The technical layer: what matters and what is theater
static HTML, fast pages and visible answers move AI citations. schema and llms.txt mostly do not. what to build and what to skip.
last updated 2 November 2026
most of the technical GEO checklist is theater. what actually matters for AI citations: static readable HTML, fast pages, structured headings, and the answer sitting in visible text. what mostly does not: schema markup and llms.txt. keep schema anyway for Google rich results, then spend the hours you saved writing better answers, because that is where the citations come from.
What actually matters
an AI crawler is a fast, impatient reader that executes nothing. serve it static HTML and it reads everything; serve it a client-rendered app and it reads a loading spinner. speed matters because crawl budgets and timeouts are real, and a slow page gets sampled or skipped. heading structure matters because retrieval works on chunks, and clean h2 and h3 boundaries are what your page gets chunked along. and the answer has to sit in visible text, in the actual paragraphs, because visible text is the only thing the whole pipeline agrees on reading.
none of this is exotic. it is the boring 2010s checklist, back with a new reader. if your product pages need JavaScript to show their content, that is the technical problem worth solving, ahead of any of the fashionable ones.
two quick self-audits settle where you stand. fetch your own page with a plain curl and read what comes back: if the answer text is in that response, crawlers have it, and if the response is a script shell, they do not. then check your server logs for the AI user agents, GPTBot, ClaudeBot, PerplexityBot, and confirm they are actually being served pages and not blocked by your robots.txt or your CDN’s bot protection. plenty of sites fail GEO right there, silently, with a firewall rule someone set years ago.
What is mostly theater
schema markup does not measurably move AI citations. an Ahrefs controlled study in May 2026 found no meaningful citation lift from schema, which matches the plumbing: assistants quote what a page says, not what its metadata declares. llms.txt is further gone. Google’s John Mueller said in April 2025 that no AI service actually reads it, and nothing since has contradicted him. it is a proposal, not a protocol.
both keep getting sold as GEO essentials because both are easy to sell: a plugin, a generator, a deliverable that takes an afternoon and produces a diff. an honest audit books them at what they are worth for AI citations, which is approximately nothing. the tell is always the same: the deliverable changes something no reader ever sees. citation systems quote pages the way readers read them, so any fix that leaves the visible page untouched should be assumed inert until someone shows a controlled test saying otherwise.
| item | effect on AI citations | keep it? |
|---|---|---|
| static readable HTML | high | yes |
| page speed | high | yes |
| clean heading structure | high | yes |
| answer in visible text | highest | yes |
| schema markup | none measured, per Ahrefs, May 2026 | yes, for Google rich results |
| llms.txt | none, no service reads it, per John Mueller, April 2025 | optional, it costs nothing |
Keep schema anyway
schema still earns its place, just not in this column. Google rich results remain real, and classic search still sends real users, so marking up articles, FAQs and products stays worth the small effort. the point is accounting honesty: file schema under SEO, where it works, not under GEO, where the measured effect is nothing. what breaks budgets is not doing schema, it is doing schema instead of writing the visible answers.
How this site runs it
this site is my own live test of exactly this shortlist: static HTML with no client framework, AI crawlers explicitly allowed in robots.txt, and every page written answer-first. the full setup is public in how i run this site. it is early and i am not claiming results yet, but the technical layer took days, not months, which is the real point: done right, this layer is cheap, and everything expensive in GEO is the writing.
the technical layer is one piece of the AI search playbook. the breakdown call is free. Let’s talk Growth.
this page is part of the AI search & GEO playbook.
More in this playbook
How AI assistants actually pick products
ChatGPT recommends a tiny fraction of the products it knows. what separates the named from the ignored, and why it differs per assistant.
AI search & GEOWriting answer-first content that gets quoted
where AI assistants pull quotes from on a page, and how to write the 40 to 60 word answers that get cited instead of skimmed past.
AI search & GEOReddit's role in AI answers, honestly
Reddit shows up in roughly 40% of AI citations and the share can drop overnight. how to use the channel without building on sand.
AI search & GEOListicles and comparison pages AI loves to cite
why lists and comparison tables get cited so often, how to write an honest one, and how to get into other people's.
AI search & GEOMeasuring your AI share of voice
a free monthly prompt panel that shows which assistants name your product, which name competitors, and which sources they read.
AI search & GEOAI directories: six good ones beat sixty
a handful of real AI directories beat mass submission. what a listing actually does for AI visibility and how to submit cleanly.
AI search & GEOThe honest GEO timeline
expect 8 to 16 weeks before AI assistants start naming you. what compounds, what spikes, and why a young domain still has a shot.