How Long Should a Direct Answer Be?

Every guide to answer-first writing hands you a word count, and almost none of them says where the number came from. This page says where ours came from, and what it is not.

We write the answer block at 40 to 75 words, and a build gate on this library rejects any page outside those bounds [our data]. That is a house specification chosen for a property we can reason about, not a window measured against citation outcomes — ours or anyone's.

Is there a published length that AI engines prefer?

No. No source on our closed list publishes a preferred passage length for AI citation, and we have found none that measures one.

Google's AI features documentation puts the mechanical bar at being indexed and snippet-eligible, with no length target attached. The one length lever it does document runs the other way: max-snippet:[number] is a cap you impose on how much of your page may be used, described in the robots meta tag reference as applying across search surfaces including AI Overviews and AI Mode. A ceiling you set is not a preference the engine published.

So treat any precise "citation window" you are offered — including one this page carried in an earlier draft — as somebody's convention until its measurement is shown. Ours is a convention too. The difference we can offer is saying so.

Why 40 to 75 words?

Because those bounds force self-containedness, which is the property that actually decides whether a passage survives being lifted. Under roughly 40 words, a block rarely manages to name its own subject, answer the question, and carry a concrete figure. Over roughly 75, it almost always contains a second idea — and a second idea is where a retrieval system cuts.

The bound also does something a looser rule cannot: it fails. A style guideline erodes quietly across a few hundred pages and a few dozen writers; a numeric bound in a build gate does not.

Set the evidence tier honestly. This is a reasoned convention, enforced consistently, on a library whose citation outcomes we cannot attribute to it. We have not run an A/B test on answer-block length, we have no way to run one that would isolate length from everything else a page changes, and no closed-list study we could borrow one from.

What does the gate actually check?

Three things, on every page, before the site can build [our data]:

The digit rule deserves its own note, because it is the one people misread. We require a figure so the claim is checkable and so the writer has to have a source — not because we have evidence that numbers get quoted more often. Adjacent evidence exists and points that way: the Princeton GEO study (KDD 2024) measured benchmark visibility gains up to about 40% from passage-level edits that added citations, quotations, and statistics. That is a lab result on the paper's own corpus about content, not a finding about word counts, and we do not stretch it into one.

The bound is also useless on its own. A 62-word block that leans on the sentence above it fails at the only job it has, and passes the gate — which is why the count is the last thing we check, not the first.

What makes a block fail even at the right length?

Dependence on the text around it. Here is the same answer written twice, at the same length, as an illustration rather than a measurement:

Fails: "This is usually the fastest option, and it costs about half as much as the alternative, which is why most of our builds start here rather than with the heavier approach described above."

Passes: "A build gate is a check that runs before a site compiles and stops the build when a page breaks a rule. On this library, 3 rules are enforced that way: answer-block length, a digit inside the block, and 3 to 6 takeaways [our data]."

The failing version is grammatical, roughly the right size, and worthless out of context. It never names its subject, its comparison points at something the reader cannot see, and the phrase "described above" is a dangling reference an engine carries into the index. The passing version names what it is defining in its first five words. The mechanics of why passages get separated from their pages in the first place are at chunking, and the property itself has its own definition at citable passage.

Where does the block sit, and what else on the page is liftable?

Directly under the H1, before any narrative, followed by 3 to 6 standalone takeaways. Those takeaways are the second liftable unit our contract enforces — not a longer "citation block" with a word count attached, because we have never measured one and will not publish a range we invented.

Everything below that is ordinary page order: question-shaped H2s answered in their first sentence, short paragraphs, at least one real table. The full layout, and the reasoning behind each position in it, is at how to structure a page so an LLM can quote it — that page owns page order, and this one owns the block's size.

When do the bounds not apply?

When the page is not answering a question. Our own glossary pages open with a definition shorter than the 40-word floor would allow in prose, procedural pages lead with an ordered list, and a page whose value is a comparison table should put the table first. The property to preserve is self-containedness; the word count is one implementation of it.

Against our own interest, as people who sell answer-first builds: no answer-block length gets a page cited. Length is a formatting decision, and formatting decisions are the cheapest, most copyable part of this discipline — everyone will have them. What is not cheap is having something to say in the block that nobody else can source, which is the argument in what the schema evidence actually shows applied to structure generally: the markup and the shape are table stakes, and the evidence is the differentiator. The wider system this fits into is our operator's guide to generative engine optimization.