What should I ask an AEO agency before hiring them?

Start with measurement: how they define and count a citation, what the fixed prompt set is and whether you can see it before anything runs, which engines they check separately, and how they handle the fact that answers change between runs. Then ask whether the methodology is published, what they do about listings and profiles, and what they will not promise.

How can I tell a real AEO practice from a renamed SEO package?

Ask what gets reported. A rebranded package reports rankings and sessions with an AI section added; an actual engagement reports answer share and citation rate per engine against a written prompt set, and can name which specific source earned each citation. Asking what they do about your listings and profiles is the fastest second test, since Yext found 42% of AI citations came from those.

Should I trust AEO agency case studies?

Read them as demonstrations that a functioning programme produces measurable movement, not as forecasts for your business. They are self-reported by the agencies that ran the work rather than independently audited. Three questions make any of them checkable: over what period, against what baseline, and measured how.

Is it a red flag if an agency guarantees AI placement?

Yes. Nobody controls how a model selects its sources at the moment it generates a response, and retrieval behaviour changes without notice. A guarantee describes something the seller does not control, which makes the question of what a vendor will not promise the most informative one you can ask.

What to ask an AEO agency before you hire one

Why the usual due diligence does not work here

Answer engine optimisation is roughly two years old as a named service. Nobody has a ten-year track record, most proposals describe the same tactics in the same language, and the category is new enough that a convincing deck is cheap to produce. The normal shortcut — compare portfolios, pick the biggest — selects for marketing budget rather than for competence.

What does discriminate is method. These are the questions worth asking, and what a weak answer to each one sounds like.

On measurement

1. How do you define a citation, and how do you count it?

There is a real difference between a company being named in an answer and one of its sources being the attached citation. The second is considerably more valuable and considerably harder. An agency that has not drawn that distinction has not been measuring carefully.

Weak answer: "We track AI visibility", with no definition offered.

2. What is the prompt set, and can I see it before you run anything?

Measurement only means something against a fixed set of questions, agreed in advance and re-run unchanged. A set that can be edited between runs produces numbers that cannot be compared, and a prompt selected after seeing the result is cherry-picking.

Weak answer: prompts assembled after the first report, or never shown to you.

3. Which engines do you check, and do you report them separately?

ChatGPT, Perplexity, Google AI Overviews, Gemini and Claude retrieve and cite differently. A single blended score hides the case you most need to see: present in two, absent in three.

Weak answer: ChatGPT only, or one averaged number.

4. How do you handle the fact that answers change run to run?

Answer engines are non-deterministic. Any honest measurement treats a single run as a sample and the trend across repeated runs as the evidence.

Weak answer: a screenshot presented as proof.

On method

5. Is your methodology published where I can read it before a call?

A named framework with no page behind it is a logo. The point of publishing a method is that it can be interrogated before money changes hands. The same applies to price, which is why I publish what an engagement costs rather than quoting it only once a call is booked.

6. What do you do about my listings and profiles?

Yext's October 2025 analysis of 6.8 million citations found 42% came from listings and profiles against 44% from first-party sites. An agency working only on your website is working on roughly half the cited surface.

7. How will you check whether my content is even visible to the crawlers?

Vercel and MERJ reported in December 2024 that no major AI crawler executes JavaScript — GPTBot requested JavaScript files 11.5% of the time and ClaudeBot 23.8%, neither ran them. If your site renders client-side, this is the first and cheapest thing to establish, and an agency that skips it may spend months optimising content nothing can fetch.

On honesty

8. What is your position on SEO?

Answer engines retrieve heavily from indexed web content, so a page classic crawlers cannot read is a page they cannot cite. "SEO is dead" is not a bold position, it is a factual error, and buyer guides list this question as a screening criterion precisely because the answer sorts vendors so quickly.

9. What can you not promise?

Nobody controls how a model selects sources at generation time. A vendor guaranteeing placement in an AI answer is either misunderstanding the technology or counting on you to. This question is the most useful of the nine, because it is the one a weak vendor is least prepared for.

On results, and how to read them

Published case studies in this category are real and worth asking for. Optimist reports 49 times growth in LLM referral revenue over 14 months for a B2B technology client; Discovered Labs reports AI-referred trials rising from 575 to 3,500 in seven weeks for a B2B SaaS client. Figures like these are self-reported by the agencies that ran the work, which is the normal standard for agency case studies and is not the same as independent verification.

So read them for what they demonstrate — that a functioning programme produces measurable movement — rather than as a forecast for your business. And ask the three questions that make any case study checkable: over what period, against what baseline, and measured how. A number without those attached is a number about nothing in particular.