lucky hit into a countable one. A cell that names the firm once in three is recorded, but it does not count as coverage; a cell that names it in two or three of three does. The rule cuts both ways: a competitor who appears once is not counted as holding that answer either.

Three is a floor chosen for cost. More passes would tighten the estimate; fewer would not produce one.

Scoring sheet · Illustrative Mention coverage
Cell Pass 1 Pass 2 Pass 3 Score
Prompt 07 · ChatGPT named named not 2 of 3
Prompt 07 · Perplexity not named not 1 of 3
Prompt 07 · Gemini not not not 0 of 3
Prompt 07 · AI Overviews named no AIO named 2 of 21 untriggered
1

Two of four cells count as covered. The Perplexity hit is logged and not counted.

Illustrative cells for one prompt. No client data. Real sheets run 100 to 300 cells per month.

What do the three measures mean?

Each answers a different question, and they move independently. An engine can name a firm without citing its site, cite a directory page about the firm without naming it in the prose, or do both. A single blended "visibility score" would hide which of these changed, so the report never produces one.

Mention coverage

The share of cells where the firm's name appears in the answer text in at least two of three passes. The question a customer would ask: did it say your name.

Owned-citation coverage

The share of cells where a page on the firm's own domain is cited as a source in at least two of three passes. The measure most directly moved by work on the firm's site.

Third-party-source presence

The share of cells where a cited source other than the firm's own site (a directory profile, a local publication, an association listing) mentions the firm. The measure moved by authority work off the site.

What does this method not tell you?

It does not tell you how many customers asked, what any one of them saw, or whether a lead came from an AI answer. It measures what a clean session in your market is told, and how that changes month to month. Those limits are stated in every report, and any vendor who claims to have removed them should be asked how.

  • No engine publishes query volume for conversational prompts. The panel measures presence in answers, not how often the question is asked.
  • Signed-in buyers see personalized answers this protocol cannot reproduce. The clean session is a shared floor, not a simulation of any one person.
  • Engines change models and retrieval behavior without notice. A step change across the whole panel in one month is flagged as a likely engine change, not as work that landed.
  • Attribution to leads relies on the client's own analytics and intake notes. The report can show correlation in time; it does not claim cause.

What do clients ask about the method?

Can a client change the prompt panel after registration?

Prompts can be added, never swapped or removed. Added prompts start their own baseline from the month they enter and are reported separately from the original panel until they have three months of history. The original registered panel is re-run unchanged for the life of the engagement, so the first comparison is never lost.

Why signed-out sessions when real buyers are signed in?

A signed-in answer is shaped by that account's history and location, which cannot be reproduced next month. A clean, signed-out session is the only condition two measurements can share. It is a floor, not a simulation of any one buyer, and the report says so.

Do you measure Google AI Overviews the same way?

Yes, with one difference noted in every report: an AI Overview does not appear for every query on every run, so a cell is scored on the passes where one appeared, and the count of passes with no Overview is recorded beside it. A prompt that triggered no Overview in any pass is reported as untriggered, not as a zero.

Want this run on your market?

The Baseline Audit is this registration filled in for your firm: 25 to 50 prompts, three passes per cell, three measures, a noise floor, and a 60-minute readout. $750, yours to keep.

Email us your business name

hello@aivispartners.com · Reply within two business days