The AEO Action Loop

How one buyer question becomes a change you can prove

Five stages, each one producing the input to the next. Every stage below states what goes in, what comes out, and the part that matters most: what it cannot tell you. Supported changes go through GitHub or Shopify review; everything else stays a clear manual action.

Measure the answer. Make the change. Prove what moved.

Measure → Understand → Decide → Ship → Prove

What AEO is, and what it can and cannot measure

The AEO Action Loop and the evidence it carries1Measure2Understand3Decide4Ship5ProveBuyerquestionObservedanswerSourceadvantageApprovedchangeComparableresultA flat or declined result is the next thing to measure
The AEO Action Loop and the evidence it carries
Five stages run left to right: Measure, Understand, Decide, Ship, Prove. Beneath each one is the evidence it produces and hands to the next: a buyer question becomes an observed answer, which reveals a source advantage, which selects an approved change, which produces a comparable result. An arrow returns from Prove to Measure, because a result is the next thing to measure rather than the end of the work.
  1. 1

    Measure reality

    Measure

    Read the site the way a crawler does, then ask a real buyer question and record what came back.

    What goes in

    • One public URL, fetched without executing JavaScript
    • Exact-URL robots.txt policy for each named search and AI bot
    • One unbranded buyer question, generated from what the page is about

    What comes out

    • A readiness score across six weighted areas
    • One live answer, printed verbatim, with provider, API surface and timestamp

    What this does not prove: The scan reads the HTML we receive. It is not a claim about what any provider's own crawler received.

    The question you were measured on is the answer the next stage reads.

  2. 2

    See why others win

    Understand

    Read the exact answer, who it named, which role each brand held, and the pages it leaned on.

    What goes in

    • The verbatim answer
    • Every source URL the answer cited

    What comes out

    • Whether you were mentioned, recommended, cited, or absent, as four different results
    • The other brands in the answer, the role each held, and the pages behind them
    • What the winning source page did that yours did not

    What this does not prove: Brands inferred from one answer are suggestions. They are provisional until you confirm them, and one answer is not a market map.

    The competitor and source that won the answer become the gap the next move has to close.

  3. 3

    Choose the next move

    Decide

    One ranked opportunity, tied to the gap that was actually observed rather than a generic checklist.

    What goes in

    • The observed gap
    • The evidence that identified it
    • Your readiness findings

    What comes out

    • One ranked next action, with the evidence it was drawn from attached

    What this does not prove: Ranking is a judgement about likely impact, not a prediction. Nothing here forecasts what an answer will say next.

    The chosen opportunity carries its evidence into the change you ship.

  4. 4

    Make the change

    Ship

    Paste-ready guidance, or a reviewable change you approve before anything is published.

    What goes in

    • The chosen action
    • Your connected repository or store, when you connect one

    What comes out

    • Paste-ready guidance, or a reviewable change you approve before anything is published
    • Confirmation that the published change is publicly visible

    What this does not prove: Review and approval are mandatory. Nothing is published on your behalf without it, and publication proves deployment, not outcome.

    The published change and its date become the baseline the next check is compared against.

  5. 5

    Prove what moved

    Prove

    Ask the same question again after the change is live, and record whether the answer improved, held flat, or declined.

    What goes in

    • The published change and its date
    • The same question, provider and API surface

    What comes out

    • A comparable later answer, and an improved / flat / declined status
    • History, trends, and alerts when a result moves

    What this does not prove: A comparable check is evidence, not attribution. Answers vary by time, surface, location and personalisation, and a flat result is a legitimate outcome.

    A flat or declined result is not a failure of the loop. It is the next thing to measure.

Which answer engines: Monitor checks the buyer questions you track against supported providers using their APIs. That is separate from Google AI Overviews, from a consumer chat session, and from Search Console data supplied by a connected Google property. A provider we did not check is shown as Not checked, never as a negative result.

The whole loop, on one question

The same chain as above, run once end to end, including the stage where the site was simply not in the answer.

One question, followed to a result

Illustrative example
  1. The question

    What are the best workflow analytics tools for remote teams?

    Perplexity · Sonar API · 12 June 2026

  2. Before

    Northloop: not named in this answer. The site was readable and indexed, so readiness was not the problem.

    A readiness score is not evidence that an assistant recommends you.

  3. Who won it, and why

    RivalTrack was recommended and cited.

    The cited page names remote-team workflow benchmarks, setup time, and the reporting method in the page body. Northloop's equivalent page describes analytics generally, so the answer had no remote-team evidence to quote back.

  4. The change, and the evidence behind it

    Publish the remote-team workflow analytics use case

    The question is about remote teams and the winning source answers it with concrete benchmarks. State Northloop's setup time, workflow metrics, and remote-team use case in the page body. Target page: /remote

  5. Approved and published

    src/app/remote/page.tsx: one line replaced, reviewed before merge.

    Published 19 June 2026, then verified as publicly visible.

  6. The comparable check afterwards

    Absent · 12 JuneCited · 3 July

    Same question, same provider, same API surface, 21 days later. A comparable check, not a guarantee that the change caused it.

Example data. The question, answer, roles, sources, change and result are the objects a real check produces; the companies are fictional and use the reserved .example domain.

What a measured answer does and does not prove

  • A readiness score is not evidence that an assistant recommends you.
  • A provider we did not check is not a negative result.
  • One answer at one moment is a sample, not a trend.
  • Mentioned, recommended, and cited are different outcomes.
  • AI answers vary by provider, API surface, time, location, and personalisation.
  • Published proves the change is live. Only a comparable later check can say whether it moved the answer.

Research, measurement and corrections methodology

Inside the Measure stage

What the readiness score actually scores

Six areas, weighted by likely impact, scored from 0 to 100. It tells you whether a crawler can read and understand the page. It is not evidence that an assistant recommends you. That is what the live answer above is for.

20%

Crawl & indexability

Does the URL return a usable response, and does its policy permit discovery? We check status codes, exact-URL robots.txt rules, sitemaps, canonicals, and noindex flags.

15%

Search basics

Titles, meta descriptions, and OpenGraph/Twitter cards: the metadata that decides how you show up in results and shares.

20%

Content depth

Whether your page clearly explains what the site, product, or service does, how it works, and who it's for, with enough real text to back it up.

20%

Supporting pages

The useful pages people expect beyond your homepage: use cases, features, comparisons, and guides.

15%

AI readability

How clearly an AI model can interpret the content and structure in the initial HTML, without assuming that a provider saw the same response.

10%

Trust signals

Structured data, trust pages, and pricing or security details that make your claims easier to verify.

Run the first two stages now

A free scan measures your readiness and asks one real buyer question of a live answer engine. No account needed.