Checklist3 sections · 2 min readBy Dezső Mező · Published October 4, 2026.mdAEOAI searchllms.txtSEO
AI search is not one channel but three behaviours: assistants that fetch pages live to answer a question (ChatGPT search, Perplexity, Claude), answer engines that compile citations from their index, and the SEO fundamentals that still decide what gets found at all. Each needs different plumbing, and most sites are set up for exactly one of them.
The honest framing before the list: nobody has proven which of these moves the needle and how much, because the systems change monthly and attribution is near-impossible. The items marked as proven also earn their keep on plain SEO and accessibility; the ones marked as bets are cheap enough to place anyway. That honesty is itself one of the items — an answer engine asked 'is this accurate' rewards content that says what it knows.
01Proven: the crawlable, quotable surface
Server-rendered content the crawler sees without executing JavaScript; a clean sitemap and robots.txt; definitional content where the answer is the first sentence (the part that gets quoted); structured data restating the page honestly (Organization, Article, FAQPage, DefinedTerm); real author and organization entities with sameAs links so the machine knows who is speaking.
02Working so far: the machine-readable extras
llms.txt and llms-full.txt as a site map for models; .md twins of key pages so a fetch gets markdown instead of a canvas; speakable markup marking the passage to read aloud; named AI-crawler rules in robots.txt documenting a deliberate allow. Early signs say the .md twins get fetched by assistant crawlers — the rest is inexpensive enough to keep.
03Unproven bets, cheap enough to place
Directly answering the phrasing of questions people ask assistants; keeping facts consistent everywhere the entity appears (Knowledge Graph work by another name); fresh content so crawlers re-check often; and content so quotable that being cited is the default outcome. None of these has a guaranteed return — all of them cost less than a month of ads.
What to take away
- AI search = live-fetch assistants + compiled answer engines + classic SEO plumbing.
- Proven: server-rendered quotable content, honest schema, real entities.
- Working: llms.txt, .md twins, named crawler rules — assistants do fetch markdown.
- Say what is proven and what is a bet — honest hedging is itself citable.
More from the lab
Browse all entries
Structured data an answer engine can use
What we did for the AI crawlers
One site, three languages, zero duplicate-content problems
Move the site, keep the rankings
Want this looked at on your own system?Start a conversation
