The short answer
LLM SEO is the practice of making a website a source that large language models find, quote and name. It is the same discipline as generative engine optimisation, viewed from the model's side rather than the product's.
The thing you are optimising for is retrieval and attribution, not position. There is no ranked list to climb.
Training or retrieval: which one are you optimising for?
This is the distinction everything else hangs on, and it is frequently muddled.
| Training | Retrieval | |
|---|---|---|
| When it happens | Once, when the model is built | Live, on each question |
| Can you influence it? | Not meaningfully, after the fact | Yes, directly |
| What it covers | General knowledge, often dated | Current, specific, commercial |
| How you affect it | By existing widely over years | By being fetchable and quotable now |
For any question involving a supplier, a price, a comparison or anything that changed in the last year, assume retrieval. That is where the work pays, and it pays on a timescale you can measure.
People sometimes ask how to get into the next training run. The honest answer is that you cannot target it, and that the things which would help, being described consistently across many independent sources, are the same things that help retrieval. Do those, and treat training as a slow by-product.
How retrieval actually picks a passage
Simplified, but close enough to act on:
- The system turns the question into one or more searches.
- It gets candidate documents back.
- It fetches some of them, usually without running JavaScript.
- It splits what it fetched into chunks.
- It scores the chunks for relevance to the question.
- It writes an answer from the best ones, sometimes citing their sources.
Step three is where a surprising number of sites fail silently. Step four is why the passage, not the page, is the unit of work.
What makes a chunk get chosen?
Four properties, in rough order of how much they matter.
- It answers the question by itself. No dependency on the paragraph above, no pronouns pointing at something out of frame.
- It is specific. Contains the nouns, numbers and qualifiers the answer needs. Reassurance is not retrievable.
- It sits under a matching heading. Headings are a strong and cheap signal.
- It is short enough to lift. Two to four sentences keeps your wording intact; a long block gets paraphrased and your framing disappears.
If you want one rule: write so that any section of your page could be pasted into an answer without an editor having to fix it first.
Why what others say about you outweighs what you say
This is the part people find hardest to accept. A model treats a claim repeated across independent sources as more reliable than the same claim made once, on the website of the company that profits from it.
That is sensible behaviour, and it explains a pattern that otherwise looks unfair: a thin directory listing being cited instead of a far better page on the company's own site. The directory is not better written. It is corroborated, and it agrees with other sources.
The practical consequence is that entity consistency is not a tidy-up task, it is the main event. If your service names, locations and descriptions differ between your website, your profiles, your Companies House record and the trade directories, you are asking the model to resolve a conflict, and it will resolve it using something other than you.
What about llms.txt?
A proposed convention: a plain-text file at your site root that describes the organisation and links to the pages worth reading. It is not a ratified standard and support is unverified.
Our position is that it is cheap insurance. It takes minutes, it does no harm, and if adoption grows you are already there. It is not a ranking factor, there is no evidence that it moves answers today, and anyone presenting it as a lever is guessing.
Does blocking AI crawlers protect anything?
It depends entirely on what you are trying to protect, and the two goals pull in opposite directions.
| If you want | Then | Trade-off |
|---|---|---|
| To be cited in answers | Allow the search and user-facing crawlers | Your content is used |
| To stay out of training | Disallow GPTBot and Google-Extended | May reduce long-run familiarity |
| Both | Allow search crawlers, block training ones | The usual sensible middle |
| Neither | Block everything | You will not appear in AI answers |
What we see most often is none of these: a robots.txt that blocks something by accident, usually because a named user-agent group was added and the existing rules were not repeated inside it. A named group replaces the wildcard group rather than adding to it.
How would you know if any of this worked?
Build a fixed prompt set from the questions your buyers actually ask. Run it across the assistants on a schedule. Record, every time: were you named, what was said, and which URL was cited.
Because outputs vary run to run, one result is an anecdote. The number worth watching is the rate at which you are named across repeated runs, and the direction it moves over months.
The service side of this is on our GEO agency page, and the wider programme on AI SEO. If you want a baseline first, the free AI visibility audit produces exactly this and the findings are yours to keep.