How to get cited by ChatGPT and AI search: write answer-first, data-rich pages

Content gets cited when a model can lift a sentence from it that answers the question on its own. In practice: the answer in the first two paragraphs, figures with a date and a source, links to external sources, and headings phrased as the questions people ask. The GEO study (KDD 2024) measured lab visibility gains of up to 40%.1

To check it on your own site: AI visibility audit

What makes a paragraph citable?

A citable paragraph is a block of text that answers a specific question without relying on what surrounds it. It names its subject, contains a fact that can be checked and, if the fact is someone else's, gives the source. If your best sentence opens with "this" or "as mentioned above", it cannot stand alone, and the assistant will pick another page's sentence instead.

Assistants work with fragments. Google explains that AI Overviews and AI Mode may run several related searches across subtopics, a technique it calls query fan-out, to build a response.2 Each of those searches is a chance for one section of your page to fit, provided that section makes sense in isolation.

This article is the writing side of our AI search optimization guide, which covers the other pieces: crawlers, structured data and measurement.

What did the GEO study actually measure?

The 40% figure gets quoted a lot, so it is worth reading in full. The paper "GEO: Generative Engine Optimization" (Aggarwal et al., KDD 2024), which gave generative engine optimisation its name, built GEO-bench, a set of 10,000 queries, and simulated a generative engine with GPT-3.5 answering from the top five Google results.1 For each query the authors rewrote one of those five sources using one of nine techniques, then measured how much of the final answer came from it.

The authors sum up that the three best techniques achieved a relative improvement of 30 to 40% on that metric. Three caveats rarely make it into the summaries:

• It is a lab, not ChatGPT. On Perplexity.ai, with a 200-query sample, the best technique improved the word-count visibility metric by 22% and keyword stuffing came out 10% worse than the baseline.
• The effect depends on the topic. Statistics worked best for law, government and opinion queries; quotations for history, people and society. The authors recommend domain-specific adjustments.
• It helps lower-ranked pages most. Citing sources raised the visibility of sites ranked fifth by 115.1%, while the top-ranked site lost 30.3% on average.

With those caveats, the pattern matches what Google asks for. According to John Mueller of Google, the advice for its AI search experiences is to create "unique, non-commodity content" that visitors find helpful.3

1. How do you answer in the first two paragraphs?

Write the complete answer in the first paragraph, in 40 to 60 words: subject, answer and the nuance that makes it accurate. Context, background and examples come afterwards. If the question is "how long does an office move take?", the first sentence gives a timeframe rather than saying "every move is different".

I wrote this example for this article: a typical corporate blog paragraph and its rewrite. The figures in the new version come from the GEO study.

Before. In recent times, generative engine optimisation has become absolutely essential for any business that wants to stay relevant in the future of search. Many experts agree that these techniques work and can make a huge difference to your results.

After. Generative engine optimisation (GEO) is the set of changes to a text that aim to make an AI assistant use it in its answers. In the study that coined the term (KDD 2024), adding quotations, statistics or sources increased the share of the answer drawn from a page by 30 to 40%.

2. What counts as citable data?

Citable data is a figure, a date, a price or a specific comparison that can be copied without rewording. "Next-day delivery to mainland UK", "three-year warranty" and "updated 12 September 2026" all qualify; "fast delivery" and "generous warranty" do not. For a local business, opening hours, address and phone number are its citable data, as our guide to local SEO for AI search explains.

Every third-party figure carries three things, in the same sentence or its footnote: who published it, when, and on what sample. "According to SISTRIX, ChatGPT swaps between 42% and 74% of the domains it cites every week, depending on the country (82,619 prompts, December 2025 to April 2026)" stands on its own and can be verified.4 If a figure circulates without a methodology, leave it out.

3. Why link to sources and attribute quotations?

Linking to the original source turns a claim into something checkable. In the GEO study, citing sources lifted visibility from 19.5 to 24.9. Link to the primary source, such as the official documentation or the study with its methodology, not to the blog post that summarised it.

An attributed quotation works the same way: a name, a role or organisation, and the words in quotation marks. Google's guidance on helpful content asks whether it is self-evident who wrote each page and encourages bylines where readers expect them.5 A visible byline and date are the minimum version of that attribution; our guide to E-E-A-T author signals goes further.

