LLM SEO: what actually gets a model to mention you
How large language models decide which companies to name, why your blog is rarely the deciding factor, and the work that changes the answer.
Most advice about optimising for large language models is either a rebranded SEO checklist or wishful thinking about prompt engineering. The useful version starts by understanding what the model can actually see.
Two different moments
A model can mention you for two entirely different reasons, and the work is different for each.
From training. The model absorbed text about you when it was trained. You cannot influence this quickly, it is frozen at a cutoff date, and it favours businesses with a lot of written coverage.
From live retrieval. The model runs a web search while answering and reads what comes back. This you can influence, within weeks, and it is where nearly all practical effort should go.
Almost every buying question triggers retrieval, because models are tuned to avoid stale commercial claims. So the retrieval path is the one that matters.
What the model reads before it answers
When retrieval fires, a small number of pages get read. In our own testing across four assistants and twenty commercial questions, the pages that got cited skewed heavily towards four types:
- Listicles and comparisons. "Best X for Y", "top 10 X", vendor comparison articles.
- Directories and marketplaces. Category listings with multiple options and structured fields.
- Review platforms. Anywhere with ratings, counts and recent dates.
- Community threads. Reddit especially, where a question has real answers from apparent practitioners.
Brand websites appeared, but usually as a supporting citation after the shortlist had already been formed elsewhere.
That single observation should reshape where you spend effort. Writing a better page about yourself competes for a slot that is mostly not available. Getting into a listicle competes for the slot that decides the answer.
First, check you are not blocked
Before any optimisation, check the boring thing.
Open your robots.txt and look for disallow rules affecting AI user agents. Several site builders and security plugins added blanket blocks over the last two years, often without the owner knowing.
There is a real distinction worth understanding. Some crawlers collect training data. Others fetch a page live because a user asked a question right now. Blocking the first is a legitimate business decision. Blocking the second removes you from answers today.
We have audited sites that spent months on content while quietly refusing every assistant at the door.
The work, in order of return
1. Get into the sources being cited
Run your buying questions, list the cited sources, and go and get into the ones that recur. One listing in a frequently read directory can put you into dozens of answers across many phrasings.
2. Make your facts liftable
Plain text answers to the obvious questions: what you do, for whom, where, what it costs, what is different. Models summarise text. They do not infer from a hero image or a carousel of logos.
3. Be consistent across the web
Same description, same category, same contact details everywhere. Contradiction is a reason to prefer somebody else.
4. Accumulate reviews and state them
Count and rating turn up in answers repeatedly. Get them, and write the number on your own site as text.
5. Publish the comparison content yourself
The one place your own site genuinely competes is honest comparison. "X vs Y", "alternatives to Z". Models like these pages because they are the shape of the answer being requested. Write them fairly, including where a competitor is the better choice, or you will not get cited.
6. Structured data and llms.txt, last
Both are cheap and neither is decisive. Do them when the four items above are done.
How to know if it is working
Ask the same buyer questions weekly across several models, record who is named and what is cited, and watch your share of answers over a month. Anything less frequent is not measurement, and any single check is noise.
Track competitors in the same run. A month where your share drops and every competitor's share drops is a model update, not a failure on your part. You cannot tell the difference without the comparison.
The summary
LLM SEO is mostly getting into other people's pages, saying the same clear true thing everywhere, and measuring on a schedule. The technical layer is real but small. The distribution layer is where the answer is decided.
Questions people ask
Do LLMs read my website?
Sometimes, if it ranks for one of the searches the model runs. More often the deciding sources are third party lists, directories and review sites, because they read as neutral and cover several options at once.
Does blocking AI crawlers hurt me?
Yes, if you block the ones that fetch pages to answer live questions. Blocking training crawlers is a separate decision. Check your robots.txt before you conclude you have an optimisation problem, because you may have a permission problem.
Is there an llms.txt standard I should adopt?
Adoption is still low and no major assistant treats it as authoritative. It costs little to publish one, but it is not a substitute for being in the sources that get read.
How is this different from normal SEO?
The work overlaps, but the target changes. Classic SEO aims your own page at a query. LLM SEO aims at the handful of third party pages a model will read before it answers.
See it on your own competitors
Figo checks their ads, pages, rankings, reviews and AI answers every week, then tells you what to do in plain words. Set up in two minutes.
Keep reading