Answer Extractability

No search engine defines the term. It names a set of properties that the engines’ own guidance describes one by one. Google’s guide to generative AI features says there is no requirement to break content into tiny pieces for AI to understand it, because its systems can understand the nuance of multiple topics on a page and show the relevant piece to users^[1]. So extractability is not a format to impose on every page. It is whether the piece a system would show can stand by itself.

Microsoft’s guide to AI search answers is the nearest thing to a checklist. It lists concise answers of one to two sentences that directly address a question, and sentences that make sense even when pulled out of context, among the things that make content eligible for featured snippets, and it says assistants can often lift question and answer pairs word for word into AI-generated responses. It adds that important answers hidden in tabs or expandable menus may be skipped, because AI systems may not render hidden content^[2].

Featured snippets are the oldest form of the same idea: a box in which the descriptive snippet comes first. A page cannot mark itself as one. Google’s systems decide whether a page would make a good featured snippet for a search, and if so they elevate it^[3]. Extractability is therefore a condition that a passage can meet, not a way to be chosen.

Snippet controls can close a passage off, on purpose or by accident. Google’s documentation says nosnippet applies to all forms of search results, AI Overviews and AI Mode among them, and prevents the content from being used as a direct input for them; max-snippet limits how much of the content may be used as a direct input; and data-nosnippet leaves marked sections of text out^[4]. Google adds that, to be shown in its generative AI features, a page must be indexed and eligible to be shown with a snippet^[1].

A refusal in robots.txt closes off more than a passage, because a crawler that is turned away never reads the page. In QueryRev’s study, 896 of the 6,294 sites whose robots.txt could be read, 14.2%, block at least one AI crawler by name^[5]. Whether a passage can be lifted depends on whether the system that would lift it can fetch the page at all.

Read a passage the way a system meets it. Copy the sentence that answers a question out on its own and ask three things: does it still name its subject, does its first sentence give the answer, and does it carry its own condition or date? A passage that begins “It also depends” or “As above” fails the second test whatever its content.

Then check access, for the same passage. It has to be in the HTML a crawler receives, not behind a tab that needs a click or inside an image or a PDF^[2]. Its page has to be indexed and allowed a snippet^[1], and no nosnippet, max-snippet or data-nosnippet rule may cover it^[4].

Count across a set. For the questions that matter, record each passage as standing alone or not, in the text or not, and open to snippets or not. The share that passes all three is the extractability of that set. It is a count of your own pages, not a score from a search engine, and it says nothing about whether a passage is chosen.

Illustrative: an invented site, to show the method.

A home-improvement site checks five passages of its guide to loft insulation, one for each question the guide answers. Two open with the answer in a sentence of their own and sit in the page text. One starts “It also depends” and means nothing without the paragraph before it. One is a table inside a tab. One sits in a section that carries a nosnippet rule, added to the whole template years ago.

Two of five can be lifted as they stand. The fixes differ: rewrite the opening of the second, move the table out of the tab and into the page text, and take the rule off the section. The site reads the five passages again after the changes and records which moved. Whether any of them is then shown is for the search engine to decide.