The Google, OpenAI and Perplexity guidance checked for this article sets no word-count target for AI citations. Google says outright that there is no ideal page length. The useful distinction is between the length of your page and the passage that answers the question. An AI answer may quote or summarize that passage without the rest of the page, so the passage is what you edit.
A long page can hold a clear, self-contained answer. A short page can leave out the condition that makes its answer true. So judge the answering passage on its own: does it name its subject, state the answer, and carry the qualification and evidence a reader needs? Then decide what the rest of the page has to explain.
What do the engines actually say about length?
Google’s guide to generative AI features in Search (last updated 10 July 2026) states: “There’s no ideal page length, and in the end, make pages for your audience, not just for generative AI search.” The same guide adds: “There’s no requirement to break your content into tiny pieces for AI to better understand it.”
Its guidance on helpful content asks writers a self-check question and answers it in the same line: “Are you writing to a particular word count because you’ve heard or read that Google has a preferred word count? (No, we don’t.)”
OpenAI’s publishers FAQ, as checked on 24 September 2026, covered access and tracking. It told publishers who want ChatGPT to summarize their content not to block OAI-SearchBot, the crawler OpenAI uses for ChatGPT search. Neither that FAQ nor the OpenAI and Perplexity crawler documentation gave a word count.
The conclusion is narrow: these documents give you no number to write toward. That does not prove length never matters to any system or query. But if someone hands you a target, ask where it came from. Not from these documents.
Page length versus the passage an answer uses
Retrieval is the lookup step before an AI tool writes an answer from web sources. A system can retrieve a page and use only part of it. That is why the distinction matters when you edit: a reader may meet one claim from your page without the explanation around it.
Google describes a passage ranking system that identifies “individual sections or ‘passages’ of a web page to better understand how relevant a page is to a search.” Its AI guide adds that Google’s systems “are able to understand the nuance of multiple topics on a page and show the relevant piece to users.” Neither statement gives a passage length.
Research offers a concrete example. In the 2020 paper that introduced RAG models (retrieval-augmented generation: look up a source, then write), Patrick Lewis and colleagues split a 2018 copy of Wikipedia into 100-word chunks, 21 million of them. A chunk is a slice of text that is stored and searched on its own. Those chunks were the study’s unit of retrieval, not a recommended length for your pages or paragraphs.
Pew Research Center analyzed Google searches made by 900 US adults in March 2025. In that sample, the median AI summary was 67 words long, and 88% of summaries cited three or more sources. That is the length of the answers, not of the pages behind them. It does not tell you how long a source page should be or how much of it an answer will use.
None of these sources says how a commercial answer engine splits pages or how large its pieces are. The research examples show a principle, not any vendor’s settings.
Three decisions instead of a word count
For an editor, three decisions are more useful than a target length.
1. Where the answer sits. An introduction gets in the way when it delays the answer or leaves its subject unclear, and a short page can make that mistake as easily as a long one. Answer-first writing puts the usable conclusion, with its conditions, in the first screen and explains it afterwards. That is an editing choice for clarity, not a documented citation requirement.
2. Whether the conditions and evidence travel with it. A passage may be used without its neighbors. If the “except when” sits two paragraphs later, the claim can mislead on its own. Keep the main qualification beside the claim, and name the supporting source inside the passage. A footnote elsewhere serves the reader of the full page, but it can leave an extracted passage without its evidence.
3. Whether the page answers the next question. Google says its AI features may use “query fan-out”: the model generates several related searches at once to fetch more results. Its example starts from a question about fixing a lawn full of weeds and fans out into searches about the best herbicides for lawns, removing weeds without chemicals, and preventing weeds. When a follow-up question like that is useful to your reader, answer it in its own clearly headed section. Let that need, not a target, set the extra length.
That is not a license to build a page for every phrasing of a question. Google warns that creating separate content for query variations primarily to manipulate rankings or generative AI responses violates its scaled content abuse policy.
Clear sections also help the people who build retrieval systems. In The RAG Engineer I describe a chunk that covers twelve topics: a question about one of them matches it weakly, because the chunk’s meaning is spread across all twelve. That is a point about systems a team builds and tests. It does not tell you how ChatGPT or Google will process your page.
GRAIN: a builder’s view of the answering passage
GRAIN, from my book The RAG Engineer (available in English and Dutch), is a framework for building and checking retrieval systems. Its five skills help explain where an answer can lose its context. They describe a builder’s work, not the internal process of any particular search engine.
