Get the audit

Guide

What is LLM SEO?

How language models learn about your business, and how to check and fix what they say.

On this page
  1. Two channels, two speeds
  2. Your first week
  3. The four test questions
  4. When a model quotes the wrong price
  5. Your site, from a model’s point of view
  6. Getting mentioned where models look
  7. Is llms.txt worth adding?
  8. Tracking progress
  9. FAQ
Short answer

LLM SEO is the job of making sure that ChatGPT, Claude, Gemini, Perplexity and other language models describe your company correctly and name it when a buyer asks about your line of work. A model picks up information about a business through two channels: the material it was trained on, and the pages it looks up while composing a reply. Good LLM SEO covers both channels.

Two channels, two speeds

Trained knowledge Lookup during the reply
When the model gets it months before the question, up to a cutoff date at the moment of the question
Source a large snapshot of public texts a handful of pages found right then
How soon your fixes count after the provider ships a newer model once the page is crawled again
What you can work on the number of public texts that name you, and whether they agree your site, plus other sites that name you
Typical sign usually a reply with no links a reply that lists its sources

Today’s assistants mix both. Linked sources nearly always mean a lookup took place. No links does not prove there was none, so ask the same question twice, once with browsing off and once on, and compare.

Your first week

  1. Day 1

    Ask the four test questions below in ChatGPT, with browsing off and on, twice each, and save every reply.

    Read more

    Add Perplexity and Gemini next month.

  2. Day 2

    Mark each wrong fact and, where links appear, the page it came from.

  3. Day 3

    Fix or redirect outdated pages on your own site: old price lists, archived posts, help articles, PDFs, campaign landing pages.

  4. Day 4

    Update your profiles on directories and review sites so they match your site.

  5. Day 5

    Email the owners of third-party pages that state something wrong about you and ask for a correction.

Then repeat the test questions once a month.

The four test questions

  1. “Tell me about [your company].” Reveals whether it has heard of you and has the basics straight.
  2. “Which [your category] would you recommend in [your market]?” Reveals whether you sit on the shortlist with rivals.
  3. “Compare [your company] with [main competitor].” Shows how you are framed.
  4. “How much does [your service] cost at [your company]?” Shows whether prices and other facts are current.

Run each one three times, because replies change between runs. This is a short version of the method in our GEO guide, which uses a longer list of buyer questions.

Linked sources nearly always mean a lookup took place.

When a model quotes the wrong price

The fix depends on where the old number comes from.

  • The reply links to a page. Open it. If the page is yours, edit it or point it to the current one, and make sure the new figure is plain text that appears even with scripts off, not part of a picture. If the page belongs to another site, write to whoever runs it and request an update.
  • No links, and turning browsing on gives the same wrong figure. Look for the old price on your own site first: price lists, blog archive, help center, PDF brochures, landing pages for past campaigns, partner pages. Search engines may still hold copies of these.
  • The wrong figure shows up only with browsing disabled. Then it is most likely part of the model’s trained memory. It stays put until the vendor ships a fresh version, and vendors give no dates. Meanwhile make sure every live source is correct, so that answers with search show the right number. For your manager, a fair summary is: “replies that browse improve once the sources are fixed; the model’s built-in memory changes only with the next release.”

We have no published before-and-after numbers on this topic yet, so treat the timelines in this guide as expectations, not guarantees.

Your site, from a model’s point of view

Page layout, such as answering first and using tables, is covered in our AEO checklist. For LLM SEO, also check:

  • Crawlers. OpenAI runs two separate bots. OAI-SearchBot finds pages for ChatGPT’s search answers, and according to OpenAI’s documentation sites that block it will not be shown in those answers. GPTBot gathers material for training, and blocking it does not affect ChatGPT search answers. Other providers also run separate bots for training and for search, so decide on each bot separately.
  • Facts as text. A price, a city or a service list that exists only in an image or appears after a script runs can be missed by a model.
  • No stale copies. An outdated page left online is a page a model can still quote. Refresh it or send it to the current one.

Getting mentioned where models look

A model forms its picture of a brand from many sources, not only from the brand’s own site. In practice:

  • Profiles on directories and review platforms in your industry, with the same name, description and facts as your site.
  • Round-up articles that come up for your buyers’ questions: offer their authors a free look at your product or data. Do not pay for a place in a list. Our GEO guide shows how one such reply was assembled from round-ups.
  • Articles under your name in trade media: they give models more places that tie your name to your field.
  • Corrections on third-party pages that get your facts wrong.

The fix depends on where the old number comes from.

Is llms.txt worth adding?

llms.txt is a suggested text file, placed at the site root, that points language models to the pages a company considers most important. It takes little time to add, but we know of no public evidence that it changes how often assistants mention or cite a site. We see it as optional and do the work above first.

Tracking progress

Each month, after running the four test questions, note:

  • does it recognize the company, and are the facts right;
  • does it put you on the shortlist for your category and region;
  • which URLs does it cite when browsing.

Corrections on pages that browsing reaches can show up sooner. The trained memory moves only with a new model release, on a schedule vendors keep to themselves.

FAQ

LLM SEO or GEO: what is the difference?

There is a big overlap. GEO is about being named and linked in AI replies, and in most products those replies browse the web first. LLM SEO takes that and adds a second question: what does the model believe about you without browsing, and is it accurate. The day-to-day tasks are largely the same.

Should we block AI bots?

Block training bots only for a clear reason, such as protecting paid content. Leave OAI-SearchBot, PerplexityBot and similar search crawlers allowed if you want your pages linked in browsing replies.

When will the model have it right?

For replies that use search, changes often appear within weeks of the sources being fixed, but providers set their own refresh schedules and we cannot promise a date. For the model’s trained knowledge, only after the next model release. We cannot promise what the next version will say, but accurate pages and profiles give it better material to learn from.

Can an agency help?

Mostly with monthly tracking and the work on other sites. See our AI search optimization service, or our white label offer if you run an agency.

Keep reading

More guides

AI visibility auditReply within one business dayGet the audit