What AEO Cannot Do: The Honest Limits, With Our Own Failures as Evidence

Every page selling answer engine optimization lists what AEO achieves. Almost none lists what it cannot. That asymmetry is itself a buying signal: a discipline described only by its wins is being sold, not explained. This page lists five limits of AEO, and evidences each one with our own measurements rather than a disclaimer.

The short version

What AEO cannot do, in five sentences. It cannot guarantee a position, because positions move without any change to your site. It cannot convert being cited into being recommended, because those are different events bought in different ways. It cannot win answers that live off your domain, and 27 of the 56 answers on our own panel that cited anything cited a platform we do not own. It cannot work quickly, and our own score on 8 target questions has been 0% for 52 days. And it cannot be verified by the vendor selling it, which is why the proof standard matters more than the promise.

Limit 1: it cannot guarantee a position

What the measurement showed

We ran 8 Dubai buyer questions through Google AI Mode on 21 July 2026 and the identical 8 on 12 September 2026. Between those dates, the agencies the engine favoured changed completely. Push Group's pages held a 50% citation share in July and 0% in September. Digital Nexa was named in 37.5% of July answers and none in September. Only 20% of the July field appeared at all in September.

Why that matters to a buyer

Nothing visible changed about those agencies. Both still publish the same service pages. The engine re-sampled the category and returned a different answer. Any agency promising you a fixed position in an AI answer is promising something the engine does not offer, and our two-date receipt is the cheapest available proof.

Limit 2: it cannot turn citation into recommendation

Two different events

Being cited means the engine used your page as a source. Being recommended means the engine named you as the answer. Our July panel separated them cleanly: Push Group's pages were cited 4 times while the brand was named once; Digital Nexa was named 3 times while its pages were cited zero times. One agency supplied the words, the other got the credit.

What each one buys

Citation is earned with structure, freshness and machine-readable pages. Recommendation is earned by being distinct enough that the model has a reason to prefer you. That second one is a Voice problem before it is a technical one, and no amount of schema markup fixes a brand that reads as the category average. The full separation is worked through here.

Limit 3: it cannot win what lives off your domain

Where the answers originate

Across our 62-query panel (16-18 September 2026), the surfaces the engine cites most are not agency websites. YouTube, LinkedIn, Reddit and Quora together account for 54 of the 301 citations (18%) - the largest bloc, and present in 27 of the 56 answers that cited anything. Agency domains sit in a long tail below them, none above 2%.

The uncomfortable implication

On-domain work cannot win an answer sourced from a Reddit thread. That is a placement problem wearing a content problem's clothes, and an AEO retainer that only touches your own site is structurally unable to reach it. Ask any agency what proportion of your target answers are currently sourced off-domain before you buy on-domain work.

Limit 4: it cannot work quickly

Our own score

We publish our own numbers because a measured zero is more useful than a flattering estimate. Across those same 8 Dubai questions, Ivanooo scored 0% named and 0% cited on 21 July, on 7 September, and again on 12 September. Across the broader 62-query panel, our Share of Recommendation reads 0.0, with 59 of 62 questions returning us absent.

What that tells you about timelines

We run this method on ourselves and have not yet won these questions. Anyone quoting you a 30-day guarantee is quoting a sales cycle, not a measurement. The honest answer is that AI Search Visibility compounds from coverage held over time, and the time is not optional.

Limit 5: it cannot be verified by the seller

The circular-proof problem

Of the 10 agencies on our Dubai panel, none publishes pricing and only 1 shows any measured AI-answer evidence on its own service page. The rest sell AI visibility on claims. A dashboard supplied by the vendor being graded is not independent evidence, and neither is a case study with no method attached.

The shape of independent proof

A date. A named engine. The exact questions asked. The full answer saved. Counts anyone can reproduce by asking the same questions themselves. That standard costs nothing to meet and almost nobody meets it.

What AEO can do

Claim	Can AEO deliver it?	Evidence
Guaranteed position in an AI answer	No	100% turnover of leaders in 52 days
A page cited as a source	Yes, with structure and freshness	4 of 8 answers cited one agency's pages
Brand named as the recommendation	Only with distinctiveness	Named and cited diverged on 2 of 10 agencies
Winning an off-domain answer	Not with on-domain work	Platforms are cited in 27 of 56 answers
Results inside 30 days	No	0% on our own 8 questions across 52 days
Independently checkable proof	Yes, and it is rare	1 of 10 agencies shows measured evidence

Five things to do instead of buying a guarantee

1. Get a baseline before the first sales call. Ask the engines your buyers' questions, save the answers, and walk in knowing which ones you lose.
2. Ask for two measurements, not one. Any vendor can show a favourable snapshot. The gap between two dates is the only evidence of method.
3. Split your target questions into on-domain and off-domain. Buy on-domain work only for the answers your own pages can plausibly win.
4. Separate citation goals from recommendation goals in the brief, because they are bought differently and cost differently.
5. Set the proof standard in the contract. Named engine, dated runs, full answers stored, counts reproducible by you.

Questions buyers ask

Is AEO a scam, then? No. It is a real discipline with real mechanics, and the primary research from Aggarwal et al. at Princeton found that adding citations and statistics to a passage lifted its visibility in generated answers by up to 40% in controlled tests. What is unreliable is the promise layer sold on top of it.

Why publish your own zero? Because the alternative is asking you to trust a number we would not show. As Firoz Azees puts it: "a measured zero is a work order; a flattering estimate is a story you tell yourself." Our absence on 8 questions tells us exactly which answers to attack.

If positions churn, is any of this durable? Coverage is durable; position is not. The engines re-sample constantly, but a site that answers a subject completely gets re-sampled into more answers over time. That is why topical coverage held over time is the asset, not any single ranking.

How do I know whether my answers are off-domain? Ask your buyer questions and read the source list under each answer. If the sources are Reddit threads, YouTube videos and review sites, your own pages are not the battlefield.

Does this mean I should not hire an agency? It means hire on method, not on promise. An agency that shows you two dated measurements and tells you which of your questions it cannot win is describing reality. One that guarantees a ChatGPT recommendation is describing a sale.

What is the single most common overclaim? Guaranteed inclusion in AI answers. Our 52-day churn data shows why it cannot be honoured: the engine changed its Dubai shortlist entirely without any of those agencies doing anything wrong.

Can I check any of this myself? Yes, and that is the point of publishing the method. Ask the same 8 questions, save the answers, and compare them to ours. If your results differ from ours, ours were a snapshot and yours are newer.

At Ivanooo, Firoz Azees runs Distinctiveness Engineering for the AI-answer era: measuring who the engines name, cite and recommend, then engineering the gap between listed and chosen. We publish the limits because a method you cannot audit is a claim. Start with a free AI visibility check and find which answers you are absent from.