Footnote Works AI agency

LLM optimization

LLM optimization, without the guesswork.

LLM optimization is the practice of making a website readable, quotable and trustworthy to large language models, so that assistants like ChatGPT, Claude and Perplexity can use it as a source. It is mostly an engineering problem rather than a writing one: the model has to be able to fetch the page, parse meaning from the HTML without running JavaScript, and find a passage that answers a question on its own. Access, structure and extractable answers are the three things that decide it.

What actually moves it

In rough order of impact, and every one of these is observable from outside your site without any access to your server.

  • Crawler access. If your CDN or robots.txt turns away GPTBot, ClaudeBot or PerplexityBot, nothing else matters. This is the most common single failure and it is usually accidental.
  • Server-side rendering. A model fetching your page does not run your JavaScript. Content that only appears after hydration is content the model never sees.
  • Answer-first passages. Models quote text that stands alone. A paragraph that answers the question in its first two sentences is quotable; the same fact buried in paragraph six is not.
  • Structured data. Schema.org markup tells a model what a page is about and who published it, rather than leaving it to infer both.
  • Entity clarity. Consistent name, address and identity across your site and the wider web, so the model can tell which business you are.

What does not move it, whatever you have been told

llms.txt is worth having and costs nothing, but it is not a ranking mechanism. Ahrefs studied 137,000 domains and found 97% of llms.txt files received zero requests. We ship one on every build and we do not sell it as the reason anything works.

Keyword density, word count targets and the rest of the 2012 toolkit do not transfer. A model is not counting your keywords, it is deciding whether a passage answers a question well enough to repeat.

Test your own site first

Our free scan runs 24 checks and fetches your homepage twice more while identifying as GPTBot and as PerplexityBot, then reports what your server actually served each of them. That is the difference between policy and enforcement, and almost nobody tests it. The result is a shareable URL, there is no signup, and the scoring rubric is published on the page.

What it is worth, and who measured it

+8.4%

A 0.1 second improvement in load speed lifted retail conversion by 8.4% and average order value by 9.2%. Same traffic, same spend, more revenue.

Deloitte and 55, commissioned by Google, “Milliseconds Make Millions”, 2020. 37 European and American brand sites, over 30 million user sessions, measured across four load metrics.

These are findings from published research on the wider market, measured by the people named beside each one. They are not results we are promising you, and we do not have client results of our own to show yet, which we would rather say plainly than dress up.

Questions

What is LLM optimization?

Making a website readable, quotable and trustworthy to large language models so assistants can cite it. It covers crawler access, server-side rendering, answer-first content structure and structured data. It overlaps with SEO but optimises for being quoted rather than for being clicked.

Is LLM optimization the same as SEO?

They overlap and they are not the same. SEO earns a link in a list of ten. LLM optimization earns a sentence inside an answer. A page can rank well and never be quoted, and it can be quoted without ranking, because the two systems select sources differently.

Does llms.txt help?

Slightly, and much less than it is sold as. Ahrefs found 97% of llms.txt files across 137,000 domains got zero requests. Ship it because it is free, but if someone quotes it as their main mechanism, ask what else they have.

How do I know if AI can read my site right now?

Run our free scan. It fetches your page while identifying as GPTBot and PerplexityBot and reports exactly what your server returned to each, alongside 27 other checks. It takes about ten seconds and needs no access to your systems.

Related

Start here

See what the machines see, free.

24 checks on any URL, including a live fetch as GPTBot and PerplexityBot that reports what your server actually serves them. No signup, and every check is shown [1]