LLM SEO

When someone asks an assistant which tool, agency or product to pick, the assistant usually searches, fetches a handful of pages, and writes one answer grounded in them, often with citations. Three doors have to open in sequence: the crawler must be able to fetch your page, the model must be able to read what it fetched, and the page must contain something worth repeating. Most brands fail at a door they have never tested.

• Let the crawlers in
  GPTBot, ClaudeBot, PerplexityBot and Google-Extended each knock with their own user agent. Your robots.txt decides on paper; your CDN and firewall decide in practice, and the two disagree more often than anyone expects. Test the real behaviour, not the config.
• Survive without JavaScript
  Most AI crawlers read the HTML they are handed and run little or none of your JavaScript. If your text only exists after hydration, an assistant sees an empty page. Server render the words that matter.
• Say what you are, in text
  One page that states plainly what the product is, who it is for, and what it costs. Models quote sentences, not hero animations. Structured data and a canonical URL tell the retriever it found the right page.
• Publish direct answers
  For every question a buyer asks an assistant, somebody owns the page that answers it in the first paragraph. If that page is not yours, the answer cites whoever wrote it.
• Be on the pages assistants already cite
  Answers lean on comparison articles, review sites and forums. Which ones depends on your category, and it is measurable: collect answers, read the citations, and go where they point.

The first three are testable on your live domain right now with the free check: fifteen checks including whether a request identifying as GPTBot actually gets a page, no account and no card.

A single answer is an anecdote. The same question to the same engine returns different answers on different days, so a screenshot proves nothing in either direction. An honest measurement asks the same questions repeatedly across engines, counts how often you are named and cited, and states how many answers each number rests on. That is the method behind the sample report, where every metric carries its n.

If the acronyms are the confusion, AEO vs GEO vs SEO settles the vocabulary and where the budget goes.

• Is LLM SEO different from AEO or GEO?
  It is the same discipline seen from the machine side. AEO and GEO name the goal, being in the answer. LLM SEO names the audience, the language models doing the answering. If you are optimising for one you are optimising for all three.
• Do LLMs read my website at all?
  Two ways, and they fail differently. Training bakes in what the model absorbed months ago, which you cannot change quickly. Retrieval happens at answer time: the assistant searches, fetches pages, and grounds its answer in them. Retrieval is where you can move the needle this quarter, and it only works if the crawlers can actually fetch and parse your pages.
• Does blocking AI crawlers protect my content?
  It protects it from being read, which also removes it from consideration when an assistant writes an answer about your category. That is a legitimate trade for a publisher selling content. For a brand that wants to be recommended, blocking GPTBot while a competitor admits it is handing over the answer.
• How fast does any of this show up?
  Retrieval changes can show up as soon as an assistant next fetches your pages; reputation on third party sources builds over weeks. Nobody serious will promise a date, because answers are sampled from a model and vary day to day. What you can do is measure weekly and watch the rate move.