What LLM SEO means
LLM SEO is optimising for large language models — the systems behind ChatGPT, Gemini, Claude, Perplexity and Copilot — so they name your business when someone asks a question you could answer.
It is the same discipline as GEO and Answer Engine Optimisation, framed around the models rather than the answers. The distinction is marketing, not method. What genuinely differs is the target: optimising for a model that writes an answer is a different job from optimising for an index that ranks pages.
Two ways a model can know about you
This is the distinction that makes the whole subject make sense, and most explanations skip it.
| How it works | What influences it | How fast you can change it | |
|---|---|---|---|
| Training knowledge | Baked in when the model was trained, then frozen | Broad, long-standing presence across the open web | Slowly, indirectly, and never on demand |
| Retrieval | The model searches and reads live pages before answering | Whether you are findable and quotable right now | Weeks |
Retrieval is where essentially all your leverage sits. A model's training knowledge of a Poole plumber is thin, was fixed at a cutoff date, and is not something you can edit. What you can influence is whether — when it goes looking — it finds you, and finds you clear enough to quote.
There is a corollary worth internalising: if a model answers without searching, you are at the mercy of its training data and there is little to be done. The businesses that get named in those answers are typically national brands with decades of web presence. Competing there is not a realistic goal for a local business, and any agency implying otherwise is selling a fantasy.
Where each model retrieves from
They differ, and the differences change where you should spend effort. Treat this as current rather than permanent — it moves.
| Retrieves via | What that means for you | |
|---|---|---|
| ChatGPT | Bing’s index when it searches | Bing coverage matters far more than its market share suggests |
| Microsoft Copilot | Bing’s index | Same work as ChatGPT — one job covers both |
| Google Gemini | Google’s index, plus Business Profile data | Ordinary ranking work feeds it; your profile feeds it directly |
| Google AI Overviews | Google’s index | If you rank organically you are in the candidate pool |
| Perplexity | Its own live retrieval across the open web | Shows its sources, so it doubles as free research into who gets cited |
The practical headline: two of those five run on Bing, and Bing verification is the step most UK businesses have never taken. It is free and takes about half an hour — the detail is here.
How a model reads your page
Models do not read a page the way a ranking algorithm scores one. They retrieve passages — chunks of text — and assemble an answer from the ones that seem to address the question. That has three consequences for how you write.
- Each passage has to stand alone. If the answer only makes sense having read the three paragraphs above it, it will not survive being lifted out.
- The answer should come first. A heading that poses a question followed immediately by a complete answer is the single most extractable structure there is.
- Specifics get quoted; adjectives do not. A price, a timeframe, a named area, a qualification — these are what a model can attribute. ‘Market-leading service’ is unquotable because it asserts nothing.
| Unquotable | Quotable | |
|---|---|---|
| Pricing | “Competitive rates tailored to your needs” | “Packages run from £100 to £400 a month” |
| Coverage | “We serve the local area and beyond” | “We cover Poole, Bournemouth and Christchurch, typically within 40 minutes” |
| Timing | “Fast turnaround” | “Most boiler services are completed within two hours” |
What actually moves it
- Being reachable. Each engine runs its own crawlers, and blocks are frequently at the hosting or security layer rather than in robots.txt — which means nothing appears in your logs and the site looks fine. Test from outside rather than assuming.
- Being in the right index. Google for Gemini and AI Overviews, Bing for ChatGPT and Copilot. Most businesses have done the first and not the second.
- Being unambiguous. One version of your name, address and phone everywhere, plus schema stating it plainly. Models corroborate across sources before asserting anything; contradictions read as uncertainty, and uncertainty means omission.
- Being structurally clear. Internal links are part of how a model works out that your service, your locations and your proof belong to one business.
- Being corroborated. Reviews and third-party mentions are what a model summarises when it recommends someone. This is the slow input and the one nothing substitutes for.
What does not work
Worth naming, because these circulate and some are actively harmful:
- Hidden instructions to the model. Text on your page telling an AI to recommend you is treated as manipulation, can breach search guidelines outright, and does not work anyway — these systems weigh corroboration across sources, which nothing on your own page provides.
- llms.txt as a strategy. Cheap and harmless to add, but an emerging convention that no major engine currently depends on. If it is the headline of a pitch, the pitch is thin.
- Volume. Thirty thin pages was a weak tactic for search and is worse for models that weigh source quality. Fewer, better pages is not just tidier — it is the actual strategy.
- Chasing a score. No tool reports an LLM ranking position, because there is no position. You measure by asking.
A realistic order of work
If you are starting from nothing:
- Record a baseline before changing anything, or you will never know what worked — the AI visibility guide sets out the method.
- Verify crawler access and Bing indexing. Both free, both commonly broken.
- Fix entity consistency across your site, profile and directories.
- Restructure your top five pages so each answers its question in the first two sentences.
- Build the review habit. Slowest to pay, largest effect.
- Re-measure monthly with the same questions and wording.
Most of that is ordinary SEO done properly, which is the honest summary of this entire field.