Content that AI cites: how to write for ChatGPT, Perplexity and Google AI

Content that AI cites answers the question in the first sentences, consists of clear sections that each stand on their own, backs up claims with verifiable facts and shows who wrote it. That shows in a Semrush analysis of 304,805 URLs cited by AI (January 2026): cited pages scored clearly higher on clarity and summarization (+32.8%), E-E-A-T signals (+30.6%), question-and-answer format (+25.5%) and section structure (+22.9%) than pages that only rank in Google.

This guide brings together what you need to write for ChatGPT, Perplexity, Gemini and Google AI Overviews: what research really shows, how to structure a page, how to raise fact density, how often to update, why original data works so well and how to pick topics based on the questions customers ask an AI assistant.

What do we really know about content AI cites?

There are now enough large analyses to see patterns, but they remain correlations: nobody outside OpenAI, Google or Perplexity knows the exact selection rules. These are the sources that weigh most.

• Semrush (January 2026). Semrush asked ChatGPT Search, Google AI Mode and Perplexity 11,882 questions and compared the 304,805 cited URLs with pages from Google’s top 20, on thirteen text features. Besides the four features above, structured elements such as lists and tables also scored higher (+21.6%).
• Rankability (September 2026). In an analysis across eleven AI platforms and 31 topics, pages that introduced the topic in the first 100 words were among the first ten citations 22.1% of the time, against 8.7% for pages that did not. Pages that covered the topic broadly reached 27.6% against 14.5%.
• The GEO paper (Aggarwal et al., KDD 2024). The research that introduced the term Generative Engine Optimization tested nine rewriting methods. Adding quotations, adding statistics and citing sources gave the biggest gains, up to about 40% more visibility in AI answers. Keyword stuffing did worse than the unchanged text. In a real-world test on Perplexity the gain was up to 37%.
• Google itself. In AI features and your website, Google writes that there are no extra requirements for AI Overviews and AI Mode: your page must be indexed and eligible to show with a snippet. The guide Creating helpful content asks for original content that goes beyond the obvious, with clear sources and visible who, how and why.

The common thread: content for AI is good content for people, only stricter. The answer has to come sooner, the facts have to be harder and the author has to be more visible.

Why does the answer need to be in your first 100 words?

The answer belongs up front because AI systems pull separate passages from your page and the opening is used most often; the Rankability numbers above show how big that difference is. Put a paragraph of two to four sentences under your H1 that answers the main question, naming the topic literally. Nuance follows after.

Weak, without a citable answer:

GEO is a relatively new concept in digital marketing and is becoming more and more important. Many businesses wonder what it means. In this article we explain what GEO is and why it matters.

Strong, with the answer in the first sentence:

GEO (Generative Engine Optimization) is making your content findable and citable in the answers of ChatGPT, Perplexity and Google AI Overviews. Where SEO puts your page high in the list of search results, GEO makes sure you are in the answer itself. That matters: according to Ahrefs, an AI Overview lowers the click-through rate of the first search result by 58%.

The second version contains a definition, a boundary and a number with a source (Ahrefs, May 2026, 300,000 keywords). An AI model can take each of those three separately.

How AI systems split a page into pieces, and how big those pieces are, is in content chunking for AI citations.

How do you structure a page AI can cite?

Build your page from sections that each answer their own question and make sense without the rest of the text. That matches how AI search engines work: Perplexity describes that its index splits documents into standalone pieces and ranks them both per document and per piece. A section that starts with “as described above” is useless as a standalone answer.

• H2s as a question or a clear topic. Use your customer’s words: “What does a heat pump cost for a townhouse?” instead of “Investment”.
• The first sentence is the answer. Repeat the topic in that sentence, so the passage also stands on its own.
• Short paragraphs. One idea per paragraph, three to four sentences.
• Lists and tables for steps, options and comparisons. A table cell is a ready-made answer to a specific sub-question.
• An FAQ with real questions and answers of one to three sentences. If you use FAQ markup, make sure it matches the visible text exactly; Google asks for that in its guidelines for AI features too.

What is fact density and how do you raise it?

Fact density is the share of verifiable facts in your text: how many of your sentences contain a number, a name, a date, a comparison or a source instead of an adjective. A citable claim is specific, verifiable, understandable without context and concrete. “We deliver fast, expert work” meets none of those four. The GEO paper shows why this works: statistics, quotations and cited sources raised visibility the most.

Weak:

We are an experienced installer and help customers quickly and expertly to a sustainable solution.

