Disclosure: CoreAEX provides AEO services, so this guide is written by one of the providers it describes. Its own model is labeled where it appears.

An AEO engagement should leave you with evidence about the current problem, a scoped plan, changes with named owners, and a way to check that those changes reached the pages they were meant to. Which of those you receive depends on what you bought: a diagnosis, an architecture, implementation help or ongoing review. A dashboard, or a list of recommendations for making content "AI-friendly", is not on its own proof that the underlying work was done.

This guide is about deliverables and acceptance: what to ask for in writing, and how to tell whether you received it. "Agency" is used because it is how buyers search; the providers doing this work range from individual consultants to larger teams, and the same questions apply to all of them.

The short answer: what should you receive?

Categories of work to agree in a proposal, not a standard package every provider offers.

StageTangible outputWhat you can check
DiagnoseEvidence, affected URLs and buying questions, stated limits, prioritized findingsCan your team reproduce or inspect each finding?
ArchitectRequirements, owners, an ordered backlog, acceptance criteriaCould a developer or content owner implement it without guessing?
ExecuteThe agreed changes and evidence of their releaseDid the specified change reach the intended pages?
Review (CoreAEX calls this Accelerate)Dated observations and a decision on the next actionWhat changed, and what remains uncertain?

Not every engagement covers all four stages, and providers name them differently. What matters is that each stage you pay for produces something you can inspect.

Before deliverables: define the question and the scope

Outputs can only be checked against an agreed scope. Before work starts, expect to agree:

  • the product and market in question;
  • the affected page templates and representative URLs;
  • a set of representative buying questions, if AI answers are in scope;
  • which platforms will be examined;
  • what access the provider gets to analytics, Search Console or server logs, where relevant;
  • the stakeholders who own content, engineering and publication;
  • what is explicitly excluded.

"Improve AI visibility" is a goal, not an acceptance criterion. It names no questions, platforms, metric or date range, so no piece of work can pass or fail against it.

Hypothetical example: a fictional company used throughout this guide. A mid-market accounts-payable software company finds it is absent from some AI answers when buyers ask for a shortlist of vendors in its category. Separately, its own site states plan information inconsistently: the pricing page, two plan pages and a comparison page describe seat minimums and billing terms differently. The example is illustrative; it has no real traffic, citation or revenue figures, and it does not describe a CoreAEX client.

Diagnostic outputs: findings your team can challenge

A diagnosis is useful when your own team can inspect the evidence and disagree with it. For the hypothetical company, an evidence package might include:

  • the list of URLs and templates tested, with dates;
  • crawler and rendering observations, such as whether key plan details appear in the HTML a given crawler receives;
  • an inventory of conflicting product claims, each linked to the page where it appears;
  • a dated prompt panel, where AI answers are in scope, with the raw responses kept;
  • a trail from each finding to the recommendation it supports.

Each finding should state four things: what was observed, how it was checked, what it does and does not establish, and who owns the next check. "Plan pages disagree on seat minimums" is checkable. "Your brand lacks authority with AI" is not, unless it comes with the evidence and the definition behind it.

Findings should also keep outcomes apart. Being mentioned in an answer, being recommended for the stated need, being cited as a source, and receiving referral traffic are different outcomes, and one can occur without the others. CoreAEX's guide to measuring AI citations sets out those definitions and how vendor tools define them differently.

A prompt panel is a sample. It describes how specific platforms answered specific questions on specific dates under recorded conditions. It cannot describe every question buyers ask, and on its own it cannot show that a later change caused a later result.

Architecture outputs: something the implementing team can use

The architecture stage turns findings into instructions. A usable specification typically names:

  • the affected templates and URLs;
  • the exact product facts and the approved source of truth for each;
  • the proposed page or content changes;
  • technical requirements, where applicable;
  • dependencies and the order of work;
  • an owner for each item;
  • acceptance tests that say how each item will be checked.

The difference between a recommendation and an implementable instruction is usually specificity. "Make pricing clearer" is a recommendation. "On these plan pages, display the unit, billing period, commitment and eligibility condition from the approved pricing source, then check the rendered page against that source" is an instruction someone can build and someone else can verify.

Hypothetical example: a requirement in an illustrative format.

ID
PLAN-03
Pages
/pricing, /plans/team, /plans/business
Requirement
Each plan displays its price unit, billing period, minimum seat commitment and eligibility condition in visible text, taken from the approved pricing source (version agreed in the proposal).
Owner
Content: product marketing. Template change: web engineering. Publication: CMS owner.
Pass condition
On all three live URLs, the rendered page shows all four values for every plan, each matching the approved source; the previous seat-minimum wording is absent; any pricing markup states the same values as the visible text.
Not covered by this test
Whether any search or AI system will repeat the corrected values.

Execution outputs: who changes what?

Execution is where responsibilities need to be most explicit, because several people touch each change. There are three common arrangements:

Illustrative. Whatever the arrangement, production deployment needs a named owner.

