Hiring an AEO agency? Seven questions to ask before you sign

No AEO agency can sell you a guaranteed citation. What you can buy is work that helps answer engines find, understand, and cite your pages, and a record of what changed. Read the proposal, the contract, and every monthly report with that distinction in mind.

AEO stands for answer engine optimization. Answer engines such as ChatGPT search, Perplexity, Google’s AI Overviews, and Microsoft Copilot write their own answers and can link to the sources behind them. An agency can improve your pages. It cannot decide which sources an engine selects. OpenAI’s search guidance puts it in four words: “Placement is not guaranteed.”

Before you sign, get seven answers in writing.

What does an AEO agency actually do?

The work has three parts. Access: checking that each engine’s crawler, the program it sends to read web pages, can reach your content, and that your pages can be indexed. Content: putting a clear answer, with its source, where a reader and a machine can both find it. Measurement: recording, with dates, what appeared, where, and for which question.

None of this is secret. Google, OpenAI, Perplexity, Anthropic, and Microsoft publish guidance on their crawlers, features, and reports. An agency earns its fee by applying that guidance to your site, doing the work, and showing you what it did. Because the guidance is public, you can check an agency’s advice against the source.

Google’s hiring advice already names AEO

Google’s “Do you need an SEO?” page, last updated in June 2026, names AEO and GEO services directly. It tells you to ask whether a provider’s advice on optimizing for AI experiences matches Google Search’s official guidance on generative AI features. It is blunt about guarantees: “No one can guarantee a #1 ranking on Google.” And it puts the responsibility on you: you answer for the actions of any company you hire.

The seven questions below apply the same discipline to the other engines. Each should produce something you can inspect before you grant access or approve a page.

The seven questions

Question	A good answer	A red flag
1. Which questions will you measure, on which engines, and can I see the baseline first?	An agreed list of buyer questions, the engines named, and dated answers saved before anything changes.	A proprietary visibility score with no answers underneath.
2. What do you check before you touch the writing?	Crawler access, then indexing and snippet eligibility, then the page itself.	An llms.txt file as the first fix, before anyone has checked access.
3. What will you change, and will I see the reason?	A change log: what, where, why, and the evidence behind each recommendation.	A request for admin access without a written scope.
4. What will you not promise?	No guaranteed citations, rankings, or traffic, and a plain account of what the agency does control.	Guaranteed placement in ChatGPT by a set date.
5. Which advice comes from official documentation, and which from your own tests?	Linked guidance, with the agency’s own tests and data marked as such.	Claimed inside knowledge of how models choose sources.
6. How will you report mentions, citations, and clicks?	Three separate measures, with coverage and gaps stated.	One AI share-of-voice number with no breakdown.
7. Who writes the content, and who checks the facts?	A named editor, a source for every factual claim, and your approval before publication.	Page volume offered in place of an editorial process.

1. Measurement and baseline

A baseline is a dated record of how you appear before the work starts. Ask for the record itself, not a summary: the prompt set with the exact wording of each question, the engines, the dates, and the saved answers. Agree how the checks will be repeated, so every later report compares like with like. A single answer is an observation, not a trend. Even a clean before-and-after comparison does not, on its own, show that the agency caused the change.

2. Access before writing

Check each engine separately. OpenAI names OAI-SearchBot as its search crawler. Perplexity uses PerplexityBot, and Anthropic uses Claude-SearchBot for search. Google’s AI features guidance says a page must be indexed and eligible to show with a snippet before it can appear as a supporting link in AI Overviews or AI Mode.

Then ask how the agency finds and fixes a blocked crawler. For ChatGPT search, OpenAI says to allow OAI-SearchBot and to let through traffic from its published searchbot IP addresses. Its crawler documentation says its systems can take about 24 hours to adjust to a robots.txt change. That is a processing time, not a citation deadline. GPTBot vs OAI-SearchBot explains which OpenAI bot does what.

3. Changes and access rights

One of Google’s hiring questions asks whether a provider will share every change it makes to your site and explain the reasoning behind its recommendations. If the provider offers an audit, Google advises granting only read access to Search Console, Google’s dashboard for your site’s search performance, and no write access at that stage. Agree the scope in writing before you widen access, and keep a record of every change you approve.

4. What they will not promise

OpenAI does not guarantee placement. Google says indexing and serving are not guaranteed for its AI features, and its hiring page tells you to find someone else if a provider guarantees first place in search results. No agency can guarantee what the engines themselves do not. Ask the agency to commit to what it controls: the work, the dates, and the reporting. A guaranteed citation is a reason to reject the proposal.

