How to improve AI visibility: the AEO loop

A five-stage AEO loop for measuring an answer, understanding who won, choosing and shipping one supported change, then checking what moved.

Quick answers

What if the answer does not move?

That is a legitimate outcome and should be recorded as one. It means the change did not close the gap you thought it closed, which is information: re-read the winning source, check whether the page is being retrieved at all, and consider whether the question is won off-site by a directory or community thread rather than by anybody's own page.

Can this be automated end to end?

Measurement, comparison, and verification can be. The decision and the change should not be, because both involve factual and brand claims that need a person accountable for them. The useful automation is everything either side of the approval, not the approval itself.

The short answer

AEO fails as a checklist and works as a loop, because each stage produces the input to the next. Measure a real answer. Understand who won it and why. Decide the one change that closes that specific gap. Ship it. Then ask the same question again and record what happened.

The discipline that makes it work is refusing to skip the first two stages. Most AEO advice starts at Decide — add schema, write comparison pages, improve trust signals — without ever establishing what the answers currently say. That produces a plausible backlog with no way to tell which item mattered.

Stage 1 · Measure reality

Two measurements, kept separate. The first is technical: can a named crawler fetch this exact URL, and what does the initial HTML contain before JavaScript runs? The second is observational: ask an unbranded buyer question of a live answer engine and record the answer verbatim, with the provider, API surface, and timestamp.^1,2

They must not be blended. A readable, indexed, permitted site that appears in no answers is a completely different problem from an unreadable one, and only separate measurements distinguish them.

In: one public URL, robots policy for each named bot, one unbranded buyer question.
Out: a readiness reading, and one verbatim answer with its provenance.
Cannot tell you: what any provider's own crawler received. You measured what you were served.

Stage 2 · Understand why others win

Read the answer as a document. Which brands does it name? What role does each hold — mentioned in passing, put forward as a recommendation, or cited as a source? Which pages did it lean on?

The useful comparison is between the winning source page and your equivalent. Usually the difference is concrete and small: the winning page states a fact in its body text that yours puts behind a form, or answers the question as asked while yours answers a broader one.

In: the verbatim answer and every source URL it cited.
Out: your state (mentioned, recommended, cited, absent), the other brands and their roles, and what the winning page did differently.
Cannot tell you: the model's reasoning. Compare the pages, not the intentions.

One question, carried to a result you can check›

A buyer question is asked of an answer engine. The answer names other companies and cites the pages it used. Comparing the winning page with your own produces one ranked change, which is reviewed as a diff before anything is published. After it is live, the same question is asked again on the same surface, and the result is recorded as improved, flat or declined — with the interval between the two checks stated, because that is what makes them comparable.

Stage 3 · Choose the next move

One change, chosen because of the gap you just observed rather than from a generic list. The test of a good decision at this stage is whether you can state the evidence that selected it in one sentence. If you cannot, you are working from a checklist again.

In: the observed gap, the evidence that identified it, and your readiness findings.
Out: one ranked action with its evidence attached.
Cannot tell you: what the answer will say next. Ranking is a judgement about likely impact, not a forecast.

Stage 4 · Make the change

Ship it, and keep it reviewable. Whether the change is pasted in by hand or opened as a pull request, a human should approve the factual, legal, and brand content before it is published. Automation that publishes on your behalf without review is trading a small amount of time for a category of risk that is difficult to bound.

In: the chosen action, and a connected repository or store if you have one.
Out: a published change, and confirmation that it is publicly visible.
Cannot tell you: whether it worked. Publication proves deployment and nothing else.

Stage 5 · Prove what moved

Ask the same question again, of the same provider, on the same API surface, and state the interval between the two checks. That is what makes a before and after comparable. Record the result honestly, including when it is flat.

A flat result is not a failed loop. It is a measurement that sends you back to stage two with better information than you had, and a loop that only reports improvements is not a measurement system.

In: the published change and its date, and the same question, provider, and surface.
Out: a comparable later answer, and an improved, flat, or declined status.
Cannot tell you: causation. Answers vary by time, surface, location, and personalisation, so a comparable check is evidence and not attribution.

Why the loop closes rather than ends

The output of stage five is the input to stage one. Answers move on their own as competitors publish, sources change, and providers update models, so a position held today is a measurement rather than a settlement. Teams that treat AEO as a project finish it once and quietly lose the ground back.

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.

Run the first two stages free

Sources and further reading

A new resource for optimizing for AI features — Google Search Central. Acknowledges AEO and GEO as terms in common use while stating that established search fundamentals still apply rather than being replaced.
ChatGPT Search — OpenAI. Describes query rewriting and live retrieval from search providers, which is why a measured answer is a sample at a moment rather than a stable ranking.

Continue learning

AEO evidence: what can it prove?AEO vs GEO vs AI visibility AEO glossary: 45 AI visibility terms How to track AI citations over time

See where your website stands

Run a free scan and get your AI readiness score across all six categories, with the gaps to fix first.

Run a free scan More guides