LLM SEO is a contradiction in terms

"LLM SEO" has a problem before you even start: it cannot mean what it says. SEO is optimising for a search engine. You learn how the ranking system weighs things, and you give it more of what it rewards. A large language model is not a ranking system. It is a frozen set of weights trained on a snapshot of the web, plus, sometimes, a live web search bolted on the side. There is no dial inside it to turn. So LLM SEO is a contradiction, and the interesting part is what the contradiction hides. Take it apart and everything sold as LLM SEO turns out to be one of two things: impossible, or just SEO.

How a large language model actually works

Two mechanisms, and keeping them apart is the whole game.

The first is training. A model like ChatGPT, Claude or Gemini learned language and facts from an enormous snapshot of text, frozen at a cutoff date. That snapshot is where most of its default sense of "who is good at what" comes from. You cannot edit it. You could not edit it if you worked there. It is done, shipped, baked in. Whatever a model "thinks" about your brand straight out of the box is a photograph of the web as it was months, sometimes years, ago.

The second is retrieval. When a model searches the live web mid-answer, what the industry calls RAG, or retrieval-augmented generation, it reads current pages, usually through an ordinary search index like Bing's, and quotes what it finds. This part is live. This part you can influence.

Almost every confident claim about LLM SEO quietly blends the two together. One popular guide says the practice "focuses specifically on the model layer." (At the time of writing this is laughably being returned as a featured snippet in Google search results for "LLM SEO"). That is the contradiction in a single phrase. There is no model layer you can reach.

You cannot optimise the training data

The training snapshot is the part people most want to change, because it is the model's default opinion of you. It is also the one part you have no access to whatsoever. You cannot submit to it, tweak it, or buy your way into it. The only thing that puts you in a training snapshot is having already existed, widely and credibly, before the cutoff. That is brand and time. It is not a tactic you run this quarter.

So spending effort to "optimise" the static training data is optimising for material you do not control. It is a fool's errand, and a lot of LLM SEO advice is exactly that errand dressed up as a method.

The part you can influence is just SEO

I am not telling you nothing can be done. The opposite. The live retrieval layer is wide open, and it is where almost all of your current, changeable visibility actually comes from. When a model searches the web to answer a question about your category, you want to be what it finds and trusts: a clear, well-structured, genuinely authoritative source on the thing you do. And if you have never checked what they currently say about you, a free LLM visibility audit will show you in about thirty seconds.

But look at what that is. Being the page a live search picks up and quotes is being findable, credible and quotable in search. That is search engine optimisation. The same crawlability, the same structured clarity, the same earned authority, pointed at the same retrieval in the same way. The "LLM" in front adds nothing you can act on. You are doing SEO, plus a little structure that makes you easy for the robots to quote cleanly, plus the entity work of being unambiguous about who you are. I wrote about that entity problem already, in why GEO is a bad name for it.

So LLM SEO splits cleanly down the middle. The training half is impossible. The retrieval half is SEO. There is no third thing, and there is no lever on the model.

The grift the contradiction feeds

Because the term implies a lever that does not exist, it breeds tactics that look like cleverness and are really just gaming. The one I see most right now is the self-published "best of" list.

You will have seen them. An agency publishes "the 10 best marketing agencies in [city]" on its own blog and, what are the odds, they're ranked number one. The logic is that models read listicles when they recommend, so if you write the listicle you control the recommendation. I have watched this spread ridiculously rapidly over the last year, and it is obvious for the same reason it has always been obvious: a company ranking itself first on its own website is not evidence of anything except that the company has a website (and a pretty high opinion of itself). Even burying your own company further down the list among other brands you don't actually consider competitors (and certainly wouldn't provide a link to) isn't going to pull the wool over anyone's eyes, though nice try.

There is a real signal nearby, however, which is why people reach for it. Being independently included in other people's "best of" lists, the ones you didn't write, genuinely does feed how both models and search judge you. But that is the opposite of publishing your own. The self-serving version is not future-proof in either direction. Google spent two decades learning to discount sites that rank themselves, and it catches the pattern quickly now. The live indexes the latest AI models lean on will learn the same lesson, because the tell is so easy to spot: a source that always concludes the source is the best is not a source, it is an advert. Optimising around it is, once again, dressing up something you do not control as something you do.

So what do you actually do

Drop the idea that there is an LLM to optimise, because there isn't. Do the two real things instead. Be genuinely, demonstrably good and well-documented now, so that when a model searches live it finds a clear, quotable, trustworthy source and picks you. And earn real third-party credibility, the mentions and inclusions you did not write yourself, so that both today's retrieval and tomorrow's training snapshot have a reason to tie your name to your subject.

That is less exciting than a 24-hour hack to rank on ChatGPT, and the people selling those will keep selling them, because the contradiction gives them just enough cover to. But it is the work that compounds, and it always was. We just keep giving it new names.

There is no LLM to SEO. We've always been optimising for robots. We've just got more of them to optimise for now.