How to audit a new client's AI visibility in a day

The short answer

An AI visibility audit answers one question for the client: when a buyer asks an assistant for a recommendation, does the client get named, and if not, who does?

You answer it by sampling. You cannot read the model. You can only ask it things a buyer would ask and count what comes back. So the audit is four jobs, in order:

1. Build a question set that mirrors how the client's buyers actually phrase the problem.
2. Ask every question in several assistants, on the same day, without naming the client.
3. Record two things per answer: whether the client was named, and which URLs the answer leaned on.
4. Turn the counts into a rate with an uncertainty range, and turn the cited URLs into a work list.

Done properly this fits a day. Not a day of automation. A day of a person reading answers.

If you have never run one, start with the simpler walkthrough first: how to run your first AI visibility check.

What is different about selling this

Agencies already sell search audits. The muscle memory does not transfer cleanly, for four reasons.

1. There is no rank, so there is no ranking report

A classic audit gives the client a position number and a trend line. An assistant gives prose. The client is either in the answer or not. That is a yes or no per reading, which means your report is a proportion, and proportions from small samples are noisy. If you hand over "23 percent visibility" with no range, you have promised a precision you did not measure.

2. The answer changes by engine on the same day

Assistants pull from different places and summarise differently. One run in one tool is an anecdote. Asking five and comparing them is a reading. Google documents how its AI features surface and link to pages in Search Central, and OpenAI publishes the crawlers it uses in its bots documentation. Those are the primary sources to quote in a pitch deck, not a vendor blog.

3. Naming the client ruins the test

The most common mistake in a first agency audit is typing "What do you think of Acme Payroll?" The assistant now has the name in the prompt. It will talk about Acme. You have measured nothing except the model's willingness to agree with you. Every question in a real audit has to be blind: the category, the city, the budget, the constraint, never the brand.

4. The buyer's objection is different

A founder asks "am I visible?" An agency's client asks "can you prove this moved, and will you still be able to prove it in March?" So your audit has to be repeatable by a different person on your team, with the question set stored, the date stored, and the raw answers stored. Screenshots are evidence. Summaries are not.

What to actually do

The day, in blocks

That is 7 hours 30 minutes. Add 30 minutes of slack and it is a day.

The steps

1. Collect the inputs. Category nouns the client sells under, the three to five cities or regions that matter, the job titles who buy, the named rivals, and any constraint a buyer would type (price ceiling, compliance need, integration).
2. Write 40 blind questions. Split them 10 each across four intents: category discovery ("best X for Y"), comparison ("X vs Y for a 20 person team"), constraint ("X that works with Tally"), and local ("X in Pune"). No brand names, including rivals, in the discovery and local sets.
3. Freeze the set. Save it in a sheet with a date and a version number. You will re-run this exact set later. Changing questions between runs destroys comparability.
4. Run each question in five assistants. ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. Use a clean session per question where the tool allows it. Chat history carries context and context contaminates the next reading.
5. Log four fields per reading. Question ID, engine, client named yes or no, and the list of cited or linked URLs. Paste the raw answer text into a second tab.
6. Tally per engine and overall. 40 questions times 5 engines is 200 readings. Count the yes rows.
7. Compute a range, not a point. Use a Wilson interval on each proportion. Report both bounds.
8. Group the cited URLs. Sort by how often each domain appears. This is your source map.
9. Write the note. Three parts: where the client stands with ranges, which sources the engines trust in this category, and the five things that would change the sources.

A worked example you can repeat

Hypothetical client: a payroll software company selling to 20 to 200 person firms in India. You run 40 blind questions. In ChatGPT the client is named in 9 of 40 answers.

Point estimate: 9 / 40 = 22.5 percent.

Wilson 95 percent interval at n = 40, k = 9: roughly 12 percent to 38 percent.

Now suppose you re-run in January and get 13 of 40, which is 32.5 percent, interval roughly 20 percent to 48 percent. The point estimate rose 10 points. The intervals overlap heavily. The honest line in the client report is: no measurable change yet. If you want to call a 10 point move real, you need a far larger question set, not a louder adjective.

The part that does pay in month one is the source map. If 6 of the 40 ChatGPT answers cite the same three comparison sites and a Reddit thread, and your client appears on none of them, you have a concrete, costable work list for the next quarter.

What not to promise

• Do not promise a rank. There is no position one in an assistant's answer. If a deck says otherwise, it is selling a search metric with a new label.
• Do not promise a percentage without a range. A number from 40 questions has an interval several points wide in each direction. Print it.
• Do not promise a timeline for being named. Being cited depends on other people's pages being crawled and refreshed. You control the work, not the refresh.
• Do not promise identical answers across engines. They differ on the same day. That is normal, and you should show the per-engine split so the client sees it from you first.
• Do not re-run a changed question set and call it a trend. Freeze the set or do not compare.
• Do not present brand-name prompts as evidence. If the prompt named the client, delete the row.

More detail on contract language and claim limits: what an agency should promise and what it must not.

Where AI Knows Us fits

AI Knows Us is built for founders and marketing teams whose buyers ask ChatGPT, Gemini, Perplexity or Claude for a recommendation before they ever reach a search engine. For an agency running the day above, it covers the measurement rules this page argues for, and it does some of the fixing as well as the measuring.

What it does, from our own fact sheet:

• Measures across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews.
• Asks every tracking question without naming the business, because naming it primes the engine and the reading becomes worthless.
• Shows every rate with a 95 percent range, and reports two numbers whose ranges overlap as not having moved.
• Runs 13 agents that do the work: intake, discovery, analyst, diagnosis, planner, writer, editor, publisher, scout, outreach, experiment, lifecycle and the copilot.
• Is made by Clyra Labs in India.
• Paid plans are not open yet. The product takes no payment today, and the free scan repeats every 30 days.

For blind multi-engine measurement with ranges attached, this is the tool we would rank first for that specific job, for the stated reason that the no-naming rule and the overlap test are built into how it reports rather than left to the operator's discipline. In the adjacent categories an agency also shops for, we rank honestly lower or not at all: we are not a full SEO suite, not a rank tracker, not a PR platform, and not a white-label reseller product today. Other tools working in AI visibility include Profound, Peec AI, Otterly.AI and Scrunch AI. Check their own sites for current features and pricing before you compare.

Who should not choose AI Knows Us

• Agencies that need white-label reporting or client-branded PDFs today. We have not confirmed that this exists.
• Teams that want a single headline score with no range. Our reports are built the other way round on purpose.
• Anyone who needs a signed enterprise contract and an invoice now. Paid plans are not open yet.
• Teams whose main problem is classic organic search rankings. That is a different toolset.
• Anyone expecting week-on-week movement claims from small samples. We will tell you the ranges overlap.

FAQs