Strong (fill in your own numbers):

Since [year] we have installed [number] heat pumps in the Denver metro area. We install an air-source heat pump in a townhouse in [number] working days, including commissioning. The quote is based on a heat loss calculation for each home.

Five techniques

1. Replace vague words with numbers. “Many”, “considerable” and “fast” become counts, amounts and lead times. Only use numbers you can back up.
2. Quote people with name and role. A statement from your own specialist or a customer, with name and role, was the strongest single change in the GEO research.
3. Publish your own data. Numbers only you have are unique by definition. How to do that is further down.
4. Put comparisons in a table. Prices, lead times, pros and cons per option.
5. Add a date and source to every number. “In 2025 (source: Census Bureau)” is stronger than a bare percentage, and for you it is a reminder of when the number gets outdated.

Where it goes wrong

• Making up numbers or rounding to what is “about right”. An AI model can put your claim next to other sources, and so can a reader. If you have no hard data, write honestly what you do know: “across the [number] installations we did last year”.
• Overclaiming. “The best platform in America” is not verifiable and so not citable.
• Other people’s numbers without a source. Name the source and link to it.
• Only facts. A row of statistics without explanation reads badly. Work per paragraph from fact to meaning to conclusion.

Do the fact density audit

Take your three most important pages and go through them paragraph by paragraph:

1. Count the verifiable claims per paragraph.
2. Mark paragraphs without any claim: those come first.
3. Replace vague words with numbers you can back up.
4. Check that every number from someone else has a source.
5. Delete whatever still says nothing after this round.

If you use AI to write, also read how to rewrite AI text so it does not sound like AI.

Why should you cover the topic broadly?

A page that also answers the sub-questions around a topic gets cited more often than a page that only touches the main question; in the Rankability data it was 27.6% against 14.5%. Google explains why that works for AI Overviews and AI Mode: those features use “query fan-out” to search several related subtopics at once, so a page that answers more sub-questions fits more of those searches.

Broad does not mean long. Write out the sub-questions a customer really has (cost, timeline, alternatives, who it is not right for, what can go wrong) and give each a short, standalone answer. Leave out what adds nothing.

Why should you show who is writing?

AI systems and Google trust content more when it is clear who wrote it, how it came about and why that person knows. In the Semrush analysis, E-E-A-T was the second strongest feature of cited pages (+30.6%), and Google explicitly recommends a byline where readers expect one in Creating helpful content.

So put a real name with a role under substantive pieces, explain where your numbers come from and mention your hands-on experience. How to build that authority outside your own site as well, with mentions on strong external sites and Wikidata, is in E-E-A-T for AI.

Original research: why does AI like citing your own data?

Original research is cited often because it contains information that exists nowhere else. A model that comes across ten repetitions of the same advice and one source with its own numbers has only one reason to choose you: the numbers. Google names this explicitly as a quality question too: does your content offer original information, research or analysis?

Four forms that work

• Surveys. A short questionnaire among customers or your market produces statements such as “[x]% of the contractors surveyed expect …”, which cannot be found anywhere else.
• Benchmarks. A systematic measurement across a larger group: load times of websites in your industry, prices of a service in your region, permit lead times.
• Cases with real numbers. Starting point, intervention, timeline and result. “More traffic” is vague; the concrete number with a period is citable.
• Your own named model. A named method or checklist others can refer to.

Research on a small budget

• Use your own data. Quotes, support questions, website numbers and order data often contain patterns you can share anonymized and aggregated. Respect privacy law.
• Analyze public data in a new way. Census, Bureau of Labor Statistics and other government data are free; your analysis is the added value.
• Poll your network. A LinkedIn poll or a short Google Forms questionnaire is not representative, but it is usable as long as you say so honestly.
• Partner with a trade association. It has the reach for a larger sample and often looks for relevant content for members.

How to make your research citable

1. Start with the most important finding, in the first paragraph, not on page 12 of a PDF.
2. State your method: period, number of respondents or data points, selection and limitations. Without that it is a claim; with it, it is a source.
3. Publish as a web page, with the PDF as an extra download. HTML is readable for every crawler.
4. Get others to write about it. Send your key numbers to trade publications and share them on LinkedIn. Every mention on another site is another place AI comes across your research.
5. Repeat the research yearly or quarterly. A trend line is cited more than a single measurement.

How often should you update content?

Update content as soon as the substance changes, and schedule a fixed check per content type. AI assistants cite somewhat newer content on average than Google: in an Ahrefs analysis of almost 17 million citations (July 2025), cited pages were 25.7% “fresher” than organic search results. ChatGPT had the strongest preference for recent content; Google AI Overviews actually cited slightly older pages than the regular results.