- Gather: choose the sources, cut them into pieces the system can store, and decide what belongs in the index (the searchable store) at all.
- Rank: find the candidate pieces that match the question, then sort them so the best come first.
- Assemble: pack the top pieces, with their sources, into the limited space the model reads before it writes.
- Inspect: when an answer is wrong, trace which step failed instead of blaming the model.
- Navigate: keep answers fresh by noticing when a source has changed and a stored piece is out of date.
Take a hypothetical project-management software company with a 2,500-word billing page. A customer asks: “Can I cancel an annual plan halfway through the year?”
At Gather, a builder might split that page along its headings. At Rank, a section headed “Canceling an annual plan” that opens with “Yes, you can cancel at any time and we refund the unused full months” is a more explicit answer than a section called “Our billing philosophy”. If it is selected at Assemble, that section enters the context the model reads before it answers.
If the exception (“the discounted launch plan is non-refundable”) lives elsewhere and is not retrieved, the model may answer without it, and be wrong for exactly the customers the exception was written for. Inspect means tracing that gap back through the system. Navigate means keeping the stored policy current when it changes; a visible “last updated” date on the page does not show that the retrieval system has refreshed its copy.
The editorial repair is one moved sentence: put the exception beside the refund rule. The passage becomes complete, and the page gets no longer. The builder still has to test retrieval and freshness. The framework has its own page: GRAIN.
What the GEO study measured, and what it did not
GEO stands for generative engine optimization: work aimed at making a source more visible in AI answers. A 2023 paper by Pranjal Aggarwal and colleagues, accepted at KDD 2024, proposed the approach and tested edits to source content.
The researchers built GEO-bench, a benchmark of 10,000 queries. Their simulated engine took the top five Google results for a query and had GPT-3.5-turbo write an answer from them. The paper also included an evaluation on Perplexity.
Here is the distinction that matters for word counts. The paper’s “Word Count” metric is not the length of your page. It measures your source’s share of the AI answer: the words in answer sentences that cite the source, relative to the whole answer. “Position-Adjusted Word Count” also gives more weight to citations earlier in the answer. Neither metric is a recommended page length.
The researchers tested nine edits. They reported roughly 30 to 40% relative improvement on Position-Adjusted Word Count for three of them: adding citations to credible sources, adding quotations, and adding statistics. Keyword stuffing gave little to no improvement. None of the nine methods was a length change for its own sake. Adding a quotation also adds words, but no method changed length alone, so the study cannot credit length with any gain.
These are results from the study’s setup, not predicted gains for your site. The simulated engine used a 2023-era model, and the Perplexity evaluation does not show how today’s systems select sources. What the study supports is attention to the substance of an edit. It supplies no ideal word count.
The passage test: copy the answer out alone
Pick one question your page should answer, phrased the way a reader would ask it. Copy the answering paragraph or short section into an empty document, with its heading if it needs one. Leave the introduction and the neighboring sections behind. Then read the copy against five checks.
| Check | Pass | Fail |
|---|---|---|
| Names its subject | ”Annual plans can be canceled at any time." | "It can be canceled at any time.” |
| States the answer | ”Yes, and unused full months are refunded." | "It depends on your situation.” |
| Carries the main exception | ”Except the discounted launch plan.” | The exception is two sections away. |
| Names its evidence | A named source, or a number you can trace to its origin | ”Studies show” |
| Stands alone | Makes sense without the paragraph above | Needs the introduction to make sense |
Fix each fail where it occurs: name the subject, move the qualification up, name the source. Add words only where information is missing, then read the copy on its own again. Five passes mean the passage can stand alone. They do not make the claim true, and they do not guarantee that an engine will retrieve, quote, or cite it.
Try this today (15 minutes)
- Open the page you most want cited. Write down the one question it should answer.
- Run the passage test on that question. Fix the fails.
- Put the answer in the first screen, under a heading that makes its subject clear. Keep its main qualification and its source with it.
- Write down the next question a reader would ask. If it belongs on this page and is still unanswered, add a section for it. If the page already answers it, run the same passage test on that section.
Stop when the answer, its qualifications, and the useful follow-ups are complete. The word count is whatever that took. If the same claim is scattered across several of your pages, see content architecture for answer engines. If you are not sure ChatGPT can access the page at all, start with how to get cited by ChatGPT.
Cite this:Is there an ideal word count for AI citations?Len P. van der Hof. https://lenvanderhof.com/en/blog/ideal-word-count-for-ai-citations/ ·