5. Documentation versus their own tests

Both can inform a recommendation, but they support different claims. Official documentation tells you what the platform says. An agency’s test tells you what happened under its test conditions, so ask for those conditions and the data before you accept a broader claim. If a recommendation has neither a link nor test data behind it, treat it as opinion. Google’s AI guide also says that no third-party tool has access to its internal ranking or AI systems.

6. Three numbers, not one

A mention is your name in an answer. A citation is a link to your page as a source. A click is someone following that link. Keep them in separate columns. ChatGPT mentions vs citations explains why a name is not a source.

Microsoft’s Bing AI Performance announcement describes a public preview that reports citations across supported experiences, including Copilot and Bing’s AI summaries. Microsoft adds that its cited-pages figure does not indicate ranking, authority, or a page’s role in an individual answer. OpenAI’s publisher guidance describes the utm_source=chatgpt.com tag in referral URLs. Google’s July 2026 guide points to a Generative AI performance report in Search Console.

Each report covers part of the picture, so ask what it counts and what it leaves out. Mentions are the gap: record them with the agreed prompt set, because referral visits and citation counts do not show how often your name comes up.

7. Content and facts

Ask who checks the sources, who edits the draft, and who approves publication. Google’s content guidance says that using automation, including AI generation, to produce content primarily to manipulate search rankings violates its spam policies. In May 2026, Google clarified that its spam policies also apply to generative AI responses. A proposal built on page volume still has to explain how each page helps a reader.

Which red flags end the conversation?

• Guaranteed citations or rankings. No agency can make an engine select your page.
• Claimed inside access. Google says no third-party tool has access to its internal ranking or AI systems.
• llms.txt sold as a ranking switch. Google says Google Search ignores these files, and that creating one will neither help nor harm your visibility there. See what llms.txt is.
• FAQ markup sold for Google rich results. Google’s changelog says FAQ rich results no longer appear in Google Search as of 7 May 2026. Its structured data rules also prohibit marking up content readers cannot see.
• Inauthentic mentions or fake reviews. Google’s AI guide says chasing inauthentic mentions across the web is not as helpful as it might seem. The same structured data rules prohibit marking up fake reviews.
• Write access on day one. For an audit, Google’s advice is read access to Search Console. Wider permissions should follow a written task and scope.

A hypothetical: two proposals for one firm

Imagine a 12-person accounting firm that gets two proposals.

Proposal A promises “top-three visibility in ChatGPT within 60 days,” an llms.txt file, and 200 AI-written FAQ pages. It asks for admin access to the website.

Proposal B starts with a baseline on 25 questions the firm’s clients actually ask, across three engines, plus a crawler access check. In month two it rewrites five existing pages so the answer comes first, with a source for every figure. It includes a change log and a monthly report with mentions, citations, and clicks in separate columns. It says, in writing, that citations cannot be guaranteed.

Proposal A raises four red flags before anyone asks a question: the placement guarantee, an llms.txt file with no access check, admin access with no written scope, and page volume with no editorial process. The guarantee alone is reason to decline. Proposal B may still come from the wrong agency, and it promises no citations. But everything in it is work you can inspect.

What deliverables should the contract name?

1. The prompt set and the dated baseline, with the raw answers saved.
2. An access report per crawler: robots.txt rules, firewall settings, response codes.
3. A change log: every change, the reason, and the source for the reason.
4. A monthly report with mentions, citations, and clicks in separate columns, plus what could not be measured.
5. For any content: a source for each factual claim, and your approval before publication.

To check one page yourself first, the citation-readiness audit is a fifteen-minute pass or fail on one URL.

Try this today (15 minutes)

1. Open yoursite.com/robots.txt. Note any rules for OAI-SearchBot, PerplexityBot, Claude-SearchBot, and Googlebot, and read the general User-agent: * group too. Flag anything that looks like a block. A clean file does not prove access: a firewall or your host can still turn a crawler away.
2. Put five of your customers’ real questions to ChatGPT with search on, and to Perplexity. Save the exact prompt, the answer, the engine, and the date.
3. For each answer, note whether you were named, linked, or absent. Keep names and links in separate columns. This exercise does not measure clicks.

That page of notes is a first baseline, not a trend. Show it to any agency you are considering and ask what it would check next, what it would change, and how you would see the result. An answer about work and a record is something you can hold an agency to. A promise of placement is not.

Cite this:Hiring an AEO agency? Seven questions to ask before you sign.Len P. van der Hof. https://lenvanderhof.com/en/blog/hiring-an-aeo-agency/ · Published 15 October 2026.