Content that AI cites: how to write pages ChatGPT and AI Overviews pick up

Content that AI cites answers a question directly, backs up that answer with verifiable facts and shows who wrote it and why that person knows. ChatGPT, Perplexity, Gemini and Google AI Overviews build their answers from snippets of text from different pages. A paragraph that holds up on its own and looks trustworthy is easier to pick up than one that only makes sense after three screens of context.

This article is a tactical guide. It starts with E-E-A-T as Google describes it and translates that into concrete writing choices you can apply to your own site today. If you want the bigger picture first, read the complete guide to GEO. How an AI system selects its sources before it cites them is covered in the article on how AI search engines choose sources.

One caveat up front: no writing style guarantees a citation. What you can do is lower the barrier. You do that with the same qualities Google has been asking for for years, just applied more rigorously.

#What does Google mean by E-E-A-T?

E-E-A-T stands for experience, expertise, authoritativeness and trustworthiness. Google describes the concept in its documentation on helpful, reliable, people-first content. It also states there that trust is the most important element. The other three contribute to it, but a page doesn't have to demonstrate all three.

Two nuances matter:

• E-E-A-T is not a standalone ranking factor. Google says so explicitly. Its systems look at a mix of signals that together point to E-E-A-T.
• The rater guidelines don't steer rankings directly. In December 2022, Google added the E for experience to the Search Quality Rater Guidelines. Google explains that raters use them to evaluate the quality of its search systems, and that the guidelines don't directly influence rankings. They are, however, useful for taking a critical look at your own content.

The example Google gives makes the difference between experience and expertise concrete. To fill in your tax return correctly, you want an expert in accounting. For a verdict on tax software, you want people who have actually used the software themselves.

In the same documentation, Google poses three questions you can apply to any page: who created this, how was it created and why does it exist. The answer to the why question should be that the page helps people, not that it's meant to attract search traffic.

Google gives the same advice for AI search features. According to the documentation on AI features in Google Search, there are no additional requirements to appear in AI Overviews or AI Mode, and the existing SEO fundamentals still apply. In a blog post from May 2025, Google recommends focusing on unique content that can't already be found everywhere. That's why it pays to get clear on the difference between GEO and SEO first: much of what works is good SEO with more emphasis on clarity and evidence.

#What does research say about citable content?

The best-known research on this topic is the paper GEO: Generative Engine Optimization by Aggarwal and others, published on arXiv and accepted for KDD 2024. The researchers introduced the term GEO and tested nine ways of modifying a source text. They then measured how visible that text was in answers from generative search engines.

The broad findings:

• Citing sources, adding quotations and adding statistics were among the best-performing modifications.
• Writing more fluently and more clearly also gave a clear improvement.
• Keyword stuffing, repeating search terms, achieved little to nothing.
• Effectiveness varied by domain. What works for one topic doesn't automatically work for another.
• Lower-ranked websites benefited relatively more from these modifications than sites already at the top.

According to the paper's abstract, GEO can increase visibility in generative answers by up to 40%. Read that figure in the right context: it refers to a benchmark of search queries compiled by the researchers, not a promise for your site. The direction is useful, though. Concrete evidence and clear language help; keyword tricks don't.

#Start every section with the answer

A citable section opens with a sentence that fully answers the question, even if someone reads only that sentence. AI systems that work with RAG (retrieval-augmented generation) retrieve text fragments and use them as the basis for their answer. That process is called grounding. A fragment that starts with "As discussed above" or "That depends" loses its meaning as soon as it's cut out of the page.

So write according to the "answer first" principle:

1. Put the conclusion in the first sentence of the section.
2. Name the subject instead of writing "this" or "it".
3. Then give the explanation, the nuance and the example.
4. Save exceptions for the end of the section.

This isn't a new idea. Journalists call it the inverted pyramid. The difference is that an AI system doesn't keep reading to work out what you mean.

#Use headings that ask your reader's question

A question-led heading tells both the reader and the AI system exactly which question the section answers. Google describes how AI Overviews and AI Mode can use a query fan-out: the system splits one question into several sub-questions and searches for each separately. A heading like "How long does it take to convert a sole trader business into a limited company?" matches such a sub-question better than a heading like "Timeline".

Practical rules:

• Phrase headings the way someone would ask ChatGPT the question.
• Cover one question per section.
• Keep the heading and the first sentence aligned in substance, so they reinforce each other.
• Use a logical order of H2 and H3, without skipping levels.

Not every heading has to be a question. A clear statement ("Put comparisons in tables") works just as well, as long as the heading says what the section delivers.

#Before and after: a paragraph rewritten

The difference between poorly and easily citable text is clearest side by side. Both examples below are fictional and meant as illustrations. The company, the figures and the source don't exist.

