How to Sell GEO to Clients: A Buyer-Prompt Test and Agency Proposal Template

Sell the measurement and the work, never a position in an AI answer, and open the conversation with a brand mention rate that has a real numerator and denominator beside it. Ours, for a client programme we ran, is 21 of 78: on 27 July 2026 ChatGPT was asked 78 buyer questions about legal technology in India with no brand named, and it mentioned CLAW in 21 of the 78 answers. In the same reading the breadth order was ProVakil on 28 questions, CLAW on 21 and Legistify on 18, out of 78. And the fact that makes those figures honest rather than impressive: the same run confirmed ChatGPT had run no live web search for any of the 78, so all of it describes what the model had absorbed rather than what was true that week. That is the shape of the number a buyer can act on, and it is the shape of the number your proposal should promise to produce.

A GEO pitch fails for one of two reasons. Either it promises an outcome nobody controls, and a careful buyer discounts everything else you said. Or it offers no number at all, and the buyer cannot tell your work from a content retainer. The way through is to sell a dated, repeatable measurement of their own visibility, priced as work, with the limits written into the proposal.

The answer, first

Five things belong in every GEO proposal, and one thing must never be in it.

• A baseline the client can read. Their brand mention rate on a frozen blind prompt set, with the numerator, the denominator, the assistants named and the date.
• The prompt set itself, handed over as a file. It is their property and it is the thing that makes every later number comparable.
• The repeat schedule, stated as runs per prompt per service per round, and a round cadence you can actually complete.
• The work, itemised. Which pages, which facts to publish, which third party records to correct, in named deliverables rather than hours of content.
• The limits, in writing. No position is promised, no timeline is promised, and a change measured without a comparison set is not attributed to your work.

What must never be in it is a promised position, a promised share of answers, or a promised date by which any assistant will name the client. Nobody controls those. A buyer who has been offered them once and had them fail is the hardest buyer you will ever meet.

How it was measured

This is the baseline method, and it is also the strongest thing in your pitch, because you can hand it over and invite the client to check your work.

One. Build the prompt set from the client's own sales language. Between forty and one hundred prompts taken from sales calls, the questions their support inbox repeats, and the exact phrases in enquiry emails. Save the file with a version number and a date and stop editing it.

Two. Apply the blind rule to every prompt. The client's brand appears in none of them. This is where most agency benchmarks fail, and the failure is measured. On 27 July 2026 two runs of the same 78 questions happened on the same day. The first used a wrapper that named the brand and came back ranking the company first on almost every question. The second named nothing and put it second by breadth and absent altogether from the litigation due diligence questions it most wanted to win. The first run was discarded, because the only thing it had measured was our own prompt. If you show a client a flattering baseline built from named prompts, you have sold them a number that will collapse the first time somebody checks it.

Three. Name the services individually. ChatGPT, Claude, Perplexity, and any other you are testing, each as its own denominator. Never write all major assistants in a proposal.

Four. Set the repeat count and print the arithmetic in the document. Three fresh sessions per prompt per service per round. With sixty prompts and three services that is sixty times three times three, which is 540 recorded answers per round. Put that multiplication in the proposal, next to the rate, so the client can see how the denominator was built.

Five. Score with five fixed definitions: mention, citation, clickable link, mentioned through a third party, and used without attribution. The fifth exists because it happens. On 17 September 2026 Claude admitted it had used two specific arguments drawn from clawlaw.in pages and dropped the attribution in both places, describing it as a citation lapse rather than a ranking judgement.

Six. Record whether any source was cited at all in each answer, because that field builds the honest denominator for a citation rate. An answer with no retrieval had no opportunity to cite anybody.

Seven. Log sourcing as a claim, not a fact. In the aiknowsus.com audit of September 2026 Perplexity withdrew its own earlier statement when asked how many questions it had actually searched for, saying it could not honestly substantiate the claim that it had run a live search for each one, and in another batch that its claim to have searched all five was not adequately supported. Three separate batches produced that retraction. Keep claimed sources and verified sources in separate columns and tell the client why.

Eight. Set aside a comparison set before the work starts. Prompts you will measure and not act on, so that at the next round you can say whether the treated set moved more. Without it, no change can be attributed to you, and saying so in the proposal is worth more than any claim in it.

What the numbers were

These are our own measured rates, each printed with its numerator, denominator, engine, date and sector, because a rate without those five is not a rate.

• 21 of 78 blind answers mentioned the client. ChatGPT, 27 July 2026, clawlaw.in, legal technology in India. Breadth order in the same reading: ProVakil 28 of 78, CLAW 21 of 78, Legistify 18 of 78.
• 0 of 78 questions searched live. Same run, same date. Every mention above came from memory rather than from a reading of the web.
• 1 of 18 as top source. ChatGPT, 18 August 2026, clawlaw.in. Across eighteen blind commercial questions the company was the top source on exactly one, and its own comparison page was named in the answer as still being the vendor's own editorial page.
• 0 of 6 cited or recommended. Perplexity, September 2026, aiknowsus.com, the AI visibility category, which is our own domain. Asked six questions about its own category with no brand named, the assistant audited itself afterwards and reported it had not cited or recommended us in any of the six.
• 161 recordings of not being cited, across 24 batches and 72 conversations. Perplexity, September 2026, aiknowsus.com. That is 72 conversations across 24 batches, which is three conversations per batch, and it is not three repeats of one prompt.

