Answer-first writing: structure pages so people and AI can both quote them

Answer-first writing puts the answer to a page's main question in its first paragraph, gives each section one idea, and words every definition and figure so it still makes sense when lifted out of the page. Scanning readers reach the point sooner, and a system that extracts one passage gets a complete one. It needs no special markup.

What does answer-first writing mean?

It means three habits, each easy to test on a draft.

The first paragraph answers the question in the title. If the page asks how long a review takes, the opening sentence gives the duration, and the background follows. One idea per section keeps each heading honest: it promises one thing, the section delivers it and stops. Definitions stand alone: "X is Y" in a single sentence, with X named in full instead of "it" or "this approach", so the sentence survives being copied into another page, a chat answer or a search snippet.

Why does structure help people who scan?

Nielsen Norman Group (NN/g) reported in 1997 that people skim web pages: Jakob Nielsen wrote that 79% of test users always scanned any new page and 16% read word by word (opens in a new tab)^[6]. The page doesn't state how many people took part. A paper by John Morkes and Jakob Nielsen (opens in a new tab)^[7] is more specific. In its second study, 15 of 19 participants always tried to scan unfamiliar text and 3 read it word by word, which matches 79% and 16% when rounded, though the 1997 page doesn't say it uses those data. Its measurement study had 51 experienced web users each work with one of five versions of the same seven-page travel site. On an overall usability score (a geometric mean of five normalised measures: task time, task errors, memory, sitemap time and subjective satisfaction, with the promotional control set to 100), the scannable version scored 147. A version combining concise, scannable and objective writing scored 224.

Eyetracking points the same way. NN/g's article on the F-shaped pattern (opens in a new tab)^[8] (Kara Pernice, published 2017, reviewed August 2026) reports fixations clustering at the top and left of a page. It advises putting the most important points in the first two paragraphs and starting headings with the most informative words. Its own title calls the pattern "misunderstood", and it lists other scanning patterns, so treat it as a tendency, not a law.

The 1997 test used one small site and fact-finding tasks. Neither NN/g source tests answer-first pages against AI assistants.

Why does it help retrieval systems?

Question-answering systems built on retrieval often select passages, not whole pages. The Dense Passage Retrieval paper (Karpukhin et al., 2020 (opens in a new tab)^[9]) describes open-domain question answering as relying on "efficient passage retrieval to select candidate contexts". A passage that opens with its answer and names its subject can be selected and read without the paragraphs around it. One that says "as mentioned above" is harder to use on its own.

Liu et al. (2023) (opens in a new tab)^[10] tested language models on long inputs and found that "performance is often highest when relevant information occurs at the beginning or end of the input context." That concerns the order of text inside a model's input, not the layout of a web page, so it supports putting the answer early only by analogy.

These papers don't describe how any particular assistant splits web pages. The case for structure here is a design inference, not a measured result. For what studies of AI referral traffic do and don't count, see what the studies measure.

What did the GEO paper find?

The GEO paper (Aggarwal et al., arXiv 2311.09735 (opens in a new tab)^[2], submitted November 2023, accepted to KDD 2024) tested whether rewriting a page changes how visible it is in an AI-generated answer. The authors built GEO-bench, 10,000 queries across 25 domains, and a custom pipeline that retrieved the top five Google results and had gpt-3.5-turbo write a cited answer. They rewrote source text with nine methods: Authoritative, Statistics Addition, Keyword Stuffing, Cite Sources, Quotation Addition, Easy-to-Understand, Fluency Optimization, Unique Words and Technical Terms. The main results come from the 1,000-query test split, averaged over five random seeds, with one randomly chosen source per query rewritten.

Visibility was scored two ways. Position-Adjusted Word Count counts the words in sentences that cite the source, weighted by where the citation appears. Subjective Impression is a set of seven scores from GPT-3.5, such as relevance and click likelihood. The paper rescales Subjective Impression to the same mean and variance as the first metric, so both start from a baseline of 19.3 and the raw scores aren't comparable. The table below uses the Overall column for Position-Adjusted Word Count and the Average column for Subjective Impression.

Scores are from the paper's Table 1; percentages are calculated here against the unchanged baseline. The paper's text and its table don't fully agree. The text says Cite Sources, Quotation Addition and Statistics Addition each gained 30-40% on Position-Adjusted Word Count and 15-30% on Subjective Impression, but Table 1 gives Cite Sources +27% and +13%, below Fluency Optimization on the first metric. The text also reports "no significant improvement" for Authoritative, while Table 1 shows +10% and +19%. The percentages above follow the table.

On the deployed Perplexity.ai engine, tested on a 200-query subset of the test split with source text supplied as file uploads, Quotation Addition did best on Position-Adjusted Word Count with a 22% improvement over the baseline.