ArrangementContent changesEngineering changesCMS publicationFinal verification
Provider advisesBuyerBuyerBuyerBuyer, or agreed separately
Buyer implements, provider reviewsBuyerBuyerBuyerProvider reviews against agreed criteria
Provider performs assigned work alongside the buyerAs assigned in the proposalAs assigned in the proposalNamed in the proposalProvider verifies release against agreed criteria

If the proposal does not say who deploys to production, nobody clearly owns that step, and an approved specification can stall before it reaches the live site.

CoreAEX example. As the CoreAEX services page reads on November 10, 2026: the Diagnostic places implementation outside its scope; Foundation is implemented by the client's team, with two agreed implementation review rounds against acceptance criteria; and Delivery includes specialist tasks explicitly assigned to CoreAEX, alongside the client's team, with the proposal identifying who owns production deployment. Buyers work directly with Zarko Zivkovic. This is one provider's model, not a description of how every AEO provider works.

Verification: what counts as accepted work?

Acceptance checks confirm that a deliverable was produced as specified. They are narrower than outcomes, and a good proposal says so explicitly.

DeliverableExample acceptance checkWhat that check cannot prove
Updated product claimCorrect wording appears on the agreed live URLs; old wording is absentThat an AI system will repeat it
Crawl or rendering fixThe agreed crawler can fetch the intended rendered contentThat the page will rank or be cited
Structured data changeMarkup validates and matches the visible factsThat a rich result or citation will appear
Prompt-panel reportRaw answers, conditions, dates and coding rules are retainedA market-wide visibility rate
Internal handoffThe owner accepts the requirements; open issues are loggedThat implementation has happened

The third column reflects how the platforms describe themselves. Google states that valid structured data does not guarantee a rich result for a given page (Search Console Help). Its AI features documentation says there are no additional requirements, and no special schema, for appearing in AI Overviews or AI Mode, that structured data should match the visible text, and that meeting every requirement still does not guarantee crawling, indexing or serving.

Keep two records separate: work accepted, which the acceptance checks establish, and outcome observed, which later measurement may or may not show. A passed implementation check is not a promise of rankings, citations, recommendations, clicks or pipeline.

What ongoing reporting should contain

If you buy review or follow-up, each report should let you answer "what happened, and what do we do next?" without a meeting. A useful review states:

  • which URLs and questions stayed in scope, and any that changed;
  • what shipped, and when;
  • what was rechecked, and the result;
  • movement in each separately defined metric, reported per platform;
  • unresolved issues and their owners;
  • the decision on the next action.

Before-and-after comparisons have limits. If the prompt set, account state, location or platform behavior changed between rounds, a difference may reflect the measurement rather than the work. AI answers also vary between runs and over time, as CoreAEX's analysis of why AI vendor shortlists change sets out. The shortlist testing protocol covers versioned prompt panels, run counts and what to record on every run.

What should be written into the proposal?

This is the part to copy. A proposal you can hold a provider to should name:

  • The named outputs for each stage you are buying
  • The sources and access the provider will receive, and from whom
  • Sample limits: templates, URLs, buying questions
  • The markets and platforms in scope
  • Who implements each type of change
  • Who owns production deployment
  • The number of review rounds
  • Acceptance criteria for each deliverable
  • Reporting cadence and what each report contains
  • Exclusions
  • Dependencies on your team, and what happens if they slip
  • How change requests are agreed and priced
  • The handoff format: what files, in what form, at the end

What an AEO provider cannot responsibly promise

Each of these runs into a measurement problem rather than a matter of effort:

  • Guaranteed citation or recommendation. The platforms decide what appears in an answer. Google, for example, states that meeting all of its requirements still does not guarantee a page will be crawled, indexed or served.
  • A permanent "AI ranking". Answers vary between runs and over time, so the practical way to describe standing is a rate observed over a defined window, not a fixed position.
  • Complete visibility into retrieval. An outside observer sees answers, and sometimes the sources shown, not the full retrieval and generation process behind them.
  • Attribution of every conversion to an AI answer. Analytics can show some referral traffic; they do not show answers a buyer read without clicking.
  • Results from markup or formatting alone. In one CoreAEX practitioner test on fictional pricing pages, a price that existed only in JSON-LD was not reproduced in any of nine runs across three AI systems on August 30, 2026 (study and limits). That is a narrow test, not a general rule, but it is a reason to treat structural changes as necessary hygiene rather than a guaranteed lever.

Choosing the appropriate level of help

Match what you buy to what you already know:

  • You don't know the cause. Start with a diagnosis, and judge it by whether its findings can be inspected and challenged.
  • You know the issue and have a team to implement. Buy architecture and review: requirements, acceptance criteria and agreed review rounds.
  • You know the issue but lack specialist delivery capacity. Buy scoped execution, with every assigned task and the production owner named in writing.

CoreAEX's services page describes its current options and scope limits. If you would rather talk through which stage fits, share some context.


Sources

The example company and example requirement are hypothetical.

About the author

Zarko Zivkovic is the founder of CoreAEX, building technical SEO, AEO, and AI-visibility systems for B2B SaaS companies. Connect on LinkedIn.