Does AEO actually work?

We sell this service, so treat the page accordingly and check every claim on it. We have tried to write the version that survives being checked.

What is mechanically certain

A short list, and everything on it follows from how retrieval-based answering works rather than from anybody’s case study:

• A page that does not exist cannot be cited. Trivially true, and it is most of the job. In our own audit, all 90 citations went to interior pages that somebody had written.
• A page that cannot be fetched cannot be cited. Blocked in robots.txt, challenged at the CDN, JavaScript-only, behind a login — all of these remove a page from consideration entirely.
• A fact that is not in text cannot be quoted. A price in an image, a spec in a PDF, a claim only made in a video. Nothing can lift those.
• The engines do retrieve from the live web. Our own run returned 90 live URLs, several of them pages published recently. Answers are not solely a function of frozen training data.

If someone is doing only these four things for you, they are doing work that is certain to be necessary. That is not the same as certain to be sufficient.

What is plausible but unproven

• That the shape of a page changes whether it is chosen. We observed that 74 of 90 cited URLs carried a word from the question in the path in run A, on a strict literal rule (75 of 90 in run B), and that not one was a homepage in either run. That is a strong correlation in a sample of 90, and it is not a demonstration of cause.
• That structured data helps. No published evidence exists that we can find. We mark up every page anyway, for reasons set out on structured data for AI search, and none of those reasons is “it has been shown to work”.
• That llms.txt does anything. No major engine has publicly confirmed reading it. We publish one and say so.
• That being named in more answers sells more. Plausible, and nobody in this category has published attribution data connecting a citation to revenue. Including us. We have no clients long enough to have it.

What nobody has published, and why that matters

We looked for a controlled study — a fixed question set, a treated group and an untreated control, measured before and after, with the raw data available. We could not find one from any vendor in this category, and we have not run one either. What circulates instead is single-brand before-and-afters with no control, which cannot distinguish the intervention from the category simply getting more coverage over the same period.

What would count as proof

1. A fixed question set, published. If the prompts change between measurements there is no comparison, and a report that does not print its prompt list cannot be trusted to have kept them stable.
2. A baseline with the raw answers stored, not just a score. A score you cannot audit is a number somebody chose.
3. Repeat runs at each point. Engines are non-deterministic: our two runs a minute apart moved individual domain counts by one or two while the headline held. One run is not a measurement.
4. A control. Questions you deliberately do not work on. Without them you cannot separate your effort from the whole category moving.
5. Publication either way. A method that only gets published when it worked is marketing.

The test you can run yourself, in thirty days, for nothing

1. Day 1. Write ten questions a buyer asks before they know you. Split them into six you will work on and four you will deliberately leave alone. The four are your control.
2. Day 1. Ask all ten, twice, in a logged-out session. Record for each: were you named, in what position, and the full list of cited sources. Save the raw text. This is your baseline and its value depends entirely on being honest now.
3. Days 2 to 20. For the six, write one page per question. Title it as the question, answer it in the first forty words, put the specifics in text. The method is on writing a page an engine will quote. Change nothing about the other four.
4. Day 30. Ask all ten again, twice, the same way. Compare the six against the four.
5. Read it honestly. If the six moved and the four did not, you have weak but real evidence. If both moved, the category moved. If neither moved, thirty days was too short or the pages are not answering the question as directly as the pages that are being cited.

This costs nothing but your time and it is a better experiment than anything currently published, which is an indictment of the category rather than a compliment to the test.

The experiment we are running on ourselves

Our audited visibility score on 19 September 2026 was 0: named in 0 of 8 answers, 0 of 90 citations. We published it in full before we had anything good to report, and we are re-running the same eight questions on the same engine every month.

The falsification condition, stated in advance: if bungad.com is not appearing in the citation list for at least one of the eight questions by the third monthly run, the thesis that publishing question-shaped pages moves AI visibility is wrong for our case, and the answer is third-party sources — directories, review platforms and publications — rather than more pages of our own. We will publish that result as prominently as we would publish a good one.

Naming the condition before the data arrives is the only way a published result means anything. It is on our score page and in our content plan, and you can hold us to it.

So should you buy it?

On the evidence available today: do the four mechanically-certain things yourself first, because they are free and their necessity is not in doubt. Measure before you start, so you can tell later. Then decide about paying anyone — including us — with a baseline in hand rather than on a promise, using tool, agency or in-house to work out which shape of purchase you are even making. If your site is under twenty pages, spend the money on writing instead; that recommendation is on AEO for a small business on a budget and it costs us sales.

Common questions