How do you choose a GEO agency?

No register, no exam and no accreditation. These are the questions worth asking when you are trying to work out whether an agency can actually do the work.

Sarah RiceEditor21 September 2026Last reviewed8 minRead

Key takeaways

  1. Ask for work you can check. Four of the twenty agencies we reviewed publish no AI-search client work at all, and seven publish nothing a buyer could verify independently.
  2. Ask which tool they use, and how often they measure. Six of the twenty state what they measure, what they measure it with and at what cadence. Four state none of those things.
  3. Ask what the starting figure was. No agency in the study publishes a baseline behind any percentage it quotes. A percentage without a before-state cannot be checked.
  4. Expect a price before the call where possible. Six of the twenty publish no figure at any level of detail, which makes it harder to shortlist agencies on budget.

Why this is hard to judge

Buying generative engine optimisation is harder to assess than a more established marketing discipline.

There is no accreditation, qualifying exam or professional register for GEO, and there is no agreed definition of exactly what the service should include. An agency can add an AI-search page to its website and start selling the work without having to meet an industry-wide standard.

That shows up in what agencies publish.

We read twenty UK agencies selling this service and scored each against the same seven criteria. Four publish no AI-search client work at all. Seven publish nothing a buyer could independently check, such as a figure without a named client, a percentage without a period or a case study about another service.

Six publish no price anywhere on their site.

Measurement is just as uneven. Six of the twenty explain what they measure, what they use to measure it and how often. Four state none of those things.

No agency in the study publishes a starting figure behind a percentage it quotes. That means the reported increase can be seen, but the starting point cannot be checked.

The information available on a website does not tell you everything about an agency’s ability. It does tell you what a prospective client can establish before making contact. That makes the questions below useful.

Seven questions to ask

Put these to every agency on the shortlist and write the answers down. The useful signal is usually how specific the answer becomes when you move past the sales language.

See how we score them

A useful answer names the platform and the engines it covers.

Seven of the twenty agencies in our study name their measurement instrument publicly. The market uses several different tools, so ask why the agency has chosen its particular one and how it handles changes between AI engines.

A useful answer describes a documented starting position.

That might include the current mention rate across a defined prompt set, which engines are being tracked, what proportion of mentions are citations and how competitors perform against the same prompts.

Ask for the baseline to be recorded before the work begins and dated. No agency in the study publishes one publicly, so this is something worth agreeing with the provider.

A useful answer explains how technical work, structured data and content changes will serve both search and AI visibility.

GEO relies heavily on the same technical and editorial foundations as SEO. The two should therefore make sense together rather than becoming two unrelated workstreams competing for the same pages.

A useful answer names the kinds of publications, directories and industry sites the agency intends to target and explains how it plans to earn coverage.

Independent information about a brand can influence what an AI system knows and cites. Ask what part of the programme will happen outside your own domain.

A useful answer shows examples rather than listing tactics.

Ask to see pages the agency has restructured and understand what changed. That might include clearer answers, factual tables, structured data or sections written so individual passages make sense when extracted from the page.

A useful answer lists the fields before you sign.

Depending on the programme, that could include mention rate, share of voice against a defined competitor set, citations by URL, referral traffic from assistants and the prompt set used for measurement.

Six of the twenty agencies set out their measurement approach publicly. Four state none of it.

A useful answer gives you a dated plan.

You should be able to see the initial audit, baseline, technical work, prompt mapping, first content changes and first report, with an indication of when each is expected.

What the first ninety days should look like

  1. DAYS 1 TO 30

    Audit and baseline

    Technical audit, entity and structured-data review, prompt set agreed and documented, and a recorded starting position across the engines being tracked.

    Competitors should be measured against the same set if comparison is part of the programme.

  2. DAYS 31 TO 60

    The work starts landing

    Technical fixes are shipped, pages are restructured where required and the first outreach begins if third-party coverage is part of the plan.

    The important thing is that there are tangible deliverables rather than another month of audits and recommendations.

The warning signs

A percentage with no starting point is one of the most common problems.

An agency reporting a 300% rise in AI visibility may have moved a brand from one mention to four. Both figures describe the same change, but the second tells you much more about the scale.

Ask for the raw numbers, the date the baseline was taken and the prompt set used. No agency in our study publishes all three publicly.

A method with no named measurement tool is another.

An agency does not need to reveal proprietary processes, but you should be able to understand what is being measured and how. Seven agencies in the study name their measurement instrument publicly, which gives buyers something concrete to compare.

A price that only appears on a call is another common issue. Six of the twenty agencies publish no figure anywhere.

Quote-on-enquiry can make sense for genuinely bespoke work. It also means you have to enter a sales conversation before you know whether the budget is anywhere near the same order of magnitude as your other options.

The final warning sign is a promise of control over the answer itself.

An agency promising a fixed position, guaranteed mention or guaranteed inclusion in an AI answer should be able to explain how that guarantee works. AI answers are generated in response to individual prompts and can change between runs. The sensible focus is on improving the factors the agency can influence and measuring the resulting visibility.

FAQ

FAQs

What should it cost?

The published figures in our study range from around £450 a month at the entry end, through £1,500 to £8,000 for a mid-market retainer, to £8,000 and above for enterprise programmes.

Six of the twenty agencies publish no figure at all, so these should be treated as published reference points rather than a universal market rate.

How long before we see anything?

There is no single timetable. Some technical and content changes can be made quickly, while changes in AI visibility depend on crawling, retrieval and the way individual engines generate answers.

Ask the agency what it expects to change in the first 30, 60 and 90 days, and agree how those changes will be measured.

What should the contract say?

The contract should make the scope and measurement clear. At a minimum, you should understand what work is included, which engines and prompts are being tracked, how often reporting happens, who is responsible for delivery and what happens if the scope changes.

Ask for the baseline and reporting approach to be documented rather than left as a verbal promise.

Can anyone guarantee a result?

An agency can commit to carrying out defined work and reporting against an agreed measurement approach. It cannot control the answer an AI system generates for every future prompt.

Be wary of guarantees around a fixed AI position or guaranteed inclusion unless the agency can explain exactly what is being guaranteed and how it can control it.

What if it does not work?

Ask before signing what happens when the agreed measurements do not improve.

A sensible programme should give you enough reporting to identify whether the problem sits with the prompt set, technical access, content, third-party coverage or the wider measurement approach. That gives both sides something concrete to change rather than relying on a general claim that the work is progressing.

Sarah Rice

About the author

Sarah Rice

Editor

Sarah Rice is a PR specialist & digital marketing expert who has spent more than 30 years working with SEO & GEO agencies, PR agencies and brands across a huge range of sectors. She is Director of Rice PR and Director & Editor of The HR World. Earlier in her career, she was Head of Public Affairs and Press Manager at RAC and Head of Campaigns at JBP, and has also worked with DC Thomas and a number of charities.

Director, Bristol Women in Business CharterMasters, Creative Writing

Get your agency listed

Being included means being read and scored against the same seven criteria as everyone else. It does not mean a position.

See how listing works