How to choose an AEO agency in the UK

This guide is written by an agency that does this work, so read it accordingly. It is structured to be useful even if you go elsewhere, and it includes the strongest arguments against paying anyone for AI visibility at all, because a provider who cannot answer those should not be selling it.

What it actually costs in the UK

Agency marketing content quotes £2,500 to £15,000 a month. The verifiable published prices cluster a long way below that. These are observed, published UK figures rather than aspirational ones.

One datapoint is worth carrying into any negotiation: it is already possible to buy AEO-inclusive SEO for less than some providers charge for AEO alone. The bundle can be better value than the specialist.

The case against the whole category

Presented at full strength, because you should hear it before you spend.

Google says it is just SEO

Forrester's Nikhil Lai has described AEO as significantly but not fundamentally different from SEO, and said proponents tend to exaggerate the differences to carve a startup-sized hole in marketers' tech stacks. That is an analyst at a major research firm calling the category a manufactured budget line.

Strip the jargon and it is a familiar checklist

Be crawlable. Render server-side. Publish clear factual content with headings. Keep it fresh. Earn third-party mentions. Keep your listings accurate. That is technical SEO plus digital PR. The genuinely novel items — llms.txt, vector embedding optimisation, speakable schema — are precisely the ones with no evidence behind them.

The flagship tactic failed its only rigorous test

Schema markup is the most-recommended intervention in the category. The one large-scale controlled test found −4.6% on Google AI Overviews and statistical noise elsewhere. If the flagship tactic fails its only proper test, the prior on the untested tactics should be low rather than high.

The base rate is grim and rarely disclosed

Attribution is structurally impossible

Between model drift, retrieval changes, prompt-set sensitivity, personalisation and plain non-determinism, no provider can demonstrate that their intervention caused a particular mention. The most credible practitioner case study located in this research moved a brand from 1 of 12 to 4 of 12 prompts over 14 weeks, and could trace only 2 of those 4 new appearances to its own actions. That was an unusually honest write-up.

Where the case against fails

Three rebuttals survive, and they are the honest core of why this work is worth anything.

1. The gates are real, binary and verifiable. Bing indexation, OAI-SearchBot access and server-side rendering materially determine whether a business can be cited at all. A business can be perfectly healthy in Google and structurally invisible to ChatGPT. That is not repackaged SEO.
2. The signal mix genuinely differs. Unlinked brand mentions out-correlate backlinks roughly three to one. Reddit and YouTube carry weight far beyond their classic SEO role. Only about 38% of AI Overview citations now come from Google's top ten, down from 76%. Optimising purely for classic rankings demonstrably does not get you cited.
3. Aggregate measurement does work. The loudest and most credible skeptic on AI rankings explicitly endorses visibility measured across many prompts and many runs as surprisingly useful. The skeptic case kills the ranking product. It does not kill the visibility product.

Questions that separate work from vocabulary

1. Which retrieval pipeline does this intervention act on? If the answer covers all assistants equally, it probably acts on none of them.
2. Show me the prompt set. Whoever chooses the prompts chooses the result, and a provider reporting improvement against prompts they picked themselves has a conflict of interest. Ask whether the set changed between reports.
3. What sample size, and what is the confidence interval? A point estimate with no variance disclosed is decoration.
4. What is your evidence grade for each thing in this proposal? Anything resting on llms.txt, vector embeddings or speakable schema should be free.
5. What is the base rate of success for a business like mine, and what does failure look like in month six?
6. Are you willing to report no detectable change? A provider who has never returned that result has either never looked or is not telling you.

Red flags

• Any guarantee of a rank, position or placement in an AI assistant. The thing being promised does not exist as a stable object.
• A score out of 100 with no explanation of what it is composed of or what a given value is worth.
• Share of voice quoted to one decimal place off a 15 or 25 prompt panel.
• A single screenshot presented as evidence.
• Claims about getting you into training data, which is unfalsifiable by construction.
• Anything involving serving different content to AI crawlers than to human visitors. That is cloaking.

What a defensible engagement looks like

1. A retrieval audit first, because those findings are binary and verifiable, and they are the part any provider can stand behind unconditionally.
2. A baseline measured on a prompt set agreed with you, run many times across engines, reported as a visibility share with the sample size and the raw answers included.
3. Interventions drawn only from the evidence-backed rows, each named with its grade in the proposal.
4. Re-measurement on an identical protocol, with model drift named as a confound and a willingness to report that nothing moved.

That is the standard we hold ourselves to, and it is the standard worth holding anyone to. If a provider will not commit to reporting a null result, the reporting is not measurement.