LLM Optimisation
A language model encounters a business in three ways: through what it absorbed in training, through what it retrieves live while answering, and through what independent sources say. The first is effectively closed to you, the second is fully open, and the third is slow, expensive and the one that decides the hard cases. Budget accordingly, which usually means spending on retrieval this quarter and on corroboration for the next two years.
Rated 5.0 from 20 Google reviews · A local SEO agency based in Haywards Heath
Three ways in, and what each is worth
| Route | How a model gets it | Your control | Time to any effect |
|---|---|---|---|
| Training | Content crawled before a model was trained | Almost none, and none retrospectively | A model generation, if ever |
| Retrieval | A live fetch while the question is being answered | Complete, and immediate | Days to weeks |
| Corroboration | What independent sources say about you | Indirect, through work rather than words | Months to years |
The first line is where most anxiety sits and where almost no return is available. You cannot get a business into a model that has already been trained, you cannot get it out, and no supplier can do either on your behalf. Anybody offering to place your brand in a model’s training data is selling you something they cannot deliver.
The second line is where the work is. Retrieval happens now, uses the page as it stands this morning, and responds to changes within days. Everything practical about being named by an assistant runs through it.
Retrieval, in plain English
When an assistant answers a question it does not consult its memory of the web. It issues searches, fetches pages, and writes an answer from what it got back, discarding the rest. This is why a page published yesterday can be cited today, and why the state of your site right now matters more than everything you published before it.
It is also why access is not a detail. A page that a fetcher cannot reach, or that renders its content only after running scripts, is a page that is not in the pool at the moment the answer is written. No amount of quality compensates for not being fetched.
Why entity consistency does more than word count
An entity is the thing a model believes your business is: a name, attached to a place, a set of services, some people and a history. Every mention of you anywhere either reinforces that or muddies it. Two spellings of your trading name, an old address on a directory you forgot, a services list that says three things on your home page and five on your LinkedIn profile, and the model has a business it is not confident describing.
A model that is not confident does one of two things: it hedges, or it names somebody else. Neither shows up in any report you are running. The remedy is unglamorous and cheap, which is why it gets skipped: list every place your business appears, make the name, address, phone number and description identical, and link them to one another through your structured data so they resolve to a single entity rather than several similar ones.
Word count does none of this. A site can publish two hundred pages and remain ambiguous about what it is, and frequently does.
Corroboration is a link building problem wearing a new hat
The third route is what the rest of the web says. For anything the model treats as a recommendation, this is what decides it, because a claim made only on your own site has exactly one source and a model asked to recommend is at its most cautious.
The work is press coverage, trade association membership, industry directories that are genuinely used, client sites, case studies published by somebody other than you, and review platforms. That is digital PR, and the fact that it is now also the main lever on AI visibility is the single most useful thing to understand about this field commercially. It is slow, it does not fit in a quarter, and it is the part that competitors cannot copy in a fortnight.
What we do, and in what order
- Confirm every assistant that matters can fetch the site, and that nothing was blocked by accident.
- Rewrite the pages that carry commercial weight so a bounded answer can be lifted from each.
- Resolve the entity: one name, one address, one description, linked and consistent everywhere it appears.
- Set up measurement before making further changes, so there is a baseline to argue from.
- Then, and only then, the slow corroboration work, which continues for as long as you want the position held.
Frequently asked questions
Related guides
This page is one of our AI search and GEO guides. Start with the complete guide to AI search visibility or see our AI search and GEO services.
- Generative engine optimisation (GEO)
- How to write content an AI will cite
- Copilot and Claude visibility
Browse all AI search and GEO guides. Would you rather we did it for you? Start with an AI visibility audit.
Sources
- Google Search Central. Optimizing for Generative AI Features on Google Search. Updated 10 July 2026. Google’s description of retrieval augmented generation over its Search index, and the statement that no special files or markup are needed.
- OpenAI. OpenAI Crawlers and User Agents. Read 3 September 2026. The separation between the training crawler and the search crawler, which is what makes the training and retrieval routes distinct.
- Anthropic. Anthropic web crawlers and how to block them. Updated 7 April 2026. ClaudeBot for training and development, Claude-User for user initiated fetches, blocked separately.
- Ofcom. Adults’ Media Use and Attitudes 2026. 2 April 2026. 54 per cent of UK adults use AI tools such as ChatGPT, Copilot or Gemini, rising to about 79 per cent of 16 to 24 year olds, up from 31 per cent a year earlier.
Every figure on this page was read on the source page listed above. This page is reviewed monthly and the last checked date is updated when it is.
Frequently asked questions
Can we get our content into a model's training data?
Not deliberately and not on any useful timescale. Training sets are assembled from broad web crawls and closed before a model ships, and no supplier has a route in. Anyone who offers one is describing something that does not exist. Spend the budget on retrieval instead.
What is retrieval augmented generation, in plain English?
The assistant searches the web while it is answering you, reads a handful of pages, and writes its reply from those. It is the difference between recalling something and looking it up. Almost all of your practical influence sits in what it finds when it looks.
Does blocking AI crawlers protect our content?
It stops future crawls by the crawlers that honour it. It does not remove anything from models already trained, and it does not stop a user initiated fetch, which most providers document as ignoring robots.txt. Block on principle if you hold one, but not in the belief that it undoes anything.
How does this relate to our link building?
They are now the same work. What earns a link, which is being genuinely worth citing, is also what makes an independent source say something about you that a model can verify. If you already run digital PR, you already run half of an AI visibility programme.
Ask us how your business shows up in AI search
A friendly 15 minute call, no obligation.
or call 01444 523105