How to Write Content That AI Engines Cite

If you want to know how to get cited by AI engines such as ChatGPT, Perplexity and Google AI Overviews, start with two things: be retrievable for the searches the engine runs, and write passages that state a specific, verifiable answer early, backed by numbers, named sources and first-hand detail. Research on generative engines found that adding statistics, quotations and citations raised a source’s visibility by up to 40%, while keyword stuffing did worse than doing nothing. None of this needs a special AI format; it is clear writing that is easy to quote.

This page is about the content itself. For the wider discipline, see our overview of generative engine optimization, and for how engines pick sources before they read a word, see how AI search engines choose which sources to cite.

What makes content citable for AI engines

An AI engine that searches the web works in two stages. First it retrieves a handful of pages, usually through a search index. Then it reads parts of those pages and decides which sentences to rely on and credit. Citable content has to pass both stages: it must be found, and once found, it must contain a passage that answers the question better than the alternatives.

The first stage is mostly ordinary SEO. When Lily Ray tracked 11 site sections that lost Google traffic in early 2026, their average AI citations fell 22.5% across engines. Google says its AI features rely on the same core ranking systems as Search. If you are invisible in search, better sentences will not save you.

The second stage is where writing matters. Engines do not cite pages; they cite passages. The rest of this guide is about making your passages the ones they pick.

What the GEO research tested and found

The 2023 paper that coined generative engine optimization, by Pranjal Aggarwal and colleagues at IIT Delhi and Princeton (published at KDD 2024), tested nine ways of rewriting a source on 10,000 queries. The GEO paper measured how much of each answer was drawn from a source, weighted by position. Against a baseline score of 19.3, the results on its test set were:

Two caveats keep this honest. The “engine” was a simulation (top five Google results summarized by GPT-3.5), not live ChatGPT or AI Overviews. And the authors found that the best method varied by subject area. Treat the results as evidence of what language models prefer to quote, not as a recipe with guaranteed percentages.

How much of your page an AI engine actually reads

Engines have a limited reading budget. Dan Petrovic of DEJAN analyzed 7,060 queries and over 880,000 grounding snippets from Google’s Gemini grounding (December 2025). As summarized by PPC Land, the median query drew on about 1,929 words of source text in total, split across sources by rank:

Coverage fell as pages grew. Pages under 1,000 words had about 61% of their text used; pages over 3,000 words had about 13% used, and selection plateaued around 540 words per page. Long pages are not penalized, but most of a long page is never read for any single query. The raw data is not public and the analysis did not control for authority or structure, so read it as a strong hint rather than a law.

A separate analysis by Kevin Indig found that 44.3% of ChatGPT citations came from the first 30% of a page’s text, based on 18,012 citations, as reported by Practical Ecommerce in March 2026. Put the two together and the lesson is simple: what you say first, and what you say in each section’s opening lines, carries most of the weight.

What Google says about chunking and writing for AI

Google’s guide to optimizing for generative AI features (May 2026) lists several tactics it calls unnecessary for Google Search: llms.txt files, special markup, “chunking” content into small pieces, and rewriting content for AI. It says AI features rely on core ranking and quality systems, and it values unique, non-commodity content built on first-hand or expert insight. It also warns that creating separate pages for every sub-query variation can break its scaled content abuse policy.

At first glance this contradicts the research above. It does not, once you separate format from substance:

• Google is warning against artificial formats. Cutting an article into disconnected 50-word blocks, or writing a robotic version for machines, does not help and can make pages worse for readers.
• The research describes what gets quoted. Specific statistics, named sources and a direct answer near the top are what a reader wants too. They are good writing, not a separate AI format.
• Google is one engine. Its guide covers Google Search. Other engines read pages through their own fetchers, and the reading-budget data shows that the opening of a page and of each section matters wherever text is selected passage by passage.

So if you want to optimize content for AI Overviews and other engines at once, the reconciled advice is: write one page for people, with clear headings and sections that each make sense on their own, and put the answer before the explanation. Do not produce a second, chopped-up version for machines.

How to get cited by AI engines, step by step

1. Answer first, then explain

Open the page, and each H2 section, with one to three sentences that directly answer the question in the heading. Follow with the reasoning, evidence and exceptions. A passage that starts “It depends” is harder to quote than one that starts with the answer and then gives the conditions.

2. Add numbers with their source and date

Statistics addition was one of the strongest methods in the GEO paper. Use specific figures (“58% lower click-through at position one, Ahrefs, February 2026”) rather than vague claims (“much lower”). Name the source in the sentence and link to it. A number without a source is less useful to an engine that is trying to credit someone.

3. Quote and cite people who know

Quotation addition topped the GEO results. Attribute claims to named experts, standards or studies, and keep your own first-hand evidence visible: your test, your data, your customer count. This is also what Google means by non-commodity content.

4. Define terms in one sentence

Write definitions in the form “X is a Y that does Z.” Engines answering “what is” questions look for exactly this shape, and it helps readers who arrive mid-page from a citation link.

5. Use tables for comparisons and attributes

When you compare options, list specifications or show steps, use an HTML table or list instead of burying the data in paragraphs. Tables hold many facts in little space, which suits a limited reading budget.

6. Make each section stand on its own

Write headings as specific questions or noun phrases, and avoid sentences that only make sense with the previous section (“As mentioned above, this…”). This is not artificial chunking; it is the same principle that makes a good reference page easy to skim.

7. Cover the sub-questions

Engines split one prompt into several searches. A page that covers the main question and its natural follow-ups (cost, alternatives, risks, how to start) has more chances to match. Our explainer on query fan-out shows how to find those sub-questions, without creating a separate page for each one.

8. Skip keyword stuffing

Repeating the target phrase was the one tested method that scored below baseline. Use the words your readers use, once, in the right place.

A before-and-after passage

Here is a typical opening paragraph rewritten using the steps above.

The second version is the one an engine can lift and credit, and it is also more useful to a human reader.

How to check whether your content earns AI citations

Knowing how to get cited by AI is only half the job; you also need to confirm it is working. Content changes need measurement, and single checks mislead because AI answers vary from run to run. Run a fixed set of prompts on each engine several times before and after you publish, and track two separate numbers: how often your brand is mentioned, and how often your URLs are cited. Our guide to building a prompt set shows how to choose those prompts. Asheard, as one example, tracks mentions, citations, position and sentiment across ten AI engines and includes a free citability analyzer for individual pages; see the AI visibility tracking page. In Google, the Search Console generative AI reports show impressions for AI Overviews and AI Mode, though not clicks.

Frequently asked questions

How do I get my website cited by AI?

First make sure your pages rank and can be crawled, because most AI engines retrieve sources through a search index. Then write each section so it opens with a direct answer, supported by specific statistics, named sources and first-hand detail. Track citations across engines over several runs, since one check can mislead.

Do I need to chunk my content for AI search?

No. Google’s May 2026 guide says chunking content into small pieces is unnecessary for Google Search. What helps is ordinary clear structure: specific headings, sections that make sense on their own, and the key answer stated near the start. Write one version of the page for people rather than a separate version for machines.

Does adding statistics help content get cited by AI?

Evidence suggests it does. In the original GEO research paper, adding statistics raised a source’s visibility score by about 31% in a simulated generative engine, and adding quotations by about 41%. The test was not run on live ChatGPT or Google, so treat the size of the effect as indicative rather than guaranteed.

Does page length matter for AI citations?

Length itself is not a penalty, but engines read only part of a page. DEJAN’s grounding analysis found that pages under 1,000 words had about 61% of their text used, while pages over 3,000 words had about 13% used. Put the most important answers early in the page and early in each section.