Writing for Answer Engines

For a long time, writing for search meant writing for a crawler: put the right words on the page, mark up the headings, earn a link, and hope the ranking followed. That has not stopped being true. What has changed is what happens after the page is found, because the page is increasingly read by something that summarises it before a person ever arrives.

How the reading changed

Three things are now ordinary rather than experimental. Search results show an answer, sometimes with sources, before the list of links. Assistants answer a question in one or two paragraphs and then name where the answer came from. Some systems read a page to check a single fact rather than to understand the whole thing.

The consequence is straightforward. A page can now be evaluated on a paragraph rather than judged as a whole. If the one paragraph that answers the question is buried, hedged, or worded so that it cannot be lifted out cleanly, the page can end up being read as if it does not contain the answer at all.

This is not an argument for writing robot text. It is an argument for the same thing good writing has always needed: put the answer where it can be found, and make it possible to repeat without it becoming wrong.

What makes a paragraph quotable

A paragraph survives being quoted when it works on its own. The test is simple: copy one paragraph out of the page, paste it into a blank document, and read it alone. If it still answers the question it came from, it is quotable. If it only makes sense because of the paragraph before it, it is not.

Four habits make the difference:

Answer in the first sentence. Say what the section is about before the preamble. A reader who stops after one sentence should still have the answer.
Keep the subject in the sentence. A paragraph about response times should say response times, not they. Pronouns are fine in a sentence and expensive in a quotation.
Prefer the concrete noun. Tools, roles, and actions can be named. Vague phrases like end-to-end solutions or a range of solutions cannot be quoted because they are not verifiable.
Keep one idea per paragraph. A paragraph that answers two questions gets quoted for one of them and loses the other.

There is a further subtlety worth naming. When an assistant quotes a page, it is usually summarising across sources, and it has to decide which statement is the page’s position. If two paragraphs on the same site say slightly different things, the safe reading is that the site is uncertain. Being quotable therefore means being consistent, not just being clear.

Structure for a machine

Headings are the part of a page most likely to be read literally, because they describe what follows. A heading that says what the section is about makes the page skimmable for a person and legible for a system that is assembling an answer from several sources.

One question per heading. The heading should match the question a person would actually type. A heading that answers a question is easier to quote than one that describes a topic.
Keep the hierarchy honest. A page with one h1, sections as h2, and subsections as h3 tells a reader where they are. Flattened or skipped levels cost both readers and parsers.
Answer the obvious question early. If a heading raises a question, the paragraph under it should deal with it before moving on. Leaving the answer further down the page is how the good sentence gets missed.
Give questions a home. If the same question comes up repeatedly, give it a section or a question-and-answer block rather than scattering the answer across three pages.

One habit is worth singling out: a question-and-answer block on the page, with the answer written in full sentences. It reads slightly more formally than the rest of a site, and that is the point. The answer can be lifted out without editing, which is exactly what an answer engine wants.

Consistency beats cleverness

The most common reason a site gets quoted inaccurately is not that it is unclear. It is that it says the same thing in three different ways, and the summariser picks the wrong one. A service described as a dashboard on one page, an internal tool on another, and a reporting system on a third is three claims, and the least-supported one can win.

Consistency is mostly editorial discipline rather than technical work:

Decide the name for each thing you offer, and use it everywhere, including in titles, headings, and image alt text.
State facts the same way each time. If the answer to a question is the same on five pages, it should read the same on five pages.
Do not contradict yourself quietly when something changes. Update the old page, or remove the claim.
Keep the wording in your structured data identical to the wording on the page. Two versions of the same fact is one fact too many.

The technical layer

None of the above replaces the technical work. It makes the work usable by a machine that is doing something simpler than a reader.

Crawlable: Allow the crawlers you actually want. The common ones identify themselves clearly, and blocking them removes you from results without a warning.
Described: Structured data that matches the visible page, so a parser can tell what the page is about without inferring it.
Listed: A complete sitemap, with every page that exists in it, so nothing is left to chance.
Summarised: A short plain-text summary of the site, for systems that prefer a plain description over parsing pages.

None of these should change what the page says. They change how reliably the page can be understood and credited.

A short checklist

Before calling a page ready to be read by machines as well as people:

Does the first sentence under each heading answer the question that heading raises?
Can any single paragraph be lifted out and still stand up?
Does the page name things the same way the rest of the site names them?
Is there exactly one h1, and is the heading order sensible?
Do the structured data and the visible text say the same thing?
Are the facts repeated across your pages consistent, down to the wording?
Does the sitemap list every page that should be found?
Is the page readable by a person with JavaScript turned off, or with the images missing?

The test I keep coming back to is simple: could a careful assistant quote this page and get it right? If yes, the page is doing its job, and most of the remaining work is down to patience and measurement rather than more words.