How to choose an AEO agency

Start with the buyer problem, not the agency label

Tell each agency which products, markets, and buyer questions matter. A strong discovery process should ask about your search demand, sales cycle, product facts, technical constraints, and measurement setup. Expect the proposal to identify target pages and whether the work is technical SEO, content, citations, analytics, or a managed combination.

Ask which work would happen in the first month, what you must provide, and what can realistically be observed at each milestone. Google says standard SEO practices remain relevant to AI Overviews and AI Mode and no special AI file or schema is required. An agency that promises a secret AI switch needs to explain its evidence.

Sources for this section: Google: AI features and your website, Google: Do you need an SEO?.

Send every agency the same short brief

Comparisons work only when each provider is solving the same problem. Write a one-page brief with your product category, markets, buyer types, three to ten important pages, and the questions buyers ask before contacting sales. State whether your immediate problem is crawl and indexing, unclear product content, a lack of cited sources, unreliable measurement, or a combination. Include who can approve factual claims and who can deploy code. If you have existing Search Console or GA4 data, describe the signal without giving broad account access at the first conversation.

Ask every agency to return the same items: a first-month priority list, a sample page-level deliverable, the people who will do the work, a measurement plan, dependencies on your team, price, and what is outside scope. Require them to identify assumptions. An agency cannot know its exact impact before it has seen the site, competition, and data. A proposal that admits a missing baseline and explains how it will establish one is more useful than an impressive forecast with no method.

Keep the brief specific enough to expose a mismatch. If your product pages are not indexed, a proposal consisting only of prompt monitoring misses the immediate bottleneck. If pages are indexed but buyers cannot find answers about integrations and plan limits, a generic technical audit may leave the decision problem untouched. Ask the agency to connect every recommended activity to a named page, question, or measurement gap.

Sources for this section: Google: Do you need an SEO?, Google: AI features and your website.

Compare actual deliverables

Request a written scope with named outputs: crawl and indexing findings, page briefs, published copy or code changes, internal-link updates, answer observation logs, analytics configuration, and reporting. Specify review and acceptance points. A monthly dashboard alone is not the same as an implemented program.

For content, ask who verifies product details and whether the agency can show the before-and-after page. For technical work, ask how changes are tested and who deploys them. For outreach or citation work, ask what the agency controls and whether placements are editorial, paid, or uncertain.

Inspect a sample deliverable as if you had to use it tomorrow

For a technical finding, look for an exact URL, the tested response and render, the indexing or canonical evidence, a proposed fix, priority, and a post-deployment check. A screenshot of a tool score alone does not tell a developer what to change. For a content brief, look for the buyer question, existing page, factual gaps, primary sources, product-owner review, proposed structure, and intended internal links. The brief should tell an editor what to publish without asking them to invent a capability or outcome.

For an answer observation report, ask whether the agency saved the question, date, region, language, search mode, answer, linked source, and the number of observations behind a percentage. If it sampled a model API, it should say so; that sample is not the live consumer search answer in ChatGPT or another app. For Google AI-feature performance, ask which Search Console report produced the impressions and whether the report was available for the property. For referrals, ask for the GA4 source or channel definition and which key events are configured.

Also inspect how the agency handles an unfavorable result. Does the sample report call out uncertain attribution and changed prompts? Does it list the page changes actually shipped? Can a reader follow a metric back to its source and date range? A report should support a decision about the next URL or question. Colorful charts without definitions can hide that nothing was implemented or that two unlike metrics were combined.

Sources for this section: Google: Generative AI performance report, Google: Do you need an SEO?.

Demand a reporting method you can audit

Ask for a sample report that separates consumer answer observations, Google AI-feature impressions where available, identifiable referral sessions, and leads. It should state the date range, question set, platform and search mode, and limitations. A single 'AI visibility score' can be useful internally only if its construction and sampling are explicit; it should not hide estimated values inside measured results.

Google's hiring guidance recommends asking how success will be measured and how changes will be communicated. It warns against guarantees of a number-one ranking and against providers that will not explain what they intend to do. Apply the same skepticism to promises of guaranteed AI citations.

Sources for this section: Google: Do you need an SEO?, Google: Generative AI performance report.

Check access, ownership, and the exit path

Start a diagnostic with the least access needed. Google specifically recommends read-only Search Console access for an initial SEO audit. Confirm who owns Search Console and GA4 properties, whether new content and assets remain yours, and what export you receive if you stop working together.

Compare pricing against the actual scope: who writes, who develops, who approves, and what reporting is included. Ask what happens when a page is not indexed or an answer sample worsens. A credible response describes diagnosis and iteration instead of a guaranteed outcome.

1. 01

   Request a sample deliverable

   A page brief, audit finding, or report with sensitive details removed is more informative than a slide of claims.

2. 02

   Define review points

   Agree on the first set of URLs, owners, acceptance criteria, and meeting cadence.

3. 03

   Record exit rights

   Keep access, analytics, code, copy, and observation logs under your control.

Sources for this section: Google: Do you need an SEO?.

Agree on responsibilities and ownership before signing

Write down who provides product facts, who writes and edits copy, who approves legal or compliance language, who implements technical fixes, and who publishes. A content plan can stall for months if the agency assumes it can edit the site while your team assumes it is only supplying recommendations. Give each deliverable an acceptance check and an owner on both sides. Specify whether outreach, analytics setup, development, image creation, and localization are included or quoted separately.

Keep the business's Search Console, GA4, domain, content management system, and code repository under business-owned accounts. Use role-limited access, and confirm that access can be removed. Ask whether drafts, research notes, original graphics, analytics configuration, and observation logs are delivered in portable formats when the contract ends. Review any recurring software fees and what happens to historical reports if you stop paying for a vendor's dashboard.

Compare the contract with the proposed work rhythm. What counts as a completed page, an approved finding, and a monthly report? What is the notice period? How are corrections handled when the product changes? These operational details matter because search work depends on accurate facts and deployed changes. A lower monthly price that leaves implementation and approval outside scope may cost more staff time than a broader but clearly defined engagement.

Sources for this section: Google: Do you need an SEO?.

Use the first month to test the working relationship

A reasonable initial sequence is discovery, baseline, a small set of fixes, and a review. The provider should inspect the priority pages and answer questions you supplied, confirm source access, and explain what can be measured. It can then propose a few changes that your team can approve and publish. Save the original page and technical state, the accepted change, and the verification after release. This is an example workflow rather than a promise that visibility will change in thirty days.

At the first review, ask what shipped, what evidence supports each diagnosis, what remains blocked, and what will be checked next. A useful answer may be that a page became crawlable but has not yet been reprocessed, or that a content change is published while answer observations remain inconclusive. That is still progress you can verify. Reject a report that turns an uncertain observation into a guaranteed win or repeatedly changes the question set to show a favorable percentage.

Finally, assess communication. Can the agency explain a technical issue in terms your developer can act on and a content issue in terms your product owner can verify? Does it share the underlying report rather than only a summary? Will it tell you when a proposed tactic has little evidence? The best fit is the team whose working method you can inspect, approve, and sustain after the contract ends.

Sources for this section: Google: Do you need an SEO?, Google: AI features and your website.