LLM optimization: the method for getting named by a language model

LLM optimisation is the work of getting your business named inside the answer a large language model gives, rather than ranked in a list of links. It has two halves, because a model builds an answer in one of two ways. It either writes from what it absorbed during training, or it fetches live pages while you wait and writes from those. Training answers move when your name appears widely and consistently across the public web over months. Fetched answers move when your pages can be reached, read and quoted this week. Work out which kind of answer your buyers are getting, then do the matching work.

Step one: find out which kind of answer you are dealing with

Ask the question your buyer would ask, in the assistant they use, and then ask the assistant a second question: did you search the web for that, and which pages did you read. Every major assistant will tell you. That one follow up decides your whole plan.

The defaults differ by engine:

• Perplexity searches on every question, so almost everything there is a fetched answer with sources attached.
• Google AI Overviews and AI Mode sit on top of Google Search. Google's own documentation on AI features states that to be shown as a supporting link, a page must be indexed and eligible to appear in Search with a snippet, and that there are no additional technical requirements.
• ChatGPT and Claude decide per question. A question about a current price, a comparison, or a named local business usually triggers a search. A broad question about a category often does not. We tested the ChatGPT side of this in does ChatGPT use live search or memory.
• Gemini is attached to Google Search and behaves closest to those surfaces.

Run your buyer questions across two or three of these and mark each answer as fetched or not. Most businesses find a split: comparison and pricing questions come back fetched, and "who are the best X" questions come back from training.

What a model can reach

A fetched answer can only use a page the engine could get and parse. Four things decide that, and all four are cheap to fix.

• Named crawler access. AI engines use their own user agents, and a robots file written years ago for Googlebot says nothing about them. OpenAI documents its crawlers: OAI-SearchBot for ChatGPT search, GPTBot for training, and ChatGPT-User for fetches a person triggered. OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. The walkthrough is in which AI crawlers should you allow.
• Text in the HTML. If the words only appear after JavaScript runs, treat them as words the engine may never see. View source, search for your key sentence, and if it is missing, render it on the server.
• No login, no interstitial, no bot wall. Plenty of CDN rule sets block unfamiliar agents by default, so read the access log for the agents above rather than assuming.
• One answer per URL. Retrieval works at the passage level. A page that answers one question has a passage worth lifting. A page that answers nine has nine diluted ones.

You do not need a new machine readable file for Google's AI surfaces. Google says exactly that in the documentation linked above. Put the effort into access and into the pages themselves.

How a fetched answer chooses its sources

Three behaviours matter in practice.

One question becomes several searches. Google documents a "query fan-out" technique for AI Overviews and AI Mode: the system issues multiple related searches across subtopics, then assembles a response from a wider set of pages than a classic result list shows. So the page that gets used is often the one answering a sub question nobody else writes about, not the one chasing the head term.

The passage is the unit. The model needs a self contained chunk it can quote without the rest of the page attached. Headings written as the question, with the answer in the first two sentences underneath, is what makes a page quotable. We set out the structure in how to write a page an assistant will quote.

Established reference sites carry a lot of the weight. In a Pew Research Center analysis of 68,879 Google searches taken from the browsing data of 900 US adults in March 2025, the most frequently cited sources in both AI summaries and standard results were Wikipedia, YouTube and Reddit. That tells you where to be present, not only what to publish.

The four levers, and what each one moves

Lever	What it changes	Shows up in
Crawler and render access	Whether you are eligible at all	Fetched answers, within days
One page per buyer question, answer first	Whether there is a passage worth quoting	Fetched answers, within weeks
Checkable facts: counts, prices, dates, sources	Whether a model will commit to naming you	Both kinds of answer
Your name on sites you do not own	How many independent sources support you	Training answers, over months

Most businesses have the first two half done and the fourth not started. That is usually why a well written site still is not named.

What moves a training answer

Nothing you publish today lands inside a model that has already been trained. What you can change is what the next crawl of the public web says about you, and that is a different job from writing pages.

• Be the same entity everywhere. One legal name, one trading name, one category description, one address, repeated identically across directories, review platforms and your own site. Variation splits you into two weak entities instead of one clear one.
• Get described by other people. Trade press, category directories, review platforms, conference listings, partner pages. A description written by somebody else outweighs the same sentence on your own homepage.
• Make the category claim specific. "Insolvency software for resolution professionals" is an entity a model can file. "Business solutions provider" is not.

How to tell whether any of it worked

Ask a fixed set of buyer questions in each engine on a schedule, run each one more than once, and record two things: whether you were named, and which sources were cited. The citation list is the more useful half, because it names the pages to influence next.

Keep your own company name out of the test question. Asking "is our product good for this" primes the answer and you will see yourself everywhere. There is a protocol for testing fairly in should you ask ChatGPT about your own company, and a cadence in how often you should measure AI visibility.

The method only means something if you can tell whether it worked. That needs a reading before the work starts and the same reading after, on identical questions, because an answer varies enough between runs that one look proves nothing. AI Knows Us pins a question set, asks it without naming the business, and keeps every answer dated, which is what turns the method into a before and after rather than a theory.

This page is the method for being named by a language model. For the naming question between LLMO, GEO, AEO and AI SEO, see LLMO and how it relates to GEO, AEO and SEO. For how this sits beside ordinary search work, see GEO versus SEO.

Questions buyers ask about LLM optimization

Is LLM optimisation different from SEO?

The technical foundation is shared and the target is not. Ranking puts your page in a list a person chooses from. LLM optimisation puts your name inside the answer they were handed. Being indexed and crawlable is the entry ticket for both, which is why the technical half of SEO carries straight over.

How long does it take to see a change?

Access fixes show up in fetched answers within days of the next crawl. New pages that answer a buyer question tend to appear in fetched answers over a few weeks. Training answers move on the timescale of the next model, so treat third party presence as a standing programme rather than a campaign.

Does structured data help?

It helps an engine understand what your entity is, and Google asks that your structured data match the visible text on the page. Correct organisation and product markup is worth having, alongside the page saying the thing in plain words, because the plain words are what gets quoted.

Can I block training and still be cited?

Yes, and the controls are separate. OpenAI documents that a site can allow OAI-SearchBot so it appears in ChatGPT search answers while disallowing GPTBot so its content is not used for model training. Decide those two independently rather than with one blanket rule.

How many pages do I need?

Start with the ten questions your buyers ask before they choose, and one page for each. Ten specific pages beat a hundred vague ones, because retrieval rewards precision.

What to do first

• Pick your ten buyer questions, in their words. Ask them in two engines without naming your company, and save both the answers and the cited sources.
• Ask each assistant whether it searched. Mark every answer as fetched or from training, and split your plan along that line.
• Check your robots file and your access log for OAI-SearchBot, GPTBot, ChatGPT-User, PerplexityBot and ClaudeBot. Fix anything blocked by accident.
• Take the three questions where a competitor is named and you are not, and write one page each, title as the question and the answer in the first paragraph.
• List every site outside your own where your name appears. If that list is short, closing it is worth more than another page.