AI visibility for agencies: the service, the protocol, and how to price it

AI visibility is a sellable service line for an agency because it produces a finding the client cannot get anywhere else: the names an assistant gives when their buyers ask, and the sources behind those names. To productise it you need four things, a fixed prompt set per client, a measurement protocol you run the same way every month, a reporting format that survives a sceptical client, and a scope of work written as work delivered rather than positions achieved. This page covers all four, plus the pricing structure and the QA that keeps twenty clients consistent.

Why it sells

Three things make this an easy conversation with an existing client.

• The finding is concrete and new. A screenshot of an assistant naming three competitors, with the client absent, needs no explanation.
• It does not cannibalise the retainer. The technical and content work overlaps with SEO, so the delivery capacity you already have does most of it.
• Nobody can audit themselves easily. A client who tries it alone will ask about their own company by name, see themselves everywhere, and conclude all is well. A blind prompt set is a methodology you own.

The sales conversation and a proposal shape are in how to sell GEO to clients.

The service, in deliverables

Deliverable	What is in it	Cadence
Prompt set	15 to 30 buyer questions in the client's buyers' words, brand name in none of them, signed off by the client	Once, reviewed quarterly
Baseline audit	Crawler access, index eligibility, quotable page review, entity consistency, and a first reading per engine	Once at the start
Measurement round	Every prompt, every engine, three runs each, in clean sessions, with all cited sources logged	Monthly
Citation target list	Counted list of domains cited across all answers, with the ones the client is missing from marked	Monthly
Content production	One page per unwon buyer question, title as the question, answer in the first paragraph	Monthly, a set number of pages
Off site work	Profile claims, directory and review platform listings, pitches to the most cited publications	Monthly, a set number of placements attempted
Report	Mention rate per engine with a range, share of voice, citation share, work done, work next	Monthly

Write the scope in those units. Pages published, placements attempted, prompts measured. Those are things you control and can invoice against. A retainer benchmark is in what a GEO retainer should include.

The measurement protocol, written down once

Consistency is the whole value. Put these rules in a one page document and have every analyst follow it.

• Blind prompts. The client's name appears in no prompt. This is the rule that makes the number mean anything, and the reasoning is in why you must never name your own brand in a tracking question.
• Clean sessions. Temporary chats, memory off, custom instructions off, and a consistent locale per client.
• Three runs minimum, per prompt per engine. Report a range and compare ranges. When two ranges overlap, the report says the number has not moved.
• Same window every month. Pick the first three working days and keep them.
• Log the sources, always. Every cited domain from every answer. This is the most actionable output you produce.

Running this across many clients without drift is its own problem, and we worked through it in running AI visibility across 20 clients.

Reporting a client will accept

Four rules keep a monthly report credible for a year.

• Every rate carries a range. A jump from 20 to 30 per cent on nine runs is noise, and saying so in month two buys you trust in month eight.
• Separate the engines. One table, one row per engine. Averaging five surfaces hides the only one the client's buyers use.
• Put the work next to the numbers. Pages published, listings claimed, pitches sent. On a long cycle the work log is what justifies the invoice in the months the rate is flat.
• Show the citation list. Clients understand "these eleven sites are what the assistants read, and you are on four of them" immediately.

A full format is in how to report AI search visibility.

Pricing and white label

Price the units, not the outcome. The three cost drivers are the number of prompts, the number of engines, and the number of pages produced each month, so a tiered scope built from those three is simple to quote and simple to expand. Keep measurement and production as separate lines on the proposal, because some clients have their own writers and want only the measurement.

On white label, the practical requirements are your logo on the report, your domain on the share link, a client facing summary that names no vendor, and an export of the raw rows so you can rebuild the report your own way. Published feature and price comparisons are in white label AI visibility reports. AI Knows Us covers the measurement across five surfaces and the page production, and its free scan needs no account, which makes it a reasonable way to produce a first finding for a prospect before anybody signs anything.

This page covers productising AI visibility inside an agency. For the delivery method itself, see LLM optimization. For how buyers judge this kind of agency, see AI SEO agency.

Asked and answered: AI visibility for agencies

Can I sell this to existing SEO clients?

Yes, and it is the easiest version of the sale. Run a blind prompt round on their top five buyer questions, bring the screenshots to the next monthly meeting, and let the absence of their name make the case.

What should I promise in the contract?

Promise the work: a measured prompt set, a stated number of pages published, a stated number of listing and pitch attempts, and a monthly report in a fixed format. Google's own guidance tells buyers to beware of anyone claiming to guarantee a number one ranking, and a buyer should treat a claimed AI mention the same way. Scope in units of work and you never have that argument.

How many prompts per client?

Fifteen to thirty for most clients. Multi location or multi product clients need a set per market or per line, which is where the per prompt cost driver starts to matter in your pricing.

Do I need a tool, or can my team do it by hand?

By hand works for two or three clients and teaches your team what the data means. Past that, the run count multiplies fast, and history, competitor rows and a citation ledger are what a tool is buying you.

How do I handle a client whose rate does not move in month one?

Set the expectation at the start that access fixes show in fetched answers within days of the next crawl, new pages over a few weeks, and third party presence over months. Then report the work log every month so progress is visible while the rate is still settling.

What to do first

• Pick one existing client and write 15 blind buyer prompts with them in a 30 minute call.
• Run three rounds across two engines and build the citation list. That is your pitch deck.
• Write the protocol page and the report template once, then use them for every client.
• Price three tiers off prompt count, engine count and pages per month.
• Decide your white label requirements before you trial tooling, so you test against your own list rather than a feature grid.