Blog · Content that gets cited

Answer-first writing, and why the opening paragraph carries the answer

Published 26 June 2026 · Updated 12 October 2026 · 8 min read

Answer-first writing, and why the opening paragraph carries the answer
Photograph: Daniel Álvasd / unsplash

The short answer: Answer-first writing means the complete answer to one question sits in the opening paragraph, ahead of context, ahead of the brand story. The opening carries the weight because a retrieval system hands a model passages, not pages. The chunking defaults we can actually read, OpenAI's and Microsoft's, cut text into a few hundred tokens at a time, and an opening that clears its throat gives the engine nothing worth quoting.

That is the claim. What follows is the mechanism, the five ways an opening fails, and a test you can run on your own pages this afternoon.

Your page is not read as a page

Nobody publishes the exact indexing settings behind ChatGPT search, AI Mode or Perplexity. The public defaults in retrieval tooling are the closest honest proxy, and they agree with each other more than you would expect.

OpenAI's file search documentation sets a default chunk size of 800 tokens with 400 tokens of overlap, and returns ten results by default, with fifty as the ceiling. Microsoft's Azure AI Search guidance, last reviewed in June 2026, tells builders to start at 512 tokens with an overlap of 25 percent, which it puts at roughly 2,000 characters of text.

Do the arithmetic. An English token runs about three quarters of a word, so 512 tokens is near 380 words and 800 tokens is near 600. Eighty words is roughly 105 tokens. The opening of your page is therefore something like a fifth to a quarter of the first slice the system ever embeds, and on OpenAI's defaults that slice is competing against the nine other results the search returns alongside it. We cannot see how any consumer engine ranks those slices, so treat the arithmetic as the shape of the problem rather than a measurement of it.

The practical consequence is blunt. On chunk sizes like these, an answer held back until paragraph six lands in chunk two or three, and chunk two or three has to win retrieval on its own merits with no help from your headline, your navigation or the promise your title made.

Where a passage lands in the context appears to matter

Retrieval is only half of it. The chunks that survive get pasted into one long context and a model writes from that.

Liu and colleagues measured what happens next. Their 2024 paper in Transactions of the Association for Computational Linguistics tested multi-document question answering and key-value retrieval while moving the relevant passage around the context window. Performance was often highest when the useful material sat at the beginning or the end, and degraded significantly when the same material sat in the middle, including in models built for long contexts. That is a finding about the language models they tested on two controlled tasks in

  1. It is not a measurement of any shipping answer engine, and we have no way to confirm the

    2. same curve holds inside one.

Stack the two effects. Your sentence has to survive a cut it cannot see, then land somewhere in an assembled context where position appears to matter. Answer-first writing is not a stylistic preference. It is a cheap way to improve your odds at both steps, which is why it shows up in almost everything about what makes a page citable by an AI engine.

There is a third pressure, and one engine documents it. Google says AI Overviews and AI Mode may use a query fan-out technique, which it describes as issuing a set of related searches across subtopics and across different data sources to build one response. Whether other engines do the same is not published, so take the lesson from Google's case: a page whose opening answers exactly one phrasing qualifies for one of those searches, and a page whose opening answers the question and names its condition qualifies for more.

What actually belongs in the first eighty words

Four things, and none of them is a template to fill in.

The answer as a sentence, not a promise of one. "Most teams need between four and six weeks" is an answer. "The timeline depends on several factors" is a refusal dressed as one.

The condition under which the answer holds. A boundary you state is something a model can quote and attribute. A boundary you leave implicit is something it has to infer, and we cannot see what it infers.

One checkable specific. A number, a named standard, a file path, a version. Specifics are what make a passage worth quoting instead of worth paraphrasing, and a paraphrase rarely carries your name with it.

Enough self-containment to travel. At the chunk sizes above, the passage can arrive without your H1 attached to it. An opening that begins "It works by" or "This means that" is broken the moment it leaves the page, because the antecedent stayed behind.

