GEO audit checklist: what to check before expecting AI citations

Five checks. robots.txt has no group blocking OAI-SearchBot, ClaudeBot, PerplexityBot, or Bingbot on the pages you want cited. The CDN or WAF is not blocking AI user agents; Cloudflare in particular can block them by default. Fetching a key page with each AI user agent returns 200 and full HTML, not a challenge page. robots.txt itself returns 200 and parses. The pages are not behind authentication, geoblocking, or interstitials. SerionFlow's free AI crawler access checker and robots.txt checker cover the first four.

Five checks. Fetch the page with JavaScript disabled or with curl and count words; the number should be close to what a browser shows. The H1 and every H2 are present in the raw HTML. FAQ questions and answers are in the raw HTML, not injected on click. Tables render as table elements, not as images or client-rendered grids. The canonical, title, and meta description in the raw HTML match what the browser shows. A single-page app that renders client-side will fail all five, and no amount of content work fixes that until rendering does.

Six checks per priority page. The H1 is the prompt or the exact query. The first paragraph answers it directly in forty to sixty words with no preamble. There is a TL;DR or summary list whose items stand alone. At least one table carries real numbers. There is an honest section on limits or when the product is the wrong choice. There is an FAQ with verbatim questions and short answers. Pages that pass layers one and two but fail here are readable and still rarely cited, because engines prefer sources that state the answer plainly.

Five checks. JSON-LD validates and every marked-up element, especially FAQPage questions, is visible on the page. datePublished and dateModified reflect reality and the page has been updated within the last year for evergreen topics. The canonical is self-referencing. The page is in a sitemap that Search Console and Bing Webmaster Tools have accepted. An llms.txt exists at the root and lists the page if it is a key page. None of these earn citations on their own; wrong ones reduce trust.

Four checks. A fixed prompt set of twenty to fifty buyer questions exists. It has been run in ChatGPT with search, Perplexity, and Google AI Overviews with every cited URL logged. It is re-run monthly on the same prompts. Search Console impressions and positions for the corresponding conversational queries are tracked alongside. Without these, an audit has no baseline and no way to show that fixes worked. SerionFlow's free citation gap finder and GEO audit tool provide a starting structure; paid plans add sampled AI visibility monitoring and AI-related query signals.

Audit ten priority pages rather than the whole site; the pattern generalises. Record pass or fail per check, fix layer one and two failures site-wide first because they are usually one configuration change, then rewrite the priority pages against layer three, correct signals in layer four, and establish measurement in layer five. Re-audit quarterly. When SerionFlow published its own audit in September 2026, every page failed layer two because content rendered client-side, which explained zero AI citations more completely than any content finding; fixing the substrate came before writing a single new page.