How I think about SEO for AI search engines

I have been rebuilding my own site partly as an experiment in this, so this is a working position rather than a settled one.

Most of it is old advice, done properly#

The uncomfortable truth about "optimising for AI search" is that most of the genuinely useful practices are things that were already good practice and that many sites never did properly.

Server-render your content. Use semantic HTML. Have a coherent canonical strategy. Describe your entities with connected structured data. Write something specific enough to be worth quoting.

That is not a new discipline. It is the old discipline, executed.

Crawlable means actually crawlable#

If your article text arrives after a client-side fetch, some crawlers will never see it. Same for identity, project descriptions, and links behind JavaScript-only navigation.

This one is easy to get wrong while feeling modern. Frameworks make it trivially easy to render on the client, and the page looks perfect in a browser. It is just not there for anything that does not execute JavaScript, and you will not notice unless you check the raw response.

I check by fetching the page and reading the HTML. If the thing I care about is not in there, it does not exist for a meaningful set of consumers.

That includes animations. Reveal-on-scroll is fine as an enhancement, and I use it — but content must never be reachable only by scrolling. On this site the reveal has a failsafe that shows everything after a couple of seconds regardless, because a non-scrolling agent should not see an empty page.

Specificity is what gets cited#

A model looking for something to cite is looking for a claim that is concrete, attributable and self-contained.

"Enterprise AI adoption requires careful consideration of data governance" is true and useless. Nobody quotes it, because it says nothing that could be wrong.

"Running a model on-premise does not by itself satisfy PDPL — data residency is one input into a legal assessment" is a specific position. It could be argued with, which is exactly what makes it worth quoting.

Writing things that could be wrong is the underrated part of this.

Entity clarity does most of the work#

For a person, being citable requires being identifiable. One canonical home, one consistent name, one set of profiles that agree with each other and link back.

The technical piece is a single Person identifier referenced from everywhere rather than a fresh Person node per page. The non-technical piece is making every external profile match — and that matters more, because corroboration across independent sources is what turns a claim into an entity.

llms.txt: useful, oversold#

I publish one. It is a short machine-readable index of the site — canonical pages, roles, projects, topics, feeds.

I treat it as an additional discovery surface, not a foundation. It is not a ranking mechanism, adoption is inconsistent, and a version that duplicates the whole site is worse than none. Short and accurate, or skip it.

What I will not claim#

That any of this guarantees citations. It does not. These systems change frequently, nobody outside them knows how sources are selected, and the confident numbers circulating on this topic are mostly invented.

What I will say is that all of it is defensible on other grounds. Crawlable content, coherent identity and specific writing are worth doing whether or not a model ever quotes you — which is the right test for any advice in this area.