Five ways an opening fails

The warm-up. Forty words establishing that the topic is important and the choice is difficult. Nothing extractable. Whatever follows inside the same few hundred tokens has to carry the whole passage on its own.

The split answer. The answer exists, spread across three sections, and no single chunk holds all of it. This is the most frustrating failure because the page is genuinely good. The fix is a paragraph, not a rewrite.

The dangling pronoun. Covered above, and more common than it sounds. Read your opening aloud with the title covered. If it stops making sense, an engine has the same problem.

The contradicted answer. Your opening says one thing, a product page says another, and a third page has an old figure nobody retired. Pricing is where this bites hardest, which is why pricing pages and what AI says your product costs is worth reading before you touch anything else.

The answer that exists only in a table or a diagram. Images carry nothing to a text retriever, and a table loses its headers in some parsers. Put the finding in prose first, then mark it up. The order in structured data for AI search is the same order: say it in words, then describe it in markup.

A check you can run in ten minutes

Take one page that matters. Then do this.

Strip the navigation, the cookie banner and the hero copy, and copy the first 2,000 characters of body text into a plain file. That is a rough stand-in for chunk one at Microsoft's recommended size. Read only that file. Is the answer in it, complete, with its condition and at least one specific? If you have to scroll back to the page to defend it, the answer is no.

Second, check the robots meta tag on that page. Google's documentation states that nosnippet and max-snippet apply to AI Overviews and AI Mode as well as to ordinary results. A max-snippet:50 left over from an old content policy will crop the paragraph you just wrote, and nothing in your analytics will tell you so.

Third, ask the question in two engines, five times each, and record whether you were named and whether the sentence beside your name was true. One run is a coincidence. Five runs across two engines is weak evidence, which beats none.

The self-check walks through ten of these, including the ones that need no tooling at all.

Where answer-first writing stops helping

It cannot make an engine trust you. If nothing else on the web corroborates your claim, a clean opening gets you retrieved and still leaves you unattributed, and that gap is slow to close. Sequencing matters more than effort here, which is the argument in the first 90 days of AEO work.

It is also a weaker lever on pages with one obvious answer and heavy competition, where the decision is being made on corroboration rather than on phrasing. And the eighty-word figure is arithmetic on the two published chunking defaults cited below, not a threshold we have found stated in engine documentation, and not something we have tested against citation outcomes. Treat it as the right order of magnitude. Do not treat it as a line that switches citation on at seventy-nine words and off at eighty-one.

Honest limits

This post does not establish that answer-first openings get you cited more often. We have not run a controlled before-and-after on a set of pages, and we would not be able to separate the rewrite from everything else that moves in the same weeks. The chunking numbers come from two retrieval products, OpenAI's file search and Azure AI Search, and are used here as a proxy because the engines readers care about do not publish their own indexing settings, which is a substitution a careful reader should push back on. The position effect comes from a 2024 paper on controlled tasks, not from any live answer engine. Where this post says we cannot see something, that is literal: retrieval internals are not visible to us or to you.

FAQ

Is eighty words a real limit? No. It is an estimate derived from published retrieval defaults: around 105 tokens out of a first chunk of 512 to 800. The useful version of the rule is that the answer must be complete before a reader has to scroll, and must survive being cut away from everything around it.

Should I put the date in the opening paragraph? A date helps when the answer changes over time, such as pricing, availability or a version number. It is noise on a definition that has been stable for years. Mark the publication date in the page and in the markup either way.

Does this hurt the page for human readers? We have not run a controlled test on reader behaviour, so we cannot answer this with data. The structural argument is that people who already know the topic get their answer and leave, which was always going to happen, while people who need the detail read on. The section that used to open the page can usually move down intact.

What about long-form pages with several questions? Each section gets its own answer-first paragraph, because each section is a chunk candidate in its own right. One opening cannot carry a two thousand word page.

Sources

Related reading

What makes a page citable by an AI enginePricing pages and what AI says your product costsStructured data for AI search: what to mark up firstThe first 90 days of AEO workWriting case studies an engine can quoteChunking: how your page is cut up before it is read