LLM SEO — what it is and what actually works

The short version

LLM SEO is optimising to be named inside a generated answer. Not ranked, named. An assistant asked a question returns prose with a handful of brands or sources in it, and you are either in that set or you are invisible. There is no position eleven.

It is also called AEO, GEO, AI SEO and LLM optimisation. Those are the same discipline under different labels — the distinction between them is mostly a matter of who is selling. If you want that unpacked properly, the terminology guide does it.

This guide assumes you already do SEO and want to know what genuinely changes.

What carries over from SEO, and what does not

More carries over than the louder voices claim, and less than most teams assume.

Still works: crawlability, clean information architecture, structured data, page speed in so far as it affects crawling, topical depth, and earning links from places that matter. Google’s AI Overviews in particular run on Google’s own index, so your existing work shows up there more or less directly.

Works differently: keywords. You are not matching a query string, you are trying to be the thing a model reaches for when it composes an answer. That rewards writing that states a claim plainly and then supports it, rather than writing that circles a keyword.

Does not carry over: position. There is no rank tracking in a meaningful sense because there is no ranking. Two runs of the same prompt can return different brands. Anyone selling you a “rank” in ChatGPT is selling you a metric that does not exist.

Actively hurts: keyword stuffing. Research on generative engines has consistently found it reduces citation likelihood rather than improving it — the opposite of its effect on classical ranking, where it was merely useless.

The four engines do not read the same internet

This is the single most expensive thing to get wrong, because optimising for one and assuming the rest follow covers roughly a quarter of the ground.

• ChatGPT resolves much of its retrieval through Bing, plus a handful of directories. If your Bing presence is thin, you are thin in ChatGPT.
• Google’s AI Overviews run on Google’s own index. Your existing SEO largely applies.
• Perplexity leans heavily on community discussion and publishes its citations, which makes it the easiest engine to debug — you can see exactly what it read.
• Copilot uses Bing plus editorial sources.

Being strong in one tells you very little about the others. Measure them separately or you are guessing.

LLM visibility: the part that is actually measurement

“LLM visibility” and “LLM optimisation” usually describe the measurement half of this: knowing whether you are being named at all, and what gets named instead.

There is no console for this. No platform publishes impressions, citations or share of voice, and none has announced plans to. So measurement means running a fixed prompt set repeatedly across the engines and recording what comes back — the brand named, the position in the list, and the source cited. The third column is the one that tells you what to fix.

Run each prompt more than once. Output varies between identical runs, and a single pass will tell you a story that does not replicate.

What actually moves it

Roughly in order of how often it is the binding constraint:

1. Crawl access. Confirm you are not blocking GPTBot, PerplexityBot, ClaudeBot, OAI-SearchBot and the rest in robots.txt. A surprising number of agency-built sites block them by default, which forecloses the whole exercise. Our own crawler checker does this in a paste.
2. Third-party presence. Brands are cited through sources they do not own far more often than through their own domain. Industry directories, review platforms, community threads and genuine editorial coverage do more work here than anything you publish.
3. Extractable structure. Lead with the answer, then support it. Headings phrased the way people ask. Real specifics — numbers, names, dates — rather than adjectives. Structured data describing what the organisation and its services actually are.
4. Citations and statistics in your own content. Generative-engine research finds that citing sources is among the strongest single levers on whether a passage gets used.
5. An llms.txt. Low cost, uncertain benefit, no adoption guarantee from any engine. We publish one because it costs an afternoon, not because we can prove it works.

What does not work

Astroturfing community threads. It is detectable, it is increasingly policed, and the blast radius when it is caught lands on the brand rather than the agency.

Mass-producing near-identical pages to cover a keyword matrix. This was already risky and the March 2026 update made it expensive. Generated pages that answer nothing are the fastest way to lose the crawl access point 1 depends on.

Buying “AI SEO” packages priced per keyword. The unit does not match the work — there is no keyword to rank, and the deliverable should be a measured answer-set, not a position report.

How to tell whether it is working

Pick the prompts your actual customers would type, not the ones you wish they typed. Run them across all four engines, three times each, and record: named or not, in what position within the answer, and which source the engine cited. Re-run monthly against the same set.

Watch referral traffic from AI hosts as a secondary signal — many engines strip or omit the referrer, so treat those counts as a floor rather than a total.

The honest caveat

Nobody outside these companies knows how source selection works, and the systems change without notice. What is written above is drawn from observed output and from research published on generative engines, not from inside knowledge. Re-baseline quarterly and be suspicious of anyone, including us, who describes this with more certainty than the evidence supports.