Four names describe one practice. Here is what the term means, which of the differences from classic SEO are real rather than asserted, and how you would know whether any of it is working.
LLM SEO is the practice of trying to influence whether a large language model names your brand when someone asks it a question.
The borrowed half of the name causes most of the confusion. In search there is a position — you are third, and you can see it. In a generated answer there is no position to hold. There is only whether you were named, how consistently you are named across repeated asks, and whether anything of yours was cited. Presence, not rank.
If a tool or an agency offers to report your “rank” in ChatGPT, ask what is being ranked. The honest answer is a presence rate across repeated runs, which is a useful number wearing a misleading label.
LLM SEO, Generative Engine Optimization (GEO), Answer Engine Optimization (AEO) and AI SEO are competing labels for substantially the same work. No meaningful distinction between them has settled, and in the wild the terms are used interchangeably by people describing identical activities.
Where someone does draw a hard line — AEO is about direct answers, GEO is about generated ones, LLM SEO is about the model rather than the product — it is usually in service of selling one of them. The distinctions are real in theory and almost never operative in practice, because the same work improves all of them.
Practical advice: pick the term your audience already uses and stay consistent. The vocabulary will consolidate eventually and it will not be decided by whoever defined it most insistently.
Most claims about how AI search differs from classic search are asserted rather than measured. Three differences we can put numbers against, from our own measurement of 1,690 answers across four engines:
Perplexity included at least one source in 92.6% of its answers. ChatGPT did so in 27.8%. That is a 3.3x spread, and the ordering held for every brand we could compare across all four engines.
The consequence is strategic rather than trivial. On an engine that rarely cites, there is no link to earn and no referral traffic to measure — being recommended is the entire prize, and it is invisible in your analytics. A single “AI visibility” number that blends engines together averages this away.
Across the same answers, the share of each engine’s citations pointing at the brand’s own domain ranged from 2.8% on Perplexity to 11.1% on Claude. Even at the high end, roughly nine citations in ten went somewhere else.
This is the sharpest break from classic SEO, where your own pages are the asset you optimise. Here they are a minority shareholder in the answer.
Of the ten most-cited third-party domains in that dataset, seven were competing manufacturers’ own websites. When an engine answered a question about the category, it was substantially reading the competition’s marketing.
Publishing more pages on your own domain does not change that by itself, because your domain was never the contested surface.
Less changes than the category’s vocabulary implies. The models read the web, and they read it through much the same lens as a search crawler.
What does not carry over is the measurement. Rankings, impressions and click-through have no equivalent, which is why so much of this discipline is really a measurement problem wearing an optimisation costume.
llm.txt is a proposed file, similar in spirit to robots.txt, offering a site’s content to language models in a clean, readable form. It comes up constantly in LLM SEO discussion, so it is worth being precise about its status.
It is a proposal, not a standard. No major engine has committed to reading it, and there is no public evidence that publishing one changes whether a brand is named. It also costs very little to add. Both things are true at once: treat it as a cheap option on a possible future rather than a tactic with a known return, and do not let it displace work with evidence behind it.
Not to begin with, and the sequencing matters more than the spend.
The first useful step is free: write down the ten questions a buyer asks before they know your name, ask the engines, and record whether you appear. Most teams discover two things immediately — that the engines know them perfectly well when asked directly, and that they are absent when the question does not contain their name. That gap is the brief.
What you eventually pay for is not access to an answer, which is free, but consistency, a competitor baseline measured on the same prompts, and enough history to distinguish a real change from the ordinary variation between two runs of the same question.
The mechanics are identical; the leverage is not. A regional firm will not outrank national publishers on the sources a model reads, so publishing more on your own site is usually the weakest available move. The realistic route is being accurately and thoroughly represented wherever your category is described — directories, review platforms, trade bodies, partner and supplier sites.
Measure before assuming the worst. Smaller brands are often named more often than they expect, because the model is drawing on a category description rather than a popularity contest.
Three numbers, tracked on a fixed prompt set at a fixed cadence:
This page is deliberately about the term, the evidence and the measurement. For the practice itself — how AI systems assemble answers, the components of the work, and the strategies that follow — the Generative Engine Optimization guide covers it under the other name.