LLM SEO: how language models actually see your site

What is LLM SEO?

LLM SEO is optimizing your web presence for large language models, the systems behind ChatGPT, Claude, Gemini, and Perplexity. It is the same discipline as AEO viewed from the model's side: instead of asking "how do I win the answer," it asks "what can the model actually see about me, and through which channel."

The channel question is the useful one, because a model meets your business in two entirely different ways, on two entirely different clocks, and most confusion about "ranking in ChatGPT" comes from mixing them up.

Do LLMs learn about you from training data or live search?

Both, on different clocks. Training data bakes in what the wider web said about you months or years ago and cannot be edited, only outgrown. Live retrieval fetches current pages through search indexes at answer time, and it is where a new or newly optimized site can win visibility in weeks rather than model-generations.

You can watch both channels arrive in your server logs. Training-side crawlers (GPTBot, ClaudeBot, Google-Extended) collect content for future models; retrieval-side agents (OAI-SearchBot, ChatGPT-User, PerplexityBot) fetch pages to answer a live question. On sites we track, both kinds show up within days of launch when access is open; the verified counts by bot are published in the pillar guide.

The strategy split follows the clocks. For live retrieval: indexation on Google and Bing, extractable structure, and fresh, dated facts, all covered in how to optimize for AI search. For training data: consistent, corroborated information about your business across the web, so the next model generation bakes in the right facts. The second is slow, which is exactly why it is defensible.

What can a language model actually see on your page?

Text, structure, and structured data; nothing that lives only in pixels or scripts. Headings, paragraph order, lists, tables, and JSON-LD all survive into what a model processes. Meaning carried by images without alt text, JavaScript-only content, or visual layout is invisible. If removing the styling destroys the message, the model never got it.

This is why the unglamorous mechanics (semantic headings, real HTML tables, schema, llms.txt) keep appearing in every guide in this hub: they are the parts of a page that exist from the model's side. A quick self-test: read your homepage in reader mode. What survives is roughly what an LLM sees; if the pitch disappeared with the design, so did your AI visibility.

What is AI visibility, and how do you measure it?

AI visibility is whether AI systems can find, ingest, and name your business. It is measurable at three stages: verified crawler traffic (ingestion), recorded engine answers to a fixed question set (citation), and AI-source referrals in your analytics (payoff). Each stage is checkable in data you own, no proprietary score required.

The order matters because it tells you where to work. No verified crawls means an access problem (step 1 of the checklist); crawls but no citations means an evidence problem (structure, specifics, corroboration); citations but no inquiries means a business-page problem, not an AI problem. Skipping the diagnosis and buying "more AI visibility" as a bundle is how budgets vanish.

Does LLM SEO matter for a local business?

Yes, because assistants now answer "who should I hire near me?" directly, and they answer it from the evidence layer a local business already half-owns: the Google Business Profile, reviews, directory listings, and a site that answers local questions plainly. The businesses that structure that evidence get named; the rest are invisible.

For an Ottawa or Gatineau business the practical program is the AI search optimization service layered on local SEO: same evidence, both audiences. The local edition of the whole argument lives on the Ottawa hub, and the free audit will tell you which stage (access, evidence, or payoff) your site is stuck at today.