Answer Extractability

When search returned a list of links, the goal was to get the whole page read. Someone clicked a headline and read top to bottom, so a third paragraph could lean on the two above it and still make sense. A paragraph opening with "the second of the two options above applies here" was fine for anyone reading in order. Search engine optimization (SEO) — tuning a site so it surfaces higher in search results — only had to win the click.

AI overviews changed who reads first. An AI overview is the summarized answer pinned above the search results, with a few source sites listed underneath it. A model does not read a page in order. It picks the one chunk that matches the question and lifts it out, and if the lifted chunk doesn't stand up on its own, it drops that document and takes another.

So a well-written page can go unquoted. Pagedeck's help doc for "change document permissions" always landed on the first page of results, but ask a model the same question and some other company's doc showed up as the source. The paragraph at fault began with "in this case, use the method described above." Natural enough for a person, and no answer at all once it is on its own.

A model doesn't use a document whole. It splits the page at headings and paragraph breaks, then pulls the pieces that match the question and assembles an answer. Extractability asks whether a piece is still an answer after it is pulled.

Three things have to sit inside the paragraph itself.

1. The subject. What the paragraph is about is named in it. "This feature can only be enabled by an admin" requires the paragraph above to know which feature; "Share link expiry can only be enabled by an admin" is already an answer.
2. The conditions. When it applies travels with it. "Expiry caps at 30 days on the free plan and one year on paid plans" keeps the scope in the same chunk instead of the next one.
3. The conclusion first. The opening sentence is the answer. Lay down two sentences of background and answer in the third, and anything lifted from the front is an introduction, not an answer.

Pronouns break all three at once. Words like "this," "that," "the above," and "as mentioned earlier" only work while the previous paragraph is still next to them. More than two of them in one paragraph usually means that paragraph cannot be lifted.

Formatting decides where the cuts fall. In two thousand words without a heading, the model picks the split points itself, and a split through the middle of a paragraph leaves half an answer. Put the question in the heading and the answer in the first paragraph under it, and the cut line matches the answer boundary. Tables and numbered lists do the same job, because a block with visible edges gets taken whole. The practical version of all this is a length that fits one idea: a paragraph covering one question, its conditions, and its exception, usually three to six sentences, with the next question starting a new block under its own heading rather than continuing the same run of text.

The pages worth fixing first are the ones people search in question form.

1. Help docs. "How do I" becomes the search query verbatim here, so this is where the work starts. Procedures state the outcome in the first sentence and list the steps after it.
2. Pricing pages. Questions like "how much is it" and "how many seats on the free plan" collect here. Put the numbers only in a table and never in a sentence, and a chunk that missed the table carries no answer.
3. Feature pages. Comparisons against other products get quoted often. Instead of claiming one is better, finish the difference inside one paragraph.

Checking is manual. Copy a paragraph into an empty document and read it: if it answers, it passes; if you go looking for what came before, it needs work. Markup like structured data, which labels which block is a question and which is an answer, sharpens the cut lines, but the sentence comes before the label. Tag a paragraph that can't stand alone and the model just picks up a useless chunk more reliably.

"This paragraph isn't extractable. You have to read the one above it to know which feature it's about."

A request to strip the pronouns and put the subject back inside the paragraph.

"Let's make the heading a question and answer it in the first line underneath."

Line up where the model cuts with where the answer sits.

"Rankings are flat but nothing's getting cited. Start with extractability."

The page surfaces but never gets excerpted, so the structure is the suspect.

At Pagedeck, Ravi spent a week pasting paragraphs from sixty help docs into an empty document one at a time. The failures clustered in one shape: every time a screen was renamed, the doc said "the new name" and left what it was the new name of in the paragraph above. Fewer than forty paragraphs across sixty docs got rewritten, and two months later the number of AI overview citations had tripled.

Ravi keeps this separate from ranking work because the two touch different things. Ranking is about which page to write. Extractability is about where to cut the page that already exists.

The most common assumption is that shorter paragraphs are more extractable. The quoted chunk is short, so the paragraph feels like it should be too, and short paragraphs do get picked up often enough to reinforce it. But the test is self-sufficiency, not length. "No, that isn't supported" is one line and still not an answer, because nothing in it says what isn't supported. A five-line paragraph carrying the conditions and the exceptions is long and already an answer. Chase length and the doc ends up in fragments, and fragments leave their subject behind in the paragraph above — extractability drops.

The other one is that this trades human readers for machines. Naming the subject again in every paragraph and leading with the conclusion sounds like it will stiffen the prose. In practice most people arrive at a help doc from search, read one section, and leave, so a paragraph that stands alone serves them too. Skip the work over that worry and you end up with a document no model quotes and no visitor can read without scrolling back up.