Before any question of content quality or authority, there is a mechanical gate: can answer engines reach your pages, parse them, and hold their facts with confidence? If the answer at any step is no, everything downstream - the liftable passages, the earned mentions, the whole generative engine optimization program - is decorating a page the machines never see. This checklist is that gate, in runnable order, with the reasoning behind each check so your team can argue about priorities from understanding rather than superstition.
Why technical comes first
Recall the retrieval pipeline behind AI citations: query formulation, retrieval, selection, synthesis. Technical citability is the retrieval stage's veto. It is also the highest-leverage work in the discipline per hour spent, because it is finite - a site clears it once and coasts on maintenance - and because failures here are silent. Nothing tells you an AI crawler was turned away; you simply never appear, and teams spend quarters on content strategy while a one-line robots rule undoes all of it.
Check group one: crawler access
AI systems fetch with their own user agents - OpenAI, Anthropic, Perplexity and Google's AI fetchers each identify themselves - and robots.txt rules written years ago frequently block them by accident: a blanket disallow inherited from staging, a security plugin's bot list, a CDN rule matching "GPT" as a threat signature. The checks: read your robots.txt against the current list of AI user agents and allow the ones you want citing you; confirm your CDN and WAF are not challenge-screening them (a JavaScript challenge page is an empty page to a fetcher); and verify rate limits are not so aggressive that fetchers give up mid-crawl.
Check group two: rendering and readability
Retrieval fetchers are not patient browsers. Content that exists only after client-side rendering, behind interaction, or inside canvas is invisible or degraded to them. The standard: your substantive content should be present in the initial HTML response - server-rendered or statically generated - with semantic structure (headings that are headings, lists that are lists, tables that are tables) rather than styled divs impersonating them. The quick test is brutal and honest: fetch your page with a plain HTTP client and read what came back. If the answer to "where is the article?" is "it loads later", you have found your project.
Check group three: llms.txt
The llms.txt proposal gives language-model systems a curated, markdown map of a site: what it is, and which pages carry its most citable content. It costs an afternoon, it is increasingly fetched in the wild, and it solves a real problem for the engines - separating your load-bearing pages from your pagination. Keep it short and honest: the pages you would want quoted, described in one line each. Regenerate it when major content ships; a stale llms.txt pointing at retired URLs is worse than none.
Check group four: structured data
Schema.org markup converts prose into typed facts. The set that earns its keep for GEO: Organization (who you are - the entity anchor everything else hangs off), Article with honest dates (the freshness signal), FAQPage (question-answer pairs in exactly the shape answer engines consume), and HowTo where content is procedural. The discipline that matters more than coverage: the markup must match the visible page. Engines cross-check, and markup that diverges from the rendered content erodes exactly the machine confidence it was meant to build.
Check group five: freshness and entity consistency
When candidate sources disagree, observable tie-breakers are freshness and consistency. Freshness: real published and modified dates, in markup and on the page, updated when content actually changes - engines visibly prefer current sources on time-sensitive questions, and fake date-bumping is both detectable and counterproductive. Entity consistency: your company's name, description and category should read identically across your site, your directories and your profiles - contradictions dilute the machine's confidence about who you are, which quietly suppresses citation. The audit is tedious exactly once; after that it is a diff.
The working checklist
The interactive version below persists in your browser - work through it across as many sessions as the fixes take.
Auditing it continuously
Every check above regresses the way infrastructure always regresses: a deploy rewrites robots, a redesign moves content client-side, a CMS update drops the schema component. The countermeasure is the same as everywhere in this blog - a standing mission. The Audit agent runs exactly this pass on a schedule (the AI visibility audit playbook bundles it with the measurement layer), and the monthly diff - "what broke since last month" - is one of the highest signal-per-word reports in the entire program. However you run it: the checklist is not a project, it is a property to maintain.
Frequently asked questions
What is technical GEO?
The mechanical layer of generative engine optimization: ensuring AI crawlers can access your site, content survives plain-HTML fetching, llms.txt maps your citable pages, structured data types your facts, and freshness plus entity consistency hold up.
Do AI crawlers respect robots.txt?
The major ones identify themselves with distinct user agents and honor robots directives - which cuts both ways: a stray disallow silently removes you from AI answers. Auditing robots against the current AI agent list is the first check.
Is llms.txt actually used?
It is a proposal with growing real-world fetch traffic, it costs an afternoon, and it solves a genuine problem for engines - separating load-bearing pages from noise. Cheap insurance with plausible upside; keep it short and current.
Which schema types matter most for AI visibility?
Organization (entity anchor), Article with honest dates (freshness), FAQPage (answer-shaped pairs), and HowTo for procedures - always matching the visible page, because divergent markup erodes machine confidence.
How often should the technical pass re-run?
Monthly, and after every major deploy: regressions ship with releases. As a scheduled agent mission the re-audit costs minutes and the diff report catches breaks while they are hours old.
Sources
- llms.txt - the specification and rationale
- Google Search Central - robots.txt and crawler documentation
- Schema.org - structured data vocabulary
Every playbook on this blog ships as a runnable mission.
Open a workspace and the playbook library is waiting - describe the outcome and the agents carry it end to end, on your plan's monthly credits.