What an AI visibility audit is, and what ours contains

The mechanic is plainer than the category's marketing suggests. A buyer question goes to an engine. The engine returns a written answer that usually names a handful of vendors and, depending on the engine, a list of sources. Two things get recorded. Whether your name appears in the answer text, which we call named. Whether your domain appears in the answer's sources, which we call cited. Repeat across a set of questions and several engines and you have a named count and a cited count out of a known denominator.

Everything else in an audit is context for those two numbers. Which competitors were named instead. Which questions you were absent from entirely. Which engines behave differently from each other. The audit is worth running because the two headline numbers are usually a surprise, and because they are the numbers that decide whether you enter a deal at all.

They are also not the numbers an SEO audit reports. A ranking audit tells you where you sit in a list of ten blue links. An answer names three or four firms and shows no list. A site can rank respectably and be named zero times. We know that shape well, because it was ours. See GEO vs SEO for the mechanical difference.

There are two free things and they are different sizes. The instant check is five buyer questions through two engines, emailed back in about ten minutes, and it exists to show you the shape of the problem. The full audit is the real instrument.

The four engines are ChatGPT, Claude, Perplexity and Google AI Overviews. One run per question per engine, so ten questions produce forty answers.

The audit does not just tell you where you stand. It tells you who the engines named instead, on which questions, and how far ahead of you they actually are. That share of voice is the number most buyers want first, because two out of forty means one thing when the leader sits at four and something very different when the leader sits at thirty.

We also mark which part of that gap publishing can close. Some of it is a publishing gap and it closes. Some of it is market structure, where an engine is repeating a position a company holds by being large or long established, and no number of answer pages rewrites that. Where the data cannot separate the two, the document says so rather than handing you a plan. Selling a fixable story where the cause is market structure does not survive the first renewal conversation.

No call is required and there is no obligation. The document is the deliverable. We would rather you read it and decide nothing needs doing than sit through a call to receive a number that fits on one line.

This is the same two-page document we send prospects, run on ourselves. Two months of our own results, July and August 2026, on the same ten questions. Where a figure does not exist for our own run we have marked it [SAMPLE] rather than fill it in. Nothing here is modelled and no figure is a projection.

In July we were named in zero of forty answers. In August we were named in one, and that one mention was an engine citing our own measurement page rather than recommending us. We report it that way because the product is the measurement being trustworthy.

Forty answers per month: ten questions across four engines, one run per question per engine. These are the self-audit denominators. They are not the denominators of our four-engine category measurement, which scored 200 answers and 663 citations. The two are never combined.

Ten answers per engine per month. Zero calls failed in either month. Citation floor, disclosed: ChatGPT returned no source list on six of its ten answers in both months, and Google AI Overviews on two of ten in July and one of ten in August. That caps a citation score, not a naming score.

Who the answers named instead. Omniscient Digital was named in 16 of 40 answers in July and 17 of 40 in August. Four other agencies were each named between four and six times out of forty. In both months, 17 of the 40 answers named none of the five. Those five are not a list we chose. They are the five agencies the engines named inside the ten questions on the first run, and we reuse them every month. The full named list is published at GEO agencies measured.

In a client document every row carries the real question text and the per-engine result. We do not reproduce our own ten, because publishing them would let anyone game a series we have committed to running unchanged.

Page two: diagnosis and recommended actions. Absence ladder position: at the category door.

Every source returned across the forty August answers, including repeats across engines. This is not a count of distinct pages, and it is not the 663 citations from the category measurement. [SAMPLE] In a client document a second panel here lists the individual pages the engines cited most often, with a plain label for what each publisher is and the client's own domain highlighted in place. We have not published a per-source breakdown of our own run, so we are not going to put numbers there.

How we ran it. Ten buyer questions for the category "AI search visibility agency", asked of Claude, ChatGPT, Perplexity and Google AI Overviews, one run per question, forty answers per month. July run 27 July 2026. August run 6 August 2026. A company counts as named only on a whole-word match of its name, and as cited only on a hostname match. Prepared by Sairam Sivakumar and the Broadcastwell team. Our own self-audit is published every month at broadcastwell.com/visibility. The client re-score runs quarterly.

The audit runs on the same instrument as our published research, which is the only reason we are willing to describe it in this much detail. Three volumes of it are deposited with the data and the analysis code, so the method can be checked rather than taken on trust.

Scoring rules used in Broadcastwell's own audits and in The 2026 State of Generative Engine Optimization, Volume I. DOI 10.5281/zenodo.21537014.

Two details matter more than they sound. Brand matching is word-boundary and case-sensitive for a single plain alphabetic name, so a company called Pitch does not score a hit on the ordinary word pitch. And an engine that returns no answer is recorded as an error and dropped from the denominator, so a failed call becomes nine out of nine rather than a fake zero out of ten.

The published statistics behind the method, including the figures quoted on this page, are collected on the statistics page.

The audit ends with a document. If it shows nothing worth fixing, that is a real outcome and the correct answer is to do nothing. Most audits do not show that.

If it does show a gap, the work that follows is a monthly program: four answer pages a month, one schema and entity pass, placement in the third-party lists and buyer guides AI engines actually cite, and a monthly report on your brand mentions, your citations and the sources the engines are pulling from. Your ten questions are re-scored across all four engines every quarter. It is $4,500 a month with a three month minimum. The first three clients get a $3,000 a month founding rate in exchange for a named public case study and a review. Three months is one full quarter, which is the shortest period over which the re-score shows a trend rather than noise.

90-day proof gate: if the ten questions we track have not moved by the first quarterly re-score, we do not invoice month four. You keep every run, every answer and every source we collected.

The re-score is the part that matters and it is why the audit is worth doing properly the first time. The same ten questions, the same four engines, the same scoring rules, every quarter. If the number does not move, the report says the number did not move. The first month is a baseline rather than a verdict, and the document says so, because a first reading is where you started and not a result.

There are no case studies, testimonials or outcome claims on this page. What is published instead is three volumes of research and our own monthly score, run with the same instrument and reported the same way.

You also get a named person rather than a queue. Sairam Sivakumar is the contact on the engagement: he sends the findings, writes the recommendations and answers your email. The rest of the team does the delivery work and joins the call when he asks them to. We are not going to tell you anyone is dedicated full time to an engagement that has not started.