answer capsule

A capsule has three parts and all three are required. A framing that establishes the question, either as the subhead above it or as the opening clause. An answer stated immediately, in the first sentence, with the subject named rather than implied. And two to four sentences of support that make the claim defensible without sending the reader anywhere else. Strip any one of the three and the paragraph stops travelling: it either arrives without a question to attach to, or arrives as an assertion nobody can evaluate.

The concept it is constantly confused with is the featured snippet, and the difference is which side of the system you are on. A featured snippet is an output — Google chooses a passage from a ranked page and displays it in a box, and you influence but never author it. A capsule is an input: a unit you write deliberately, before any engine has an opinion, designed so that whatever selects it has something clean to select. One is a result you hope for; the other is a construction technique.

It is also not a summary or a TL;DR. A summary compresses a document and refers back to it — remove the document and the summary is a set of headlines. A capsule is complete at its own scale and has no parent it depends on, which is why a page carries several capsules and only one summary. The practical difference shows up in the first word: summaries tend to open with "this article" or "in short", both of which are dangling references the moment the text is extracted.

The unit of measurement is pass rate: what share of a page's paragraphs survive the self-containment test. The test has three parts and is deliberately mechanical. Does the first sentence name its subject rather than opening with a pronoun or a connective? Does the paragraph state its claim before it explains it? Can a reader who has seen nothing else on the page evaluate what is being asserted? Three yeses is a capsule; one no is a near-miss, and near-misses are usually one sentence away from passing.

Length is the other lever and it is narrower than people expect. Forty to a hundred and twenty words is the range that works — the same window that FAQ answers land in, for the same reason. Below forty there is no room for support, so the claim arrives naked. Above a hundred and twenty the paragraph almost always contains a second claim, and extraction will cut it somewhere in the middle. Capsule quality feeds the content-extractability pillar, which carries nineteen of the rubric's forty-seven checks.

Here is a paragraph as it typically ships. It is accurate, it is well written, and it is unquotable, because every referring expression in it points somewhere outside itself.

This also means it can take a while to show up. As we mentioned above, the various platforms all work differently, so you may not see the effect straight away even after doing everything right.

Count the failures. "This" refers to a previous paragraph. "It" has no antecedent inside the span. "As we mentioned above" is an explicit dependency on text that will not travel with it. "The various platforms" names nothing. Extracted, the passage asserts that something unspecified takes an unspecified time for unspecified reasons — which is why a retrieval layer scoring candidate spans will rank it below almost anything else on the page.

Citation gains from a content fix appear at different speeds on each answer engine. Every platform runs its own retrieval layer with its own crawl and refresh schedule, so a rewritten page can be cited on one engine while another has not re-fetched it. Expect days rather than hours, and compare a seven-day trailing window rather than two consecutive readings.

Same claim, same length, three changes. The subject is named in the first four words. The claim precedes its explanation. Every referring expression resolves inside the span. Nothing was added except specificity, and the paragraph now answers a real question — how long until a content fix shows up — completely enough to be quoted without the page around it.

• The dangling pronoun. The paragraph opens with "it", "this", "they" or "that approach", all of which resolve perfectly for a reader going top to bottom and resolve to nothing for a system that starts here. It is the most common failure by a wide margin and the cheapest to fix: replace the opening pronoun with the noun it stands for, every time, even where it feels repetitive to a linear reader. Repetition is the price of portability.
• The buried lede. The paragraph spends three sentences on setup, context and caveats before arriving at the claim in its final line. A reader tolerates this; extraction does not, because a span truncated at 400 tokens cuts before the payload and what escapes is preamble. The fix is inversion — put the claim in the first sentence and demote the setup to support, which also makes the paragraph better for the humans who were skimming it anyway.
• The context tax. Someone learns the lesson too well and starts restating the same three sentences of background at the top of every paragraph, so the page becomes a sequence of near-duplicates that differ only in their last line. Models see several highly similar spans and pick arbitrarily between them, and readers quit. The fix is to give each capsule its own distinct claim and let the shared context live once in the section intro, where it belongs.

The dependency chain running through that table is worth stating plainly, because teams often start at the wrong end. Capsules come first: they are prose work, and no markup rescues a paragraph that cannot stand alone. FAQPage schema is the declaration layer on top, useful precisely because it tells a parser which spans are already question-shaped — but wrapping a context-dependent paragraph in Question and Answer nodes only labels the problem. Density is the aggregate you read afterwards to know whether the page as a whole has enough of them.