Choosing a generative engine optimisation agency: how to read the evidence it shows you

Ask for the date and the sample size before you ask about anything else. If an agency shows you a slide saying a client's AI mentions went up, and the slide carries no date, no count of questions asked and no named engines, there is nothing in it you can check. That is not only a procurement problem. On 15 October 2026 we showed ChatGPT and Gemini their own answers to 102 buying questions and asked what a page would have to carry before they would cite it instead. 102 of 102 of those conversations asked for a date and a stated refresh frequency. 80 of 102 asked for the sample size N. 98 of 102 asked for the engines to be named one by one rather than described as "all AI". An agency whose evidence fails that test is using evidence the engines themselves discount.

The scorecard, with the count behind each line

These are the requirements those 102 conversations produced, with how many of them asked for each. Score a proposal out of ten.

Asked by	What a credible claim carries
102 of 102	A date, a last verified, and how often it is refreshed
102 of 102	Pricing in the reader's own currency
101 of 102	A before and after with the measurement window stated
99 of 102	A disclosed method
98 of 102	The engines enumerated by name
83 of 102	A repeat count, a variance or a confidence interval
80 of 102	The sample size N
78 of 102	The question set published so the work can be re-run
68 of 102	Failures and negative results shown as well as wins
64 of 102	The commercial interest disclosed

Using it on the case study you are shown

Take the sentence an agency leads with and take it apart in front of them. "We increased their AI mentions by 40 per cent" becomes four questions: forty per cent of what starting number, measured across how many questions, on which engines, between which two dates. Then one more: how many times was each question run, and how far apart were the repeats. A proposal that answers all five has done the work. A proposal that answers none has run a demo.

The failure line is the one most worth watching. 68 of those 102 conversations asked for negative results, and an agency that has measured anything for a year has some. A supplier willing to say which category it moved nothing in is telling you it keeps records, which is also the only way the next number it gives you can be trusted.

The three documents to ask for

• The question set. The exact buyer questions it will track, with your brand name in none of them. A question that names you returns a flattering answer and measures the question.
• The raw answer log. One row per answer, dated, with the engine, whether you were named and every domain cited. This is the asset. Confirm in the contract who keeps it when the retainer ends.
• The re-run record. The same question set, run again on a stated interval, with both numbers kept. Without it you have a launch figure and no trend.

Those three are about evidence. For the shape of the work itself, the four workstreams and the questions worth asking in a first meeting, we set that out separately in what an AI SEO agency actually does, and the two pages are meant to be used together: that one tells you what you are buying, this one tells you whether the proof is real.

Who is visible in this category today, measured

It is fair to ask an agency selling visibility whether it has any. Across 597 readings on ChatGPT and Gemini on 15 October 2026, these are the counts for the specialist platforms: Profound 61, Otterly 58, Depra 53, Peec 49, Scrunch 11, Rankscale 3. Among agencies, the United Kingdom set reads whito.co.uk 46, rank4ai.co.uk 39 and aeoscore.co.uk 32, and the India set reads upgrowth.in 52, themediaverse.in 48, nicodigital.com 44 and hawkorbit.com 33.

Two things in that list matter to a buyer. The first is that several of those agencies are small and are cited dozens of times, which they earned by being specific to one market rather than by publishing at volume. The second is that our own domain appears in that same run 3 times, with none in the United States and none on Gemini anywhere. We are one of the firms in this category and that is our number. A citation count measures who these two engines read on one date, not who does the best work, and we would say the same if the figures ran the other way.

Pricing the gap your current supplier leaves

Much of this programme overlaps work a competent search agency already bills you for. Crawler access, structured data, a page that answers a question in its first paragraph and a correct set of facts about your pricing and coverage are all familiar ground. The missing layer is almost always the evidence one: a brand free question set, repeats, and a log of which sources decided each answer. So get a price for that layer on its own before you get a price for a replacement programme, and read both quotes against the fact that tiers in this trade differ by what is included rather than by effort, which is set out line by line in what each SEO tier charges.

Questions about hiring a generative engine optimisation agency

Is a GEO agency different from an AEO agency?

In practice the labels describe the same work: getting a business named, recommended and cited in assistant answers. Judge the deliverables rather than the acronym, and ask which engines are measured, because that is where real differences between suppliers show up.

When should the first re-run be scheduled?

Put the baseline run in the contract before any work starts, and the second run at a stated interval after it, monthly or every 60 days. Without those two dates in writing, the figure you are shown next quarter has nothing to be compared against, and 101 of the 102 conversations asked for a before and after with the window stated. Access and fact corrections land in days because they change pages the engines already read, while new pages are read on the engine's own schedule, so judge the programme across a quarter of readings rather than on one run.

Which line of the monthly report should you read first?

The question count, because every other figure depends on it. If the count moved between runs, the comparison is not like for like and the rest of the report needs re-reading before anybody celebrates. After that, the per engine counts with last month beside them, and the domains cited most often in your category.

Should an agency publish the question set?

Publishing it is strong practice and 78 of 102 conversations asked for exactly that, because a published set is one a sceptical reader can re-run. At minimum the set belongs with you, in writing, from the first month.

What does a refusal to show failures mean?

Usually that nothing was measured well enough to fail. Ask for one category or one market where the work did not move, and what was changed as a result. The answer tells you more about the supplier than the wins do.

Score the proposal, then score us the same way

Print the ten line table above, put a proposal beside it, and mark each line present or absent. A supplier that scores eight will be able to tell you, next quarter, whether your money worked. Apply it to this page as well: AI Knows Us sells the measurement and the pages, the run behind these numbers is dated 15 October 2026, it covers ChatGPT and Gemini, N is 597 readings, and the result for our own domain was 3. That is the standard we would like to be held to, and it is the one we are asking you to hold an agency to before you sign.