# How to Write Content That ChatGPT and Other Answer Engines Quote

> Answer one question per page in the first two sentences, then add dated, sourced facts. What GEO research found on citations, and what you can skip.

Published: 2026-11-17
Source: https://www.itgrows.ai/business/blog/how-to-write-content-that-ai-assistants-quote
Publisher: ItGrows (https://www.itgrows.ai/business) — AI visibility for businesses

---

To get quoted, answer one question per page in its first two sentences, then back that answer with specific facts that carry a date and a named source, in passages that still make sense when lifted out of the page. In controlled studies, topic match, concrete facts and evidence moved citations while layout alone did not, and none of it counts until an assistant can actually find the page.

## What does an answer engine do with your page?

It takes pieces, not the whole page. In [guidance published on 8 October 2025](https://about.ads.microsoft.com/en/blog/post/october-2025/optimizing-your-content-for-inclusion-in-ai-search-answers), a Microsoft Bing product manager explains that assistants do not read a page top to bottom the way a person does. They split it into smaller pieces, those pieces get ranked and assembled into answers, and one answer can combine snippets from several sites.

So the unit you are writing is the passage. Microsoft's term for what survives the cut is self-contained phrasing: sentences that "make sense even when pulled out of context."

That gives you a simple editing test. Cover everything except one paragraph and ask three things: does it say what it is about, what it claims, and who says so? "It costs less and lasts longer" fails all three. A sentence that names the product, the figure and where the figure comes from passes, and an assistant can use it without the paragraph before it.

## What should the first two sentences of a page do?

They should answer the question the page is named after, in full, with no scene-setting. Microsoft's list of what makes content eligible for snippets starts with one- to two-sentence responses that directly address a question. Everything below that opening is the how, the exceptions, the evidence and the next step.

Answering first also forces you to decide what the page is about, and that decision carries a lot of weight. [A controlled study by Sprinklr researchers for the SIGIR 2026 conference](https://arxiv.org/abs/2605.25517) ran 252,000 trials across six language models, each time giving a model two sources that differed in one factor and recording which one it cited first. Topical relevance was one of the two biggest drivers, and on-topic sources beat off-topic ones in all six models.

The study compared on-topic with off-topic sources and did not test pages that mix topics, but the logic points to one question per page. A page that tries to settle five questions is partly off-topic for each of them, and its best passage has to compete with pages built around that single question. The same authors advise surfacing core topic terms early, and sources missing key query terms lost in at least four of the six models.

One question does not mean a thin page. Deep coverage beat shallow coverage in at least four of the six models too. Go deep on the costs, exceptions, mistakes and next steps of that one question, and give a genuinely separate question its own page, with a link between the two.

## Which facts make a passage worth quoting?

Specific ones, with a number, a date and a source. The KDD 2024 paper [GEO: Generative Engine Optimization](https://arxiv.org/abs/2311.09735) rewrote source pages for test queries from a 10,000-query benchmark and measured how much of each generated answer drew on each source. Adding relevant statistics, credible quotations and citations of reliable sources brought relative improvements of 30-40% on its word-count-and-position metric and 15-30% on its subjective-impression metric. On Perplexity, tested as a live engine on 200 of those queries, adding quotations did best, with a 22% improvement on the word-count-and-position metric.

Two details matter for a small business. Lower-ranked pages gained most: adding citations lifted pages ranked fifth in search by 115.1%, while top-ranked pages lost 30.3% on average. And the test gave the engine the top five Google results, so these figures describe winning more of an answer once a page is among the retrieved sources, not getting retrieved in the first place.

The Sprinklr study adds two facts that marketing pages often leave out: an explicit price and a recent date. The authors call both gatekeepers across all six models. A 2026 date beat a 2019 date, although a recent date against no date at all reached significance in only some models. The test sources were product reviews, so read the price result as applying to pages where someone is close to buying. The paper's own example of a weak page ends with "Contact us for pricing details."

In practice:

- **Give the number its unit and its condition.** Microsoft's example is "42 dB dishwasher designed for open-concept kitchens" instead of "quiet dishwasher".
- **Date anything that can age.** Write "prices as of 1 November 2026", and keep a visible page date that changes when the content changes.
- **Name the source in the sentence and link it.** If the figure is your own, say how and when you measured it.
- **Publish what only you can state.** Your prices, lead times, coverage areas and test results are facts that a competitor's page cannot supply.
- **Never invent any of it.** The research tested relevant statistics and credible quotations. A number you cannot back up is a number an assistant may repeat to your buyers under your name.

## When do tables and FAQ sections help?

A table helps when several options are compared on the same attributes. Microsoft says comparison tables "break complex details into clean, reusable segments". [A University of Toronto team](https://arxiv.org/abs/2509.08919) recommends detailed comparison tables and clear pros-and-cons lists, because an AI answer narrows the choice to a short list and needs a stated reason for each option. Every cell should hold a fact: a price, a limit, a yes or no with its condition. If a cell would say "great", or the options share no attributes, write a list or a paragraph instead.

An FAQ helps when the questions are the ones buyers really ask and each answer stands alone. Microsoft's example is "How loud is the dishwasher?" answered with 42 dB, and it says assistants can often lift such pairs word for word. Keep the answers in visible text, because Microsoft warns that AI systems may not render content hidden in tabs or expandable menus.

One caution. In the Sprinklr test, with the facts held equal, organized sections did not beat dense paragraphs, and the authors say formatting changes can be deprioritized. Microsoft describes how pieces get lifted, and the lab test measured layout alone. Read together, they say structure makes good material easy to lift but does not replace the material. Clearer writing is a separate edit: the KDD 2024 paper found that improving fluency and readability raised visibility by 15-30%.

