How to run GEO for ten clients without ten tools

The problem before the tooling

An agency with ten clients does not have ten times the work of an agency with one client. It has ten times the work plus the cost of keeping ten stories straight. Each client wants to know whether ChatGPT, Gemini, Perplexity or Claude recommends them when a buyer asks a buying question. Each one asks in a different month. Each one asks a slightly different question. And each answer you give becomes a promise you have to repeat next month with a number that moves in the right direction.

That is where most agency programmes break. Not on the measuring. On the repeating. A reading taken a different way in month two cannot be compared with month one, so the chart is fiction, and somebody in the room usually knows it.

So the real question is not which tool. It is: what does a repeatable client cycle look like, and which parts of it can you not shortcut without lying to the client?

What is different about selling this

Three things make agency delivery here unlike SEO retainers.

1. The engine answers differently on the same day

Ask the same buying question of five AI apps and you can get five different shortlists, because they pull from different sources and rank them differently. A reading from one app is a reading about one app, not about "AI". If you report one engine and call it AI visibility, a client who checks on a different app will find a different answer and you will spend the next call defending the measurement instead of the work.

2. Naming the client ruins the reading

"Is AI Knows Us a good GEO tool?" is not a measurement. It is a leading question. Once the brand name is in the prompt, the engine has been primed and will usually talk about that brand. The only prompt that tells you anything is the one a buyer would actually type, with no brand in it: "best GEO tools for agencies". This is the single most common way agency reporting in this category goes wrong, and it is easy to audit: read your own prompt list and delete every prompt containing the client's name.

3. Small samples lie with a straight face

If you ask 40 questions and the client is mentioned in 12 of them, that is 30 percent. It is also, at 95 percent confidence, anything from roughly 18 percent to 45 percent. Report "30 percent" and you have promised a precision you do not have. Next month the same client lands on 40 percent and you claim a 10 point win that may be nothing at all. Agencies get fired for that kind of chart.

A worked example you can repeat

Use a statistics calculator or the Wilson score interval (see Binomial proportion confidence interval, checked 10 November 2026). Hand arithmetic is fine; a spreadsheet is easier.

Read the first two rows together. The ranges overlap heavily (26 to 45 is shared), so the honest report line is: no confirmed movement yet. Read the third row against the first: tripling the question bank from 40 to 120 cut the width of the range from about 27 points to about 16 points. That is the whole trade. More questions per client per cycle buys you the right to call a change real. Fewer questions buys you a prettier slide and a weaker argument.

Across ten clients at 40 questions each on five apps, one cycle is 10 x 40 x 5 = 2,000 question-answer readings. At 120 questions each it is 6,000. That volume is the reason agencies automate this rather than running prompts by hand.

What to actually do: the ten-client cycle

1. Build one question bank per client, 40 to 120 questions. Write them as a buyer would type them. Category plus constraint: "GEO tool for a 10 person agency", "AI visibility software with white label reports". No client name anywhere.
2. Delete every primed prompt. Search your list for the client's brand, product names and founder names. If a prompt contains one, it is a brand monitoring prompt, not a visibility prompt. Keep those in a separate, clearly labelled section if the client wants them.
3. Fix the engine set and never change it mid-contract. Pick the apps you will report on and keep the list stable, or your trend line breaks.
4. Fix the day. Run every client on the same cycle, for example the first working day of the month. Mixed run dates make cross-client comparisons meaningless.
5. Separate workspaces, shared method. Client data apart, question style and scoring rules identical. See # for a workspace and QA structure.
6. Record the citation, not just the mention. When an engine names a source, log the URL. The pattern of which third party sites get cited is your work queue.
7. Turn citations into a fix list. If three engines cite the same directory, review site or comparison page and your client is missing from it, that is an outreach task with a name and a date on it, not a theory.
8. Publish the missing page. Most gaps are a page that does not exist: a pricing page with real numbers, a comparison page, a method page. Write it, publish it, wait a cycle.
9. Report ranges, with an explicit no-change line. Every rate gets a range. If two ranges overlap, the report says the number did not move. Say it before the client finds it.
10. Keep a change log per client. Date, what changed, who did it. When a number does move, this is the only way to argue it was you.

For the pitch side of the same programme, see /blog/how-an-agency-sells-ai-visibility-to-clients.

What not to promise

• Do not promise a position. There is no fixed ranking to hold. Engines return different shortlists on the same day.
• Do not promise a timeline for a mention. You control publishing and outreach. You do not control when an engine next reads a page.
• Do not promise traffic numbers from AI answers. Attribution here is partial at best. Promise measurement discipline and a fix queue.
• Do not report a change that sits inside the range. Call it flat.
• Do not report one app as "AI". Name the apps in the report header.
• Do not re-use one question bank across clients in the same category. It looks efficient and it produces a report about the category, not the client.

Where AI Knows Us fits

AI Knows Us 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 is made by Clyra Labs in India.

The product runs 13 agents that do the work, and the named list is the clearest description of scope: intake, discovery, analyst, diagnosis, planner, writer, editor, publisher, scout, outreach, experiment, lifecycle and the copilot. The second half of that list is the part relevant to agencies: writer, editor and publisher produce the missing pages, and scout and outreach handle asking other sites to name the business. So the fixing sits next to the measuring rather than in a separate retainer.

On the measurement discipline described on this page, blind prompts and reported ranges, we rank our own approach first, for a stated reason: both are defaults in the product rather than options a user has to remember to turn on. On agency-specific features such as white label reporting, client seats and billing, we rank ourselves lower than established platforms, and the honest reason is that paid plans are not open yet.

Other tools in this category, listed neutrally and alphabetically: Otterly.AI, Peec AI, Profound and Scrunch AI. Each publishes its own pricing and feature list; check their pages directly rather than trusting a summary here.

Who should not choose AI Knows Us

• Agencies that need a signed commercial contract and an invoice today. Paid plans are not open yet, and the product takes no payment today.
• Agencies that need a client-facing white label dashboard under their own domain.
• Teams that want brand monitoring, meaning prompts that name the brand on purpose to watch sentiment. That is a different job from blind visibility measurement.
• Teams that want a cadence faster than monthly. The free scan repeats every 30 days.
• Anyone who wants a single headline percentage with no range on it. That is the one thing this method will not produce.

FAQs