How to choose an AEO/GEO agency: a 2026 buyer's guide

Half the agencies now selling answer engine optimisation were selling links and content last year. This guide gives you the vetting questions, the red flags and the single request that separates a supplier who can actually measure and move your AI citations from one who has bolted a fashionable label onto an old SEO retainer.

Why AEO buying is different from buying SEO

Answer engine optimisation looks, from a distance, like search engine optimisation with a new coat of paint. The pitch decks rhyme, the case studies use the same charts, and half the agencies now selling it were selling links and content last year. But the underlying problem has changed shape, and that changes what a good supplier has to be able to do.

Classic SEO optimised for a ranked list of ten blue links. AEO optimises for whether a language model names your brand and cites your page inside a synthesised answer, across ChatGPT, Gemini, Perplexity, Google AI Overviews and Copilot. Those are five different systems with five different retrieval stacks, and they do not agree with each other. Ahrefs, studying 15,000 long-tail queries and published in August 2025, found that on average only about 12% of the links cited by ChatGPT, Gemini and Copilot appear in Google's own top ten for the same query. The old assumption - rank first, get chosen - no longer holds.

So the buyer's job is not to find the cheapest content shop. It is to find a supplier who can measure a moving, fragmented target honestly, then move it. Most cannot. This guide gives you the questions and the red flags to tell the two apart before you sign.

Curious how AI engines describe your brand right now? Get a free visibility audit and see where you stand across ChatGPT, Gemini and Perplexity.

The core problem: you cannot manage what nobody can measure

Every serious AEO decision comes back to measurement. If an agency cannot show you, live, which prompts trigger your brand in which engines and which sources those engines cited, then everything downstream - the content plan, the schema work, the "authority building" - is guesswork dressed as strategy. You will pay for activity and hope for outcomes.

This matters because the ground keeps shifting. Ahrefs found that citations to Google AI Overviews from top-ten organic pages fell from around 76% in mid-2025 to roughly 38% by early 2026. An engine can quietly re-weight its sources between one month and the next, and your visibility can move without anyone touching your site. Only continuous, prompt-level measurement catches that. A quarterly PDF does not.

"If they cannot show you a live citation report, they are selling you hope with a dashboard on top."

So the single most useful thing you can do in a sales call is ask to see the measurement, running, against a real brand. What follows is how to pressure-test it.

The vetting questions that actually separate the field

Generalists answer these vaguely. A measurement-first supplier answers them crisply, with definitions and screenshots. Ask all of them.

• How do you define a citation? There is a real difference between your brand being named in the answer text, your URL appearing in the reference list, and your domain merely being retrieved in the background. A good supplier draws these distinctions unprompted and tells you which they count and why. A weak one treats "mention" and "citation" as the same word.
• Do you test real prompts across engines, and how many? The answer should be a set of actual buyer prompts run repeatedly across ChatGPT, Gemini, Perplexity, AI Overviews and Copilot, not a single query typed once. Ask how they build the prompt set, how often it runs, and how they handle the fact that the same prompt returns different answers on different days.
• How do you attribute movement? When your citation share goes up, can they tell you why - a new page, a schema change, a third-party mention, or simply the engine re-weighting its index? If every improvement is credited to their work and every decline blamed on the model, the attribution is theatre.
• How do you handle engine disagreement? Because the engines diverge so sharply, "we improved your AI visibility" is meaningless unless it is broken down by engine. Ask for per-engine reporting as a baseline, not an upsell.
• What is your baseline and your control? Ask how they establish where you started and how they separate their impact from the general rise of AI search. Suppliers who cannot describe a before-and-after with a comparison group are claiming credit they cannot prove.

Ask for a live client citation report

This is the one request that ends most sales conversations quickly, which is exactly why it works. Ask the agency to show you, on screen, a current citation report for an existing client: the prompts they track, the engines they cover, which answers name the client, and how that has changed over the last few months.

You are not asking for a polished case study. You are asking to see the working tool. A supplier who genuinely measures citations has this open in a tab already and can redact a client name in seconds. A supplier who has bolted "GEO" onto an SEO site will offer a testimonial, a traffic graph, or a promise to "set that up during onboarding." Those are all substitutes for the thing you asked to see.

Watch for three specific dodges. First, showing you Google Analytics or organic rankings instead of AI citations, as if referral traffic from ChatGPT were the same as being named in its answers. Second, showing a one-off screenshot of a single lucky answer rather than a tracked trend. Third, showing their own marketing visibility rather than a client's, which proves they can rank a blog post but not that they can move a customer's numbers.

The red flags that should end the conversation

Some warning signs are structural. They tell you the supplier does not really do this, whatever the deck says.

The clearest is the recent rebrand with no new capability underneath. When almost every SEO shop has quietly added a "GEO" or "AEO" tab to its services page in the last year, the label means nothing on its own. Look for whether the measurement, the process and the reporting are genuinely built for AI answers, or whether it is the same content-and-links retainer with new headings. Ask what they stopped doing, not just what they added.

Other red flags cluster around false certainty and vanity metrics:

• Guaranteed rankings or guaranteed citations. No one controls how a model synthesises an answer. A guarantee is either ignorance or a lie, and both are disqualifying.
• One number for "AI visibility." Given how little the engines overlap, a single blended score hides more than it reveals. Insist on per-engine, per-prompt detail.
• Traffic as the only proof. AI answers frequently name a brand without sending a click. If the whole case rests on referral sessions, they are measuring the wrong thing because it is the easy thing.
• Secret methodology. If they will not explain how they detect and count citations, you cannot audit their claims, and you will never know whether the line went up because of them or in spite of them.
• Volume-based deliverables. Twenty articles a month is an input. Being named in the answer to the prompt your buyers actually type is the outcome. Retainers priced in content units optimise the input and ignore the outcome.

Why measurement-first suppliers pass and generalists fail

Put the questions and the red flags together and a pattern emerges. The work that moves AI citations - fixing how your pages are structured, earning the third-party mentions engines trust, testing which phrasings get you named - only compounds if you can see its effect per engine, per prompt, over time. Without that feedback loop you are shipping content into a void and reading the tea leaves of your traffic.

This is also why the buyer's guide is not really neutral about method, even while it stays neutral about vendors. A supplier who leads with measurement can survive every question above, because the questions are just their working process turned into an interview. A generalist who bolted a label on cannot, because the questions probe for a capability they never built.

The context makes the stakes concrete. Gartner predicted in February 2024 that traditional search volume would fall 25% by 2026 as buyers shift to AI assistants; the reality has been messier, with Google adapting through AI Overviews, but the direction of travel is not in doubt. ChatGPT alone reached roughly 800 million weekly users by September 2025. A growing share of the moment where a buyer forms an opinion about your category now happens inside a generated answer you cannot see unless you measure it.

A simple scorecard before you sign

Reduce the decision to a short checklist and score every shortlisted supplier against it. Can they define a citation precisely? Do they run a real prompt set across all five major engines on a repeating schedule? Will they show you a live client citation report on the call? Can they attribute a specific movement to a specific cause, per engine? Do they avoid guarantees, single-number dashboards and traffic-only proof?

A supplier who clears that bar is doing measurement-first AEO whether or not they use the phrase. One who stumbles on half of it is selling you an SEO retainer with a fashionable label, and you will spend a year unable to answer the only question your board will ask: are we being named in the answers our customers see, and is it getting better?

Buy the measurement first. Everything else - the content, the schema, the authority work - is only worth paying for once you can prove it moved the number that matters.