## What should you fix first?

Start with the page's topic and its missing facts, and leave layout for last. The order follows the Sprinklr authors' own checklist, with the KDD 2024 edits added to the second round.

| Edit to the page | What the research found | Priority |
| --- | --- | --- |
| Keep the page on one exact question | On-topic sources beat off-topic ones in all six models (Sprinklr) | First |
| State the price on buying pages | A gatekeeper across all six models (Sprinklr) | First |
| Update old pages and show the new date | A 2026 date beat a 2019 date, a gatekeeper across all six models (Sprinklr) | First |
| Add real statistics, quotations and cited sources | 30-40% and 15-30% relative gains on two metrics (KDD 2024) | Second |
| Add specifications and comparisons | Significant in at least four of six models (Sprinklr) | Second |
| Replace hedging with evidence | Significant in at least four of six models (Sprinklr) | Second |
| Make sentences clearer and easier to read | 15-30% visibility gain from fluency and readability edits (KDD 2024) | Second |
| Rework the layout and nothing else | No impact from formatting choices alone (Sprinklr) | Last |

## What can you skip?

- **Keyword stuffing.** The KDD 2024 paper found little to no improvement, and on Perplexity it performed 10% worse than the baseline.
- **A more persuasive tone.** The same paper found no significant improvement overall from a more persuasive and authoritative style, though it did help on debate-style questions. Microsoft adds that calling something "next-gen" or "cutting-edge" without context leaves AI unsure how to classify it.
- **New AI files and special markup.** [Google's documentation on AI Overviews and AI Mode](https://developers.google.com/search/docs/appearance/ai-features), last updated on 10 December 2025, says "You don't need to create new machine readable files, AI text files, or markup to appear in these features", and that no special schema.org data is needed. Microsoft's guidance does recommend schema, as a way to label content for machines, so standard markup is fine where it labels what the page already says, and Google asks that it match the visible text. See [what our logs show about llms.txt](https://www.itgrows.ai/business/blog/what-is-llms-txt) and [whether schema markup helps](https://www.itgrows.ai/business/blog/schema-markup-for-ai-search).
- **Moving the date without changing the page.** The date should mark a real change. A fresh date on stale facts gives an assistant nothing new to quote and can put your old numbers in front of buyers.

## Why might a well-written page still not get quoted?

Because writing only works on pages an assistant retrieves. On 19 October 2026 we asked assistants for the exact titles of 18 articles, each older than three weeks, published on three client sites. ChatGPT found 1 of 18; Claude found 7 of 18. Google had indexed 29 of 30 articles of one of those sites. Being in Google's index did not mean an assistant could find the page.

The Sprinklr study points the same way. The other biggest driver of being cited first was list position, and its authors treat absence from citations as a retrieval problem for SEO, not a writing problem. So check access first: make sure [AI crawlers are allowed in robots.txt](https://www.itgrows.ai/business/blog/how-to-allow-gptbot-in-robots-txt) and that the page is indexed.

Not every question triggers a search either. In our 23 September 2026 test of 20 German business questions, the model answered 10 without searching at all and cited nobody; where it searched, it cited primary sources (laws, ministries, OECD, large accounting firms) on 9 of 10; firms were named only on 'who should I hire / how to choose' questions; two runs of the same question named two different sets of firms. That gives you two writing decisions. On factual questions you compete with primary sources, so cite them and add what they lack, such as your prices, conditions and cases. To be named as a business, write pages for the how-to-choose questions, with your criteria stated plainly.

Your own site is also only part of what assistants read. The University of Toronto team, comparing AI search with Google in a paper dated 10 September 2025, found a "systematic and overwhelming bias" towards earned media, meaning third-party, authoritative sources, over brand-owned and social content. So give reviewers, trade publications and institutions facts they can check and repeat.

Then [measure the answers over time](https://www.itgrows.ai/business/blog/how-to-measure-ai-visibility), because a single ask proves little. To see where your site stands, the free check at [ItGrows](https://www.itgrows.ai/) generates five buyer questions from your homepage and shows two of ChatGPT's and Perplexity's answers right away, and all five once you create an account. If you want this done for you, ItGrows does the work on your own website and every two weeks puts your buyer questions to ChatGPT, Claude, Gemini and Perplexity, counting the answers that name you and the sites they cite.

## Frequently Asked Questions

**Does adding statistics really get a page cited more often?**

In the KDD 2024 paper, adding relevant statistics, quotations and citations raised a source's visibility in the answer by 30-40% on a word-count-and-position metric. That result applies to pages already among the sources given to the engine, so it does not show that statistics get a page retrieved. Use only figures you can source and date.

**Do I need FAQ schema or an llms.txt file to be quoted?**

Not for Google's AI features: its documentation says no new AI text files, markup or special schema.org data are needed for AI Overviews and AI Mode. Microsoft does recommend schema as a label for content, so add standard markup where it accurately describes what is already on the page. Either way, the visible text is what an assistant can lift, so write the questions and answers there first.

**How long should a page be for an assistant to quote it?**

None of the studies cited here sets a word count, so length is not the target. The Sprinklr test kept both versions within 5% of the same length, so its win for deep coverage over shallow was a win for depth, not for more words, while Microsoft recommends one- to two-sentence answers at the level of the passage. Write the page as long as its one question needs, in short passages that stand alone.

**How can I check whether assistants quote my pages?**

Put your buyers' questions to each assistant more than once, note who is named and which sites are cited, then repeat with the same questions. Answers to the same question can change between runs, so compare like with like. Check the citations themselves, because being named and being cited are separate results.
