Which questions should I ask an AEO agency before hiring one?

Last updated October 14, 2026. Written by the Third Coast AI team.

Ask an AEO agency fifteen questions in four groups: how they measure (which prompts, which engines, how many runs, who owns the data), how they diagnose (decision factors, cited pages, fact-check errors, crawler access), how they produce (who writes, what the brief is built from, what the publish gate checks, what off-site work is included), and how they contract (term, audit credit, reporting cadence, what happens to the work if you leave). The answers below are the ones we would expect from a competent vendor, and the ones that should worry you are noted where they matter.

You do not need to be an expert to run this list, and the point is less to catch anyone out than to find the agency whose answers you understand. Bring it to the second call, after you have seen a proposal, and give the vendor time to answer properly; a good one will welcome the questions because they are the same ones they ask themselves. The guide to choosing an agency explains the reasoning behind each group.

Which questions cover measurement?

1. Can we see the prompt set before you run it, and do we own it afterward? Expect yes to both, with a sample of twenty prompts on the call. The prompts should read like things your buyers say, grouped by persona and by type (visibility, decision factor, competitor, fact check). A vendor who keeps the prompt set private is keeping the measurement private.
2. Which engines do you run, with retrieval on or off, and how many times per prompt? Expect ChatGPT with search on, Gemini with grounding, Google AI Overviews and AI Mode, Perplexity and Copilot, with each prompt run several times. Only about 11% of cited domains overlap between ChatGPT and Perplexity, so one engine is not a proxy for the rest, and SparkToro and Gumshoe.ai found less than a 1 in 100 chance of the same brand list repeating on a single prompt, so one run is not a measurement.
3. What does the number in your report mean? Expect "the share of runs in which you were named, by engine, over the prompt set," with citation rate reported separately. If the answer is a proprietary score, ask what goes into it and whether you could recompute it from the raw answers.
4. Do we get the raw answers and citations? Expect yes, as an export, every month. The raw answers are the evidence, and a report without them cannot be checked.

How do you test the diagnosis?

1. Which decision factors do the engines use for our category, in their own words? Expect a list with frequency counts, pulled from prompts like "what should I look for in a [category]," and expect to recognize most of it from your own sales calls. A vendor who has not mined the factors will answer with generic ranking advice.
2. Which pages do the engines currently cite for our three most important questions, and have you read them? Expect specific URLs and an account of what they share: where the answer sits, how many statistics appear early, whether there is a table, how the headings are phrased. Ahrefs measured only 38% of AI Overview citations coming from Google's top ten in March 2026, so a vendor working from your rankings has not looked at the right list.
3. What do the engines currently get wrong about us? Expect a fact-check log from the proposal stage or at least a commitment to produce one in week one, because wrong facts are the fastest fixes and the most embarrassing misses. The correction guide explains what to do with the log.
4. Can the AI crawlers reach our pages? Expect the vendor to have checked robots.txt and the CDN for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Bingbot, and to have fetched a page as each one. Cyrus Shepard's May 2026 synthesis of 54 studies scores crawler accessibility 9.5 out of 10, the highest of 23 factors, so this is the first thing a competent vendor checks and the one most often skipped.

Does the production process hold up?

1. Who writes the pages, and who on our side reviews them? Expect a named writer or editor, a brief built from the cited pages and the decision factors, and a reviewer on your side who knows the business. Expect the vendor to insist on the reviewer, because a page that states a factor without proof does not get cited.
2. What does your publish gate check? Expect something like: answer in the first 150 to 200 words, every decision factor stated and proven, sourced statistics, an attributed quote, a list and a table, headings phrased as buyers ask, text in the raw HTML, nothing invented. Aggarwal and colleagues at Princeton measured quotations, statistics and cited sources lifting visibility in generated answers by about 41%, 33% and 28%, so the gate is not decoration.
3. How many pages a month, and how is that number set? Expect a number derived from the gaps found in the audit rather than a package tier, and expect the vendor to push back on volume. Glenn Gabe documented a site that lost visibility after publishing 850,000 AI-generated pages, which is the case against buying pages by the dozen.
4. What off-site work is included, and is any of it paid placement? Expect profile reconciliation, Wikidata and schema sameAs work, and outreach to the specific sources the engines cite for your category, with no paid placements described as coverage. Ahrefs' study of 75,000 brands found third-party mentions correlate with AI citations at 0.664 against 0.218 for backlinks, so this work should be in the proposal, not an upsell.

Can the contract be judged on three questions?

1. What is the term, and is the audit fee credited against a program? Expect a fixed-fee audit credited against the first month, a first quarter commitment and month to month after that. A twelve-month lock-in before any audit is a sign the vendor does not expect the audit to sell the program on its own.
2. How often do we get a report, and what is in it? Expect monthly, with prompt-level visibility by engine over repeated runs, citation rate, the fact-check log, pages shipped against the plan, and crawler status. The KPI and reporting guide lists what belongs in it.
3. If we stop, what do we keep? Expect the prompt set, the raw answers, the briefs, the published pages and the entity records, all in formats you can use. Anything the vendor keeps is a dependency you are paying for.

Should any single answer end the conversation?

A few should, though we would give the vendor a chance to clarify first. A guaranteed rank ends it, because the engines do not produce stable ranks. A report built on one run per prompt ends it, because the trend line will be noise. "Add FAQ schema to every page" as the lead recommendation ends it, because Ahrefs' difference-in-differences test on 1,885 pages found FAQ and HowTo schema made no measurable difference on AI Mode or ChatGPT. And refusing to share the raw answers ends it, because without them you are trusting a score you cannot check.

If a vendor answers most of these well, the remaining choice is about fit, which is best judged by whether you understood their answers without help. We would rather be compared on this list than on a pitch, and if you want to see how we answer it, the audit page describes the first engagement and the pricing guide covers how we charge.

Frequently asked questions