Before (hard to cite):

As you may already know, this is a really important topic that a lot of business owners struggle with. We find that it can actually happen quite quickly, as long as you go about it the right way. That's why we've built up a lot of experience with it over the years and we're happy to help.

What goes wrong here: the paragraph doesn't name the subject, contains no verifiable fact and doesn't say who "we" are. An AI system can't take anything from it.

After (easy to cite):

A fictional installer from Leeds usually fits a heat pump in an average terraced house in two working days, counted from the day the materials arrive. The preparation, such as the heat loss calculation and choosing the output, takes another one or two weeks beforehand. This is based on the fictional company's own scheduling data for 2025, published on its "How we work" page.

What works better here: the first sentence answers a concrete question, the subject is named, the duration is specific and the origin is stated. In a real text, you'd link that origin to the source. The scoping ("average terraced house", "from the day the materials arrive") also makes the claim more trustworthy.

#Back up claims with concrete facts and sources

A claim with a number, a name and a link to the primary source is more trustworthy than a general statement. This is in line with the findings of the GEO paper and with the emphasis Google places on trustworthiness.

Here's how to apply it:

• Link to the primary source. Refer to the official documentation, the legislation or the original research, not to a blog that refers to it in turn.
• Say who is making the claim. "According to Google's documentation" is more verifiable than "experts say".
• Use your own data if you have it. Your own measurements, customer questions or hands-on experience are exactly the unique information Google asks for.
• Leave out a figure if you can't back it up. A made-up percentage undermines trust in the rest of the page.
• Quote briefly and verbatim when a statement from a person or organisation carries the point, with their name and role.

#Put comparisons in tables and steps in lists

Tables and lists make the structure of information explicit, so an AI system has less guessing to do about which value belongs to which term. A comparison spread across three paragraphs of running text is hard to summarise reliably.

Use real HTML tables and lists, not images of tables. If you also want to make the meaning of your page machine-readable, read more about schema markup for AI. That article covers JSON-LD and how to link your organisation as an entity.

#Show that your content is up to date

A visible, honest date tells readers and systems how recent the information is. In its documentation on publication dates, Google recommends showing a visible date prominently on the page, with a label such as "Published" or "Last updated". You can also include datePublished and dateModified in your structured data. The visible date and the date in the markup should match.

Watch out for the flip side. Google explicitly calls changing the date without any substantive change, to make a page look fresh, a warning sign. So only update the date when you've genuinely refreshed facts, examples or advice. For time-sensitive facts, also mention the year in the text itself, so a cited sentence doesn't quietly go out of date.

#Make it visible who is writing

An author with demonstrable experience answers the who question Google asks. Google strongly encourages showing accurate author information, such as a byline, where readers expect it. That name can link through to more background about the author. At the same time, Google warns against fabricated author profiles, such as generated portrait photos or non-existent qualifications.

What to add in practice:

• A byline with the author's real name on every article.
• An author page with their role, relevant experience and links to profiles elsewhere, such as LinkedIn. In structured data, you record that link with sameAs.
• An about page that says who the company is, what it does and since when.
• Contact details that are easy to find.
• An explanation of your process where readers might wonder how something came about, including whether you used AI in writing it.

#Checklist: is your content citable?

Go through these points for each page before you publish or update an existing one:

1. The first paragraph answers the main question directly.
2. Every H2 opens with a sentence that holds up without context.
3. Headings state the section's question or claim.
4. The subject is named in the opening sentence, not referred to as "this" or "it".
5. Every statistic or external claim links to the primary source.
6. There are no figures you can't back up.
7. Comparisons are in a table, steps in a numbered list.
8. The page shows a visible publication or update date that matches the structured data.
9. A real author is named, with a link to an author page.
10. There is an about page and contact information.
11. The page contains something unique: your own experience, your own data or your own example.
12. There is no keyword stuffing.

#How to check this on your own site

You can check the key signals yourself in fifteen minutes:

1. Open an important page and read only the first sentence under each H2. Does each sentence answer the heading without needing the rest?
2. Highlight every figure and every external claim. Is there a link to the primary source?
3. Find the author. Is a name shown, and does it lead to a page with background?
4. Check that a visible date is shown and that it matches datePublished and dateModified in the source code.
5. Check that your about and contact pages can be reached from the navigation or footer.
6. Ask ChatGPT, with search turned on, and Perplexity the question your page answers. Note which sources they mention and what those sources do differently.

Flowproof's free scan automatically checks some of these E-E-A-T signals, including author, about page, contact, source citations, date and question headings. Create a free account to see where your own pages are still leaving gains on the table.