A separate test rewrote every source for a query with the same method at once. The paper's Table 2 reports relative changes in "visibility" by position in the Google results, without saying which of its two metrics it uses. Sources ranked fifth gained 115.1% (Cite Sources), 99.7% (Quotation Addition) and 97.9% (Statistics Addition). Sources ranked first lost 30.3%, 22.9% and 20.6% with the same three methods. The authors argue this favours lower-ranked sites; rank isn't the same as size, and the figures come from one simulated pipeline. The wider picture is in generative engine optimisation for expert businesses.

Four limits apply:

• The methods are model-written rewrites of wording and content: style, statistics, quotations, citations. None is described as changing headings, summaries, tables or answer order, so answer-first structure isn't tested directly. Easy-to-Understand and Fluency Optimization are the closest.
• The tests ran on one simulated pipeline and one live engine. The authors say engines will change, and they didn't measure effects on classic search rankings.
• "Visibility" means how much of an answer cites a source, not clicks or sales.
• Quotations, figures and citations help only if they are real and checkable. Invented ones would mislead readers.

What does Google say about markup and structure?

Google's documentation on AI features in Search says "There are no additional requirements to appear in AI Overviews or AI Mode" (opens in a new tab)^[3], that no special schema.org structured data is needed, and that you don't need new machine-readable files or AI text files (last updated 10 December 2025). It does recommend that important content is available in text and that structured data matches the visible text. For the file-based side of the question, see does llms.txt matter.

Google's newer guide to optimising for generative AI features (opens in a new tab)^[1] (added 15 May 2026, last updated 10 July 2026) addresses structure directly: "There's no requirement to break your content into tiny pieces for AI to better understand it." It adds that people generally appreciate pages organised by paragraphs and sections, with headings that give a clear structure. The case for answer-first writing here is that it serves readers; Google doesn't say its AI features need it.

Google's structured data policies (opens in a new tab)^[4] make the second point a rule: "Your structured data must be a true representation of the page content." Content readers can't see shouldn't be marked up (last updated 10 July 2026).

FAQ rich results are gone. In an entry dated 8 May 2026, the Search updates log (opens in a new tab)^[5] says "This feature will no longer appear in Google Search starting May 7, 2026", and a 15 June entry removed the documentation. An FAQ section still helps readers who arrive with the same few questions, so write it for them. Any markup should match the questions and answers on the page.

Which page elements carry the structure?

The 60-word and 40-to-70-word limits are editorial rules of thumb, not thresholds from the research. Named authors and consistent business details are also an entity question, covered in entity SEO for experts.

What does a before and after look like?

This illustrative example uses an invented studio and invented figures. The page question is how long a portfolio review takes.

Before

Every creative's portfolio is different, and we've always believed good feedback can't be rushed. Over the years we've refined a process that balances thoroughness with flexibility. Timelines depend on the size of the portfolio, how many rounds are needed and how busy we are, so it's best to get in touch and we'll work out what suits you.

After

A portfolio review takes five working days from the day your files arrive. You receive a written report of up to six pages and one 45-minute call. A review covers up to 20 pieces; larger portfolios are split into two rounds, which adds five working days. If you have a deadline, give it when you book and we will confirm in writing whether we can meet it.

The Before has nothing a reader or an extraction system could quote: no duration, no named subject, and an instruction to make contact. The After leads with the duration, names the thing being timed so the first sentence stands alone, gives each condition its own figure and replaces "get in touch" with a specific action.

What goes on a ten-point page checklist?

1. The first paragraph answers the title question in 60 words or fewer.
2. Every H2 is a question a reader would ask, with the informative words first.
3. Each section holds one idea and stops when it has made it.
4. Every definition reads "X is Y" and names X in full, with no "it", "this" or "as above" doing the work.
5. A 40-to-70-word summary sits at the top and makes sense alone.
6. Comparisons sit in a table with the same attributes for each option.
7. Published, updated and reviewed dates and a named author are visible.
8. Every figure carries its year, population and metric, with the source linked at the claim.
9. Any FAQ is visible text written for readers, and any markup matches it.
10. Key facts are in text, not only in images or video.

For expert businesses, the same rewrite applies to what a Digital Mind learns from. Plandora's guidance for a document in a Mind's Drive is to keep one topic per document and use headings or lists for anything visitors will scan. Test the result in the chat preview: AI replies can be wrong.

This blog follows the format described here. Each article opens with a summary, lists takeaways, ends with an FAQ and a numbered source list, and shows published, updated and reviewed dates. That describes how the pages are built. It is not evidence that the format earns citations, and this article doesn't claim it does.