4. Do question headings, lists and tables help?

Yes, when the section genuinely answers that question. An H2 such as "How much does a new boiler cost?" mirrors the way people ask an assistant, and the first sentence underneath should answer it. "Our services" or "Learn more" say nothing about what the section answers.

Steps belong in numbered lists and comparisons in tables, because an extractor can take them whole. No public study measures how much each format matters to ChatGPT, so treat this as good practice for clarity, not a guaranteed lever.

5. Do you need a summary at the top and an FAQ at the end?

A summary of three to five bullet points straight after the answer gives the reader, and the model, the key ideas with their figures. An FAQ block at the end catches the questions that do not fit the main thread, with answers of 40 to 80 words that make sense on their own.

Use the questions your customers really ask: the ones in your inbox, the ones your sales team hears, or the ones in Google's autocomplete. Questions that only suit you help nobody.

6. How long does a page need to be to cover the intent?

Long enough to settle the question and the follow-ups it triggers, and no longer. Google is blunt about it: it has no preferred word count.5 Even so, a page under 300 words rarely covers more than a definition.

Mueller adds that people using AI search ask longer, more specific questions and then follow up.3 Covering the intent means anticipating that second question: the price after the definition, the lead time after the price, the exceptions after the lead time.

What does not work?

• Repeating the keyword. In the GEO study it was the only technique that reduced visibility (17.8 against 19.5), and Google treats keyword stuffing as spam.6
• Text hidden from readers. White text on a white background, text positioned off-screen or set to size 0: Google's spam policies list these explicitly.
• Padded introductions. Three paragraphs of context before the answer push the citable sentence down the page.
• "AI files" instead of content. Google states that you do not need new machine-readable files, AI text files or markup to appear in its AI features.2 We look at the evidence in our article on llms.txt.
• Changing the date without changing the text. Google's helpful content guidance explicitly asks whether you change dates to make pages seem fresh when nothing substantial has changed.
• Blocking the search crawler. OpenAI says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links.7 Which crawlers to let in, and which you can safely block, is covered in our guide to AI crawlers and robots.txt.

And no practice guarantees a citation. Over 17 weeks, SISTRIX found AI Mode replaces between 54% and 59% of its sources every week, depending on the country, and ChatGPT between 42% and 74% (60% in the UK, 74% in Germany).4 Good writing puts you on the shortlist; it does not reserve your place.

How can you check a page in two minutes?

Do it by hand with these five questions:

1. Cover everything except the first paragraph. Does it answer the question in the title?
2. Is there at least one figure, date or price in each section, with a source if it is not yours?
3. Do you link to at least one source outside your own site?
4. Does any heading ask the question your customer asks?
5. Is there a list or table wherever there are steps or comparisons?

If you would rather automate it, a disclosure first: this blog belongs to SmoothSeen. Its analysis reviews these same points in the AI visibility half of the report:

It is a writing checklist, not a prediction: a page that passes every check is not guaranteed a ChatGPT citation. The full list is on the AI visibility features page, and the complete process, block by block, is in our AI visibility audit guide.

What to do next

Pick the three pages that matter most to you and rewrite each first paragraph to the 40-to-60-word rule. Then add one sourced figure to every section and go back to the AI search optimization guide for the rest of the pieces.

Frequently asked questions

What kind of content does ChatGPT cite?

OpenAI does not publish how it chooses the sources it cites. The only documented requirement is technical: the page has to be open to OAI-SearchBot. Beyond that, published experiments point to text that answers early, with concrete figures, linked sources and attributed quotations, although those were measured on a test bench rather than in ChatGPT itself.

How many words does an article need to get cited by AI?

Neither Google nor OpenAI publishes a minimum, and Google says it has no preferred word count. What matters is that the answer comes first and that the page settles the follow-up questions. Under 300 words a page usually falls short; over 2,000, check that none of it is padding.

Do statistics help you get cited?

In the GEO study, adding statistics was one of the most effective techniques: visibility rose from 19.5 to 25.9 on its metric, about 33% more. It worked best for law, government and opinion topics. These are lab results with GPT-3.5, so the direction is useful but the percentage is not a forecast for your site.

Is it worth phrasing headings as questions?

It helps each section state what it answers, which mirrors the way people ask an assistant. No public study isolates its effect on citations, so use question headings only when the section genuinely answers that question, and give the answer in the first sentence under the heading.