AI Slop Is an AEO Liability, Not a Shortcut

There is an obvious-looking arbitrage in AEO. AI answers are synthesised from web content. Generating web content is now nearly free. Therefore: generate a great deal of it, cover every question in your category, and become the source engines reach for.

The logic is clean and the conclusion is wrong, in a way that is worth understanding mechanically rather than moralistically. The strategy does not fail because it is lazy or because search engines disapprove. It fails because it is self-defeating at the level of how the systems actually select sources.

The mechanism it inverts

An engine composing an answer is not looking for pages about a topic. It has plenty. It is assembling a response, and the useful question is which sources add something the others do not.

The concept is information gain: how much a page contributes beyond what is already established. A page that repeats the consensus is substitutable — drop it, and the answer is unchanged. A page carrying a fact that exists nowhere else is not substitutable. If the answer needs that fact, there is exactly one place to get it.

Now consider what a language model produces when asked to write about a topic. It generates the consensus. That is what it is for — it is a compression of what has been written, and its output is a fluent restatement of the distribution it learned.

So AI-generated content about an established topic is, almost by definition, maximally substitutable. Its information gain approaches zero because it was synthesised from the existing corpus. You have produced, at scale, the exact category of page an engine has the least reason to cite.

This is not a policy failing. It is arithmetic. You cannot add information to a corpus by sampling from it.

Three ways it actively costs you

Zero benefit would be tolerable at near-zero cost. The costs are real.

Site-level quality assessment. Google’s guidance on AI-generated content is explicit that its focus is content quality rather than production method — AI use is not itself a problem, and using automation to generate content primarily to manipulate rankings is a spam policy violation. That is a real distinction and worth reading carefully, because it is more permissive than the discourse suggests and it is not a free pass. The operative risk is that assessments increasingly operate at the site level: a large volume of thin pages can affect how the whole domain is treated, including the pages you care about. You are risking your good content to host your filler.

Dilution of your own signal. If a hundred generated pages surround your one genuinely original study, retrieval has a hundred more chances to surface something forgettable with your name on it. Your average quality is what shapes an engine’s disposition toward your domain, and you have just moved it down.

Factual exposure. Generated content contains generated facts. Published under your brand, an invented statistic or a wrong specification becomes a claim you made — one that can propagate into third-party content and, eventually, into training corpora you cannot edit. This is the same asymmetry that makes original research valuable: uncontested facts travel. A wrong one travels equally well.

The uncomfortable part: it half-works, briefly

Being honest about why smart people keep trying it.

It does sometimes produce a short-term lift. Volume can catch long-tail queries nothing else covers, and a brand-new site with nothing can go from zero coverage to some coverage. Teams observe the improvement, conclude the strategy works, and scale it.

Then it decays. The long-tail queries it caught are the low-value ones. The pages generate no citations because they say nothing citable. And the accumulated liability sits on the domain permanently, because pruning a thousand pages is much harder than generating them.

The pattern is familiar from every previous content-scaling arbitrage in search: a real short-term effect, a structural long-term reversal, and a cleanup cost paid by whoever inherits the site. What is new is that the generation cost has fallen far enough that the temptation is now available to everyone rather than to teams with budget.

Where AI genuinely helps in content production

The argument here is not that using language models to write is illegitimate. It is that using them as a source of substance is.

The distinction is what the model contributes. Useful uses:

• Drafting from your material. You have interview notes, product data, an internal analysis. The model turns it into prose. The substance is yours; the assembly is automated.
• Restructuring for extractability. Taking a well-informed but poorly organised page and making its claims self-contained and quotable. Real work, and the model is good at it.
• Format conversion. A conference talk into an article, a report into a summary, one long piece into several focused ones. Content repurposing for AEO covers this.
• Editorial pressure. Asking what a draft fails to answer, where it hedges, what a sceptical reader would object to.

Harmful use is the single case where the model supplies the substance: “write me an article about X.” There, everything on the page came from the corpus, and you have published a lossy copy of what engines already had.

The test is straightforward and worth applying to any content process: could this page have been produced without access to anything proprietary to us? If yes, it has no information gain and no reason to be cited. Scaling AEO content production covers doing volume without falling into this.

What actually scales

The frustrating truth is that the things engines reward are the things that resist automation — which is precisely why they remain competitive advantages.

Original data. First-hand experience. Genuine expertise applied to a specific question. Proprietary aggregates from your own operations. Anything only you could have written, because you were the one who did the thing or has the data.

These do not scale to a thousand pages a month. They scale to perhaps one genuinely original piece a quarter for most teams, and one such piece will out-cite a hundred generated ones for years. Creating original research for AEO covers producing it without a research team.

The counter-argument

The strongest objection: this is a moral argument dressed as a technical one, and plenty of AI-assisted content ranks and gets cited perfectly well.

Partly right, and the concession matters. Content produced with AI assistance from real substance does perform, and there is no detector cleanly separating “AI-assisted” from “human-written” in a way that determines outcomes. Google’s own position is method-agnostic, and it is correct to be.

But the objection quietly substitutes a different strategy for the one under discussion. “AI-assisted content built on real substance” is not the arbitrage. The arbitrage is substanceless volume, and the argument against it is not that a machine wrote it — a human writing a thousand pages of consensus restatement would achieve exactly the same nothing. The mechanism is indifferent to authorship and sensitive to information gain.

Which is the useful reframing. The question was never “did AI write this.” It is “does this page contain anything that did not already exist.” That question has always been the right one; generative tools have simply made it cheap to answer wrongly at scale.

The one-line version

You cannot synthesise your way into a corpus you synthesised from. Publish what only you have, and use the tools to make it clearer — not to manufacture more of what everyone already has. Multi-engine monitoring will tell you soon enough which of your pages engines actually reach for, and the answer is rarely the ones that were cheapest to produce.