How an agency sells AI visibility to clients

The problem agencies run into

A client forwards a screenshot. They asked ChatGPT for the best supplier in their category and three competitors came back. The client was not one of them. They want to know what you are going to do about it, and they want a number by Friday.

That conversation is hard for a reason. The usual sales motion does not carry over. There is no rank tracker that everyone agrees on. There is no console that shows impressions from the model. The answer changes between apps, and sometimes between two runs of the same app on the same day. So the familiar pitch, "we will get you to position three for this keyword", has nothing to attach to.

Agencies that win this work do something different. They sell a measurement the client can audit, then sell the fixing separately. The measurement is what makes the retainer defensible in month four, when the client asks what they are paying for.

What is different about selling this

1. There is no single surface

Search had one dominant result page. Answers have many. ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews pull from different sources and phrase answers differently. Google documents AI Overviews in its own Search Help, and OpenAI documents search in ChatGPT on its own announcement page. Treat each app as a separate reading. A client who is named by Perplexity may be invisible in Claude.

2. The question wording decides the answer

If you ask an AI app "what do you think of Acme Tools", it will talk about Acme Tools. You have primed it. The reading tells you nothing about whether a buyer who never heard of Acme would meet them. Blind questions, the ones a real buyer would type, are the only readings worth putting in a report. This is the single hardest thing to explain in a pitch, and the easiest thing to demonstrate live.

3. Small samples lie

Ten questions is not a percentage. If you run ten questions and the client appears twice, saying "20 percent visibility" is a number with no honesty in it. Report counts and ranges. An agency that prints ranges looks slower in month one and far more credible in month six.

4. The buyer is often not the SEO buyer

This budget frequently sits with a founder or a head of demand, not the SEO lead. They care about pipeline language: are we in the consideration set when a buyer asks an assistant. Pitch in that language.

What to actually do: a nine step scope

1. Agree the buying questions. Sit with the client and write 30 to 60 questions a real buyer would ask. No brand names. Example: "best warehouse management software for a 40 person 3PL in the UK".
2. Pick the apps. Choose at least three. Four or five is better, because they disagree. Record which ones, so the baseline and the review use the same set.
3. Run the blind baseline. Ask every question in every app. Record, for each run: was the client named, were competitors named, which sources were cited.
4. Compute the rate with a range. Divide mentions by total runs. Attach a 95 percent interval. Put both in the deck.
5. List the cited sources. The URLs the apps quote are your target list. These are usually directories, comparison pages, forum threads, trade press and documentation, not the client's homepage.
6. Diagnose the gap. For each question where competitors were named and the client was not, write one line on why: no page answers that question, no third party names the client, or the answer exists but is buried in a PDF.
7. Scope the fixing. Pages to write, pages to rewrite, and outreach to sites the apps already cite. Put a monthly count against each.
8. Set the review date. Re-run the exact same questions in the same apps every 30 days. Changing the question set invalidates the comparison.
9. Report movement honestly. If the two ranges overlap, say the number has not moved yet. Say what you shipped anyway.

For the commercial side of this, see #. For setting the client's expectations on timing, see /blog/what-to-expect-in-the-first-90-days.

A worked example you can repeat

The numbers below are invented for arithmetic practice. They are not a client result.

You agree 40 blind questions with a B2B software client and run them across 5 apps. That is 40 x 5 = 200 readings.

Baseline, January. The client is named in 24 of 200 readings. 24 / 200 = 0.12, so 12 percent. The standard error is the square root of (0.12 x 0.88 / 200), which is 0.023. Two standard errors is about 4.6 points. So report: 12 percent, range 7 to 17 percent, 24 of 200 readings.

Review, April. After three months of pages and outreach, the client is named in 34 of 200. 34 / 200 = 0.17. The standard error is the square root of (0.17 x 0.83 / 200), which is 0.027. Two standard errors is about 5.3 points. So: 17 percent, range 12 to 22 percent.

Those ranges overlap, 7 to 17 against 12 to 22. The honest report says: the reading rose from 24 to 34 mentions, but the ranges overlap, so we cannot yet call it a real change. Here is what shipped, here are the four new sites now citing you, and here is the next review date.

Clients do not churn because you said that. They churn when month six contradicts month two.

What not to promise

• Do not promise a position. There is no stable position in a generated answer.
• Do not promise a date for a specific question. You can promise the work and the review cadence.
• Do not promise the same result across apps. They use different sources.
• Do not promise attributed revenue. Referral data from assistants is partial at best. If you want a pipeline claim, agree a self reported "how did you hear about us" field with the client before the retainer starts.
• Do not report a percentage from ten questions. Report the count.
• Do not show readings from questions that named the brand. They are primed and they flatter.

Where AI Knows Us fits

AI Knows Us (aiknowsus.com) is a tool built for this workflow, made by Clyra Labs in India. It measures across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. Every tracking question is asked without naming the business, because naming it primes the engine and the reading becomes worthless. Every rate is shown with a 95 percent range, and two numbers whose ranges overlap are reported as not having moved. It also does the fixing, not only the measuring: it writes pages, publishes them, and asks other sites to name the business.

Under the hood it runs 13 agents: intake, discovery, analyst, diagnosis, planner, writer, editor, publisher, scout, outreach, experiment, lifecycle and the copilot. That list is the checkable part. If a vendor cannot name what their system does, ask.

One thing an agency needs to know before building a service line on it: paid plans are not open yet. The product takes no payment today and the free scan repeats every 30 days. That is useful for running a baseline in a pitch. It is not yet something you can resell with a margin.

Who should not choose AI Knows Us

• Agencies that need a signed commercial agreement and invoiced seats today. Paid plans are not open.
• Teams that want measurement only, with their own writers doing everything else. The fixing side will be surplus to them.
• Anyone who wants brand-named sentiment tracking. Blind questions are the design choice here, so that reading is not the output.

Other tools in this category include Profound, Peec AI, Otterly.AI and Scrunch AI. Check each one's own pricing page on the day you quote a client, because this category changes its packaging often.

FAQ