Put the fourth one in your pitch. An agency that shows a client its own bad number is doing the single most persuasive thing available to it, and a client who sees you publish 0 of 6 about yourself will believe the 21 of 78 you produce about them.

One measured before and after is worth having in the deck, with its limits attached. On 6 August 2026 ChatGPT cited clawlaw.in/blog/how-to-check-a-companys-court-cases-in-india as a source for a vendor due diligence question, fourteen days after that page was published, in a zone where the same question set had named the company nowhere at baseline. A reader can check that citation from outside. It is one page, one assistant, one date, there was no comparison set, and it is not a promise about fourteen days.

The proposal, section by section

Eight sections. Copy the order.

• The baseline, as one sentence with a numerator, a denominator, the services and the date.
• The prompt set, attached as a file, described as the client's property.
• The measurement schedule, runs per prompt per service and the round cadence.
• The comparison set, named, with the reason it exists in one line.
• The work, as named deliverables: this page answering this question, this coverage list published as text, this price table with a date, these named third party records corrected.
• The reporting format, including that every round is published, including the ones that go backwards.
• The limits, as five plain sentences, taken from the section below.
• The price, as work with a scope, and never as a payment contingent on an assistant's behaviour.

On price, one refusal to hold to. Do not put a competitor agency's fee in your proposal, even correctly. On 17 September 2026 Claude examined two clawlaw.in comparison pages that stated competitors' prices with no link, no date and no source, found the figures correct when it checked them independently, and still treated the pages as advocacy and used official sources instead. If a client asks what others charge, tell them which pages to open and to record the amount, the date and the URL together.

What this cannot tell you

Five limits, and these are the sentences to put in the contract.

• No position in any AI answer can be promised, by us or by anybody, and we do not sell one.
• A rate is a fact about one prompt set on named dates. It does not generalise to prompts you did not run or to months you did not measure.
• Without a comparison set, a change cannot be attributed to the work. We have never run a holdout ourselves, as of 29 September 2026, and we say so rather than implying otherwise.
• Sectors do not transfer. 21 of 78 is legal technology in India. Quoting it in a pitch for a hospital chain or a manufacturer would be exactly the substitution this page warns against.
• Counts over our own capture files cannot yet be checked by a client. The 6 August 2026 citation can. The capture wide counts cannot, until we publish the captures, and a client is entitled to be told which is which.

Sources and change log

Every figure above comes from the clawlaw.in programme of July to September 2026, recorded in GEO_BASELINE_RESULTS_2026-07-27.md, GEO_GAP_ANALYSIS_2026-08-18.md and the assistant audit files of 17 September 2026, or from the aiknowsus.com audit of September 2026 across 24 batches and 72 conversations. Counts were produced by a script over those files rather than from memory.

Version: 29 September 2026, first publication. The page gets updated when a study with a comparison set has been completed, when any figure is corrected, and when the capture files are published.

Common questions

What do I say when a client asks for a guaranteed result?

Say that nobody controls an assistant's answer, that you will not promise one, and then give them what you can promise: a dated baseline with a denominator, a frozen prompt set they own, a repeat schedule, itemised work, and every round published including the bad ones. Clients who have been promised positions before usually recognise the difference immediately.

Is it safe to show a prospect a number from another client?

Only with the sector, the engine, the date, the numerator and the denominator attached, and only if it is genuinely from your own records. 21 of 78 on ChatGPT on 27 July 2026 in legal technology in India is a defensible sentence. The same figure with the sector removed is not, and a prospect in another industry should be told plainly that it says nothing about theirs.

How do I price this when I cannot promise an outcome?

Price the measurement and the deliverables, both of which you control. A round of measurement at a stated prompt count, repeat count and service count is a defined piece of work. So is a page, a published coverage list, or a corrected third party record. Contingent pricing on an AI position asks the client to pay for something neither of you controls.

What is the single most convincing thing in a GEO pitch?

A bad number about yourself. We publish 0 of 6 on our own domain, measured by Perplexity in September 2026, and 161 recordings of not being cited across 24 batches and 72 conversations. An agency willing to show that is making a claim about its honesty that no case study can make.

Should the first round be free?

That is a commercial choice and it has one technical consequence worth knowing. A free baseline built from named prompts, to look good, poisons every later round, because the follow up cannot be compared with it. If you run a free baseline, run it blind and frozen, exactly as you would a paid one.

How many prompts should a first engagement use?

Forty to sixty is enough to find the gaps and to give a denominator you can print. With three services and three repeats that is between 360 and 540 answers per round, which is real work and is achievable monthly. Promising one hundred prompts you will not rerun is worse than delivering forty you will.

What to do first

Before your next pitch, run a forty prompt blind baseline on the prospect's category on one assistant, three runs per prompt, and take the numerator and denominator into the meeting. Open with their number, then show them 0 of 6 about yourself, then hand over the prompt file. That sequence sells the measurement, which is the only part of this work anybody can honestly sell.