Answer engines6 min read23 September 2026

Writing passages an answer engine can lift

An answer engine does not extract your page. It extracts one passage from it. That changes what good writing means at the paragraph level: each passage has to survive being removed from its surroundings and still answer the question on its own.

Part ofAnswer Engine Optimization (AEO)

Traditional search optimisation treats the page as the unit. Answer extraction treats the passage as the unit — a paragraph, a list, a table row — and lifts it away from everything around it. A passage that reads perfectly in context and incoherently alone will not be chosen, no matter how well the page ranks.

Lead with the answer, then earn the rest

The most common structural problem is the build-up. A section opens with context, narrows through qualification, and states the answer in the fourth sentence. Extraction takes the first sentences, which means it takes the context and not the answer.

Build-up (not extractable)	Answer-first (extractable)
Many factors affect how long this takes, and it depends on your setup…	Most migrations take four to six weeks. The variables that move that number are…
Before we can discuss pricing, it is worth understanding the model…	Pricing is per environment, billed monthly. The model works like this…
This is a question we hear often, and the answer is nuanced…	Usually no. There are two cases where it is worth doing anyway…

Journalists have had a name for this for a century. The inverted pyramid — conclusion first, detail after, background last — was built for readers who stop early, and it happens to be exactly the shape an extraction system rewards.

Self-contained means no pointing outside the passage

A passage that refers to things outside itself breaks when extracted. The references that cause this are small and easy to miss because they read naturally in place.

Breaks extraction	Survives it
As mentioned above, this only applies when…	Row-level security only applies when the connection sets a tenant identifier…
It is usually slower than the alternative.	Fine-tuning is usually slower to iterate on than retrieval.
This means you should avoid it in production.	Long-lived stateless tokens should be avoided in production.
The second option is normally correct.	Tag-based invalidation is normally the correct choice.

Unresolved pronouns — it, this, they, that approach. Name the subject again; the mild repetition costs a reader nothing
Positional references — above, below, the previous section, as we saw. Meaningless once the passage is alone
Ordinals from a list — the second one, the last of these. Restate the thing
Carried qualification — a caveat stated two paragraphs earlier is absent from the extracted passage, and the answer is now wrong

Length: enough to be complete, short enough to be lifted

There is no official limit, and chasing a specific word count is a misreading of the mechanism. What matters is that a complete answer exists inside a span short enough to be shown. In practice that means the direct answer occupies roughly two to four sentences, with elaboration in the paragraphs after it rather than inside the same one.

text

## How long does a site migration take?          <- the question, as asked

Most site migrations take four to six weeks from URL         <- the answer,
mapping to post-launch monitoring. Sites with fewer than      complete and
a thousand URLs and no platform change can be done in two.    self-contained

Three things extend it: a change of content management        <- elaboration,
system, a change of URL structure, and any internationalised  in its own
variant, since each locale needs its own mapping and its own  paragraph
verification pass.

The monitoring period is not padding. Ranking changes after
a migration surface over two to three weeks, so a migration
closed at launch is one where nobody looked for the damage.

Definitional sentences, written on purpose

A large share of answer-engine queries are definitional — what something is, what it means, what it is for. These are answered by a single sentence pattern, and pages routinely fail to contain one because the author assumed the reader already knew.

X is a Y that Z — the classic definitional shape, and the one extraction systems recognise most readily
State it plainly before qualifying it; a definition wrapped in hedging is not extractable
Define the term the audience would search for, not only your internal name for it
Put the definition near the first use, not in a glossary at the end

Lists and tables are extracted as units

Where the answer is genuinely a set of items or a comparison, the corresponding structure is extracted more reliably than the same content in prose — and it is also more useful to a reader. The requirements are ordinary.

Introduce the list with a sentence stating what the items are, so the extracted unit has a subject.
Keep items parallel in grammar and comparable in length; a list where one item is a paragraph reads as broken.
Use an ordered list only for a real sequence. Numbering an unordered set implies a priority you did not mean.
Give tables a header row with meaningful labels, since the labels are what make the cells interpretable alone.
Keep tables narrow. A ten-column table is unreadable on a phone and unusable as an extracted answer.

What not to do

Do not write the same answer six times across a page in the hope of increasing the chance of extraction. Repetition reads as padding to a person and adds nothing for a machine
Do not ask a question in a heading and answer it three paragraphs later. The heading sets an expectation extraction takes literally
Do not bury the answer behind a preamble about why the question is interesting. Nobody searching for it needs convincing that it matters
Do not state an unqualified answer you would not defend. Being cited for a claim you hedged in person is worse than not being cited

The unifying idea is unglamorous: passages that stand alone are better writing. Answer engines did not create the requirement to say what you mean early and completely — they made it measurable.

Questions

How long should an extractable answer be?

Long enough to be complete and short enough to be shown — in practice two to four sentences for the direct answer, with elaboration in the paragraphs after it. Chasing a specific word count misreads the mechanism: what matters is that a complete, self-contained answer exists within a short span, not that the span hits a target length.

Why does my page rank well but never get extracted?

Most often because the answer is not in the first sentences of the relevant section, or because the passage depends on its surroundings. Check for unresolved pronouns, positional references like "as mentioned above", and qualification stated in an earlier paragraph. A passage that reads perfectly in place and incoherently alone will not be chosen.

Should I convert my content into lists and tables?

Only where the content genuinely is a set of items or a comparison. Lists and tables are extracted reliably as units and are easier to read for that kind of material, but argument, causation and nuance belong in prose. A page of bullet fragments answers simple questions while demonstrating no depth on the ones that matter.

Does repeating the answer several times help?

No. It reads as padding to a human and adds nothing for an extraction system, which is selecting the clearest passage rather than counting how often a phrase appears. One well-placed, self-contained, unambiguous answer outperforms six approximate restatements.