The retrieval layer between users and LLMs has its own preferences for what gets cited, and they don’t map neatly onto traditional SEO signals.
A debate keeps surfacing in marketing circles that reveals how poorly the industry understands what’s actually happening inside AI search. One side says: stop trying to optimize for LLMs, they’re just next-token predictors, just do good SEO. The other side says: no, there’s a retrieval layer between the user and the model, and that layer has preferences.
The first camp isn’t wrong about the technical fundamentals. Base LLMs don’t store URLs. They don’t remember where they learned things. They are, in fact, frozen statistical snapshots. None of that is in dispute.
But “just do good SEO” is the wrong conclusion to draw from correct premises. And it’s a mistake that even technically literate marketers keep making.
The part between the query and the model
When someone asks ChatGPT or Perplexity a question, the base model isn’t rifling through its training data to find your blog post. A retrieval layer, usually RAG (retrieval-augmented generation), goes out to the web, pulls candidate sources, reranks them, and feeds the most relevant chunks to the model as context. The model then synthesizes a response from those chunks.
This retrieval layer is not Google. It doesn’t care about your domain authority the way Google does. It doesn’t care about your backlink profile the way Google does. It has its own logic for deciding what gets surfaced and what doesn’t.
That’s the part most marketers haven’t internalized yet.
Why “just do good SEO” breaks down
The evidence is already visible if you know where to look. Pages that don’t rank in Google’s top ten for a given query are showing up as cited sources in AI-generated answers for that same query. Pages that do rank highly in Google sometimes get passed over entirely.
This shouldn’t be surprising. Google’s ranking algorithm weighs signals like
- click-through rate,
- backlink quality,
- and domain authority.
The retrieval layer feeding an LLM weighs different things:
- how semantically close the content is to the query,
- how extractable the information is,
- how clearly structured the page is for a system that needs to pull discrete facts rather than serve a full page to a human.
Good SEO and good AEO overlap in some places and diverge in others. Treating them as the same thing means you’ll keep optimizing for signals that the retrieval layer doesn’t prioritize and ignoring signals it does.
What the retrieval layer actually responds to
Three things consistently matter for whether your content gets pulled into an AI-generated answer:
Semantic alignment to the query pattern. Not keyword matching. The retrieval layer is doing vector similarity, looking for content that’s conceptually close to what the user is asking, even if the exact phrasing differs. Content that directly addresses a question in clear, specific language outperforms content that circles around a topic.
Extractable structure. The system needs to pull a discrete chunk of useful information from your page. Content organized into clearly delineated sections with specific, factual claims is easier to extract than long narrative paragraphs that bury the answer. Schema markup, clean heading hierarchies, and concise paragraph structure all make extraction easier.
Third-party corroboration. This is where Person A in the original debate gets something right, just for the wrong reasons. Earning mentions and coverage across other sources does matter, not because the base model “remembers” you from training, but because the retrieval layer has more candidate sources to pull from when your brand or claims appear across multiple pages. You become more citable because you’re more findable across the retrieval corpus.
The real shift
The deeper point isn’t about tactics. It’s about what kind of system you’re optimizing for.
Google serves links. The user clicks through to your site. You control the experience from there. The entire SEO discipline is built around getting that click.
AI search serves answers. The user might never visit your site. Your content’s job is to be the source the system pulls from when constructing that answer. The game isn’t getting the click. The game is being the information the model trusts enough to cite.
That requires a different editorial strategy, not just a different technical checklist. It means thinking about every page as a potential source for a synthesis engine, not just a destination for a human visitor. It means asking “can a retrieval system extract a clear, citable answer from this page?” as a standard editorial question.
Marketers who treat AEO as a subset of SEO will keep doing roughly the right things for roughly the wrong reasons. They’ll get some results by accident. But they’ll miss the structural shift underneath, which is that the retrieval layer is becoming its own channel with its own logic, and it rewards different content than the one we’ve spent fifteen years optimizing for.