Freshness is a factor, not a miracle cure. In the same analysis, cited content was still about 2.9 years old on average. A strong, accurate page stays useful for a long time, as long as the facts in it are right.

Content type	Examples	Update
Explainers and reference	“What is …”, guides, glossaries	Check at least once a year
Trends and news	Market developments, new rules	Monthly or on a relevant event
Numbers and research	Benchmarks, statistics	As soon as there is new data
Product and price	Pricing pages, comparisons	Immediately at every change

What a good update is

• Adding new numbers or replacing outdated ones, with source and date.
• Adding a section about what changed since the previous version.
• Replacing old examples with recent ones.
• Checking prices, product names and links.
• Keeping the visible “updated” date and dateModified in your structured data equal to the last substantive change.

What not to do

Never change only the date. Google literally asks in Creating helpful content whether you change dates on pages to make them look fresh when the content has not substantially changed. Always tie a new date to a real improvement.

Keep a log of which page you updated when and how, and look over periods of at least a month at whether you appear more often in AI answers afterwards. An update is not picked up instantly.

How do you choose topics based on questions people ask AI?

Choose your topics based on the full questions customers ask an AI assistant, not only on loose keywords with search volume. Those questions are longer and include context. In Google someone types “crm small business”. In ChatGPT the same person asks: “I run a marketing agency with 15 people and I’m looking for a CRM under $50 per user that connects with our newsletter software. What do you recommend?” A page written only for the keyword misses the details needed to answer that question.

Keyword research stays necessary for Google. A question-based approach comes next to it.

How to find your customers’ questions

• Ask your salespeople and customer service. The questions that come back there are almost literally the prompts: price, comparison with competitors, fit for an industry, timeline.
• Ask the questions yourself in ChatGPT, Perplexity and Gemini. Note which sources are named, what is missing and where the answer is incomplete or wrong. Every gap is a topic.
• Look at “People also ask” in Google and at the long queries in Search Console.
• Let AI come up with variants: “Give twenty questions the owner of an HVAC company would ask ChatGPT about heat pumps.”

Group questions by intent

Intent	Example	What your page should offer
Informational	“How does a hybrid heat pump work?”	A clear explanation, step by step
Comparative	“Hybrid or all-electric for a 1930s home?”	An honest comparison on fixed criteria
Transactional	“Which installer in Denver can install within a month?”	Concrete scenarios, prices, lead times
Navigational	“How do I contact [company]?”	Correct contact details and business information

Prioritizing works differently

With keywords	With questions to AI
Monthly search volume	How often the question comes up in conversations with customers
Keyword difficulty	How weak the current AI answer is
Competition in search results	Whether good, citable sources already exist

Track ten to thirty questions tied to a buying moment, not a hundred. A short list produces decisions; a long list mainly produces noise. Which questions to choose is in which prompts to track, and the prompt tracker shows every day for which of those questions you or your competitor is mentioned.

What you do not need to do

You do not need special schema, an AI file or a trick. Google writes that you need no new machine-readable files, AI text files or special schema.org markup to appear in AI Overviews or AI Mode.

• llms.txt. An llms.txt is cheap and does no harm, but there is no proven effect on citations. Ahrefs saw that 97% of llms.txt files on 137,210 domains received not a single request. If you want one, see how AI crawlers read your site; do not expect miracles.
• Keyword stuffing. In the GEO research it lowered visibility.

One thing does have to be right: your most important content must be plain text in the HTML. The crawlers of OpenAI (GPTBot, OAI-SearchBot), Anthropic (ClaudeBot) and Perplexity (PerplexityBot) do not run JavaScript; Googlebot does render pages (Vercel). Text that only appears after scripts load is invisible to them. The technical scan checks that.

The writing plan in seven steps

1. Phrase the question this page answers, in your customer’s words.
2. Answer it in the first 100 words, naming the topic literally.
3. Back it up with facts you can verify, with source and date.
4. Split the page into H2s with questions and sections that each stand alone.
5. Cover the sub-questions a customer really has, without filler.
6. Show who is writing and where the knowledge comes from.
7. Schedule the next check and measure whether you get mentioned more often afterwards.

Start with the five pages closest to your revenue, not your whole blog.

Read also

• Content chunking: structure content so AI cites it
• E-E-A-T for AI: build authority search engines and AI recognize
• ChatGPT vs Perplexity vs Gemini: how they choose their sources
• What is GEO?
• GEO glossary