What AEO Cannot Do

Every young marketing discipline goes through a period where the claims outrun the mechanisms. SEO had it — guaranteed first-page rankings, submission services, secret algorithm knowledge. It took years and a lot of wasted budget before the field settled into an honest account of what it could and could not do.

AEO is in that period now. So here is the list, written down deliberately, by a company that sells AEO tooling and would find several of these more convenient if they were not true.

You cannot buy your way into an answer

There is no ad placement inside an AI answer that you can purchase to be recommended. No inclusion fee, no submission programme, no partner tier that makes an assistant name you.

Some surfaces carry advertising adjacent to the answer, and that continues to evolve. But the recommendation itself — the part where an engine says “Acme is the strongest option for teams under 20” — is not purchasable. Any offer to place you there is either misrepresenting an ad product or selling something that does not exist. Can I pay to appear in AI answers covers the variations.

You cannot edit a model

If a model believes something false about your company, there is no correction form. No takedown, no appeal, no support ticket that reaches into the weights.

What has been learned is learned until the next model generation, and generations arrive on the vendor’s schedule. You can influence what live retrieval finds, which affects grounded answers comparatively quickly. You can change the third-party consensus, slowly. You cannot change the model.

This is the limit people find hardest to accept, and the one most likely to be quietly contradicted by a vendor pitch.

You cannot guarantee a position

Engines are non-deterministic. The same question asked twice returns different answers with nothing having changed. Anyone guaranteeing you a citation, a position, or a mention rate is guaranteeing the output of a stochastic system they do not control.

The honest version of that promise is: we will make you more likely to be selected, and here is the mechanism. That is a real service. “Cited within 30 days” is not.

You cannot attribute revenue to a mention

Someone reads an answer naming you, forms an impression, and buys three weeks later after typing your domain directly. Nothing links those events. No pixel fired, no referrer was set, no identifier persisted.

You can observe correlations — presence up, branded search up, pipeline up — and build internal evidence over time. That is genuinely useful and it is not attribution. Any model claiming to compute revenue from AI mentions is making assumptions you cannot validate, and the number it produces will be defended by nobody when challenged. Measuring GEO/AEO ROI covers what can honestly be claimed.

You cannot measure influence, only citation

When an engine composes an answer from ten retrieved documents and names three, the other seven shaped the response invisibly. Citation counts are a lower bound on influence, structurally.

And that is only for grounded answers. When an engine responds from training, there is no retrieval to observe at all — the influence is real and leaves no trace whatsoever. Every “AI cites Reddit heavily” style claim conflates these two mechanisms. Both are real; only one is observable.

You cannot see what people actually ask

There is no keyword volume for prompts, because there is no public query log and no scrapable surface. Every prompt-volume figure in the market is a panel, a search-volume proxy, or a model-generated estimate.

You are always monitoring a sample of your own choosing, not a representative slice of real demand. That is a real limitation on every visibility percentage anyone reports, including ours: the denominator is a set of prompts someone picked. There is no keyword volume for AI prompts covers the substitutes.

You cannot make results stable

A model update can change what an engine says about your entire category overnight, with no action on your part and no notice. This is the closest analogue to an algorithm update, except that it is less announced, less documented, and less analysable after the fact.

You can detect it — a category-wide shift affecting competitors simultaneously is diagnosable — and you cannot prevent it or appeal it. Model updates are the new algorithm updates covers living with it.

You cannot skip the fundamentals

No amount of AEO tactics compensates for content that does not answer the question, a site engines cannot render, an entity nobody can resolve, or a product nobody writes about. AEO is a layer over your existing substance. Applied to a thin foundation it makes thinness more legible.

You cannot do it quickly

Grounded answers respond to new content in days to weeks. Training-derived knowledge responds across model generations. Third-party consensus shifts over months. Entity representations are the slowest of all.

A programme starting today produces its first honest read in weeks and its first meaningful trend in a quarter. Anyone promising results next month is describing a timeline the mechanisms do not support. How long does AEO take is the realistic version.

Two things we cannot do specifically

Since a limits post from a vendor that only lists industry-wide limits is not really a limits post.

We cannot monitor a domain nobody signed up for. Every alert attaches to a project in a workspace. There is no way to hand us a domain from a logged-out page and have us watch it — running recurring AI queries against a domain nobody asked us to monitor is not a feature, it is an unbounded bill against an anonymous visitor.

Our own numbers are samples. Every score we show has a confidence interval we do not draw. Small week-over-week movements in our dashboard are frequently noise, and we would rather say so than let a two-point change be read as progress. How many runs before you trust an AI visibility number is the arithmetic, applied to our own product as much as anyone’s.

What is left

A shorter list, and worth having:

• You can find out what engines say about you, systematically rather than anecdotally, across every engine that matters. Nothing else in this article works without it, and most companies still do not do it.
• You can find out where the answers come from, at least for grounded ones. Which sources engines reach for in your category is knowable, and it is usually surprising — citation intelligence exists for exactly this.
• You can make your material citable. Clear claims, stated plainly, with sources. This has controlled experimental support behind it: the GEO paper found that adding quotations, statistics and cited sources measurably improved visibility in generative engines while keyword-style edits did not.
• You can publish what only you have. Original data creates a fact with one source, which is the one position no competitor can copy.
• You can be resolvable. Consistent naming and clear entity signals are cheap and they gate everything else.
• You can detect problems early. Detection latency — how long a falsehood about you circulates before you know — is fully within your control even when the falsehood is not.
• You can measure relative position. You may not be able to attribute revenue, but “we appear in 40% of the answers that matter and our competitor appears in 70%” is measurable, defensible, and moves when you work.

Why publish this

Two reasons, one principled and one self-interested, and it would be inconsistent with the article to pretend otherwise.

The principled one: a field where the claims outrun the mechanisms trains its buyers to discount everything, including the true parts. The overclaiming is bad for the honest practitioners specifically.

The self-interested one: we would rather be judged against what this can actually do. A customer who expects guaranteed citations will be disappointed by a product that measures reality accurately. A customer who wants to know what engines say about them, where it comes from, and whether it is changing is well served — and that is what exists.

The limits are not a disclaimer at the end of the pitch. They are the reason the remaining capabilities are worth anything.