When an AI assistant answers a question by citing a website, it lifts sentences from a handful of pages. Whether a given page can actually be lifted that way isn't a matter of luck; it's a measurable property of the underlying HTML, according to a recent audit walkthrough shared by an independent web developer.
The Homepage Scored 65 Out of 100
Running a custom extractability audit against his own site, the developer found his homepage scored 65 out of 100. The cause traced to five UI card components that rendered their titles as heading tags rather than plain text. Demoting those six invisible headings to ARIA-preserving paragraphs, without changing a single visible pixel or removing a word of content, brought the score to a perfect 100.
Three Stages Before a Citation Happens
The audit frames AI citation as a three-stage pipeline: a page first has to be retrieved by the engine, then extracted into a clean, self-contained answer block, and finally attributed confidently enough for the engine to name the source. Most AI-visibility advice focuses on retrieval (robots.txt, sitemaps) and attribution (schema markup, entity signals). Extraction, the middle stage, gets far less attention despite failing silently: a page that retrieves and attributes fine but extracts poorly simply never surfaces in answers, with no signal pointing to why.
What the Audit Actually Measures
The scoring method checks five properties of a page's rendered HTML on a 0 to 100 scale, landing pages in an "extractable," "partially extractable," or "not extractable" band. The critical distinction driving the six invisible failures wasn't the headings the developer wrote in his CMS, it was the headings his component library silently emitted underneath them.
Component Libraries Quietly Rewrite the Outline
A card title displaying a blog post's headline looks like a heading and gets marked up as one by many component libraries, following accessibility guidance that favors semantic HTML. But a card title is really a label pointing to a different document, not a section heading of the current page; the page's actual outline is something closer to "here are my featured posts." Because HTML has no dedicated tag for "title of another document," components default to heading tags, and every parser that walks the page inherits a false outline as a result. Since cards repeat, one index page rendering ten card instances injects ten extra false headings into that page's structural census, diluting the four or five headings the developer had actually written and tuned.
A Fix That Keeps Accessibility Intact
Rather than stripping the semantic markup entirely, which would remove a genuinely useful navigation landmark for screen reader users scanning a list of posts, the fix replaces the literal heading tag with a paragraph carrying an ARIA heading role and an explicit level. Screen readers that navigate by heading still stop at the same point and announce the same level, while extraction parsers, which chunk pages by literal heading tags, no longer count the card title as part of the page's outline. The change removed zero words of content.
Scores Drift, So the Fix Needs a Gate
After the fix, the homepage's score rose to 100, though a later redesign nudged it back down to 80 while staying comfortably inside the extractable band. That drift is the reason the developer built a CI gate into his deployment pipeline, re-running the audit against production URLs on every deploy and failing the build if any page drops below a threshold of 75, catching regressions from future component changes before they go unnoticed.
A Correlated, Not Fully Isolated, Result
Over the roughly three months following the fix, the site's reported AI citations from Microsoft Copilot and partner assistants grew substantially, rising from a few hundred to roughly 1,500 by the audit's account. The developer is careful to describe that growth as correlated with the extraction work rather than cleanly attributable to it, since other changes, including new content and continued search optimization, happened over the same window.