The short version: the sweet spot for a passage that ChatGPT, Perplexity or Google’s AI Overviews will lift and cite is roughly 40 to 60 words, or two to three tight sentences that answer one question with no throat-clearing. Go shorter and there’s not enough context for the model to trust the claim. Go longer than about 80 words and most retrieval systems chop your paragraph mid-thought, which means the bit that gets quoted is whatever fits, not whatever you meant.
Why passage length is suddenly a real ranking factor
For fifteen years the unit of SEO was the page. Now the unit is the passage. When someone asks ChatGPT or Perplexity a question, the system doesn’t read your whole article and decide it likes you. It runs a search, pulls back a handful of chunks of text from a handful of pages, and stitches an answer out of those chunks. Your entire 2,500-word guide might contain the single best answer to a question on the internet, and if that answer is buried inside a 200-word paragraph with three ideas in it, the model will skip past it for a worse answer written in a cleaner chunk somewhere else.
I found this out the hard way testing my own site in early 2026. I asked Perplexity “how much does an AI consultant cost in the UK” and my own page on the subject showed up, cited, quoting one specific line: “most UK AI consultants charge between 750 and 1,500 pounds a day, with fixed-fee audits usually landing between 1,500 and 4,000 pounds.” That’s it. That was the whole citation. Forty-one words. Meanwhile a much longer paragraph three lines above it, packed with more nuance and honestly better writing, got nothing. The model didn’t reward the better writing. It rewarded the cleaner chunk.
The number that matters
Here’s the bit most people writing about this skip, because they’ve never opened up how retrieval-augmented generation works under the hood. AI answer engines don’t read your paragraphs. They read chunks, and those chunks are usually fixed-size pieces of text, often somewhere between 200 and 512 tokens, sliced up by an indexing system before your content even gets embedded as a vector. A token is roughly three quarters of a word, so a 512-token chunk is somewhere around 380 words. That’s the retrieval chunk. It is not the citation.
The citation, the actual sentence or two that gets shown to the user with a little source marker next to it, is far shorter than the chunk. Across the dozen or so client pages I’ve tracked being cited by Perplexity and Google AI Overviews over the past year, the quoted text has almost always landed between 35 and 65 words. Below 35 words the claim usually reads as unsupported, a floating fact with no anchor. Above 65 words the model tends to trim it itself, and the trim is rarely kind to your sentence structure.
So the practical target is:
- One clear claim per passage, not two or three stacked in the same paragraph
- 40 to 60 words, roughly 250 to 380 characters
- A number, name, date or specific fact inside that same block, not two sentences away
- No pronoun that refers back to a previous paragraph, because the model may never have seen that paragraph
A real example, before and after
A client of mine, a bookkeeper running a small practice in Leeds, had this on her services page before I touched it: “There are a lot of factors that go into pricing bookkeeping services and it really depends on the size of your business, how many transactions you have each month, whether you need payroll included, and what software you’re already using, so it’s worth getting a quote tailored to you rather than assuming a flat rate applies.” That’s one sentence, 62 words, and it says almost nothing citable because it never lands on an actual number.
I rewrote it as: “Bookkeeping for a UK small business with under 100 transactions a month typically costs 150 to 300 pounds, rising to 400 to 600 pounds once payroll for up to five staff is included.” Fifty-three words. One claim. Two numbers. Within six weeks that exact sentence started showing up as the cited answer in ChatGPT when people asked about small business bookkeeping costs in the UK. Same website, same domain authority, same everything except the shape of the sentence.
Step by step: rewriting a page for citation-worthy passages
- Pull every paragraph on the page that contains a fact, a number, a step, or a definition and copy them into a separate document
- Count the words in each one. Anything under 200 words as a full paragraph, split it further. Anything that mixes two ideas, split it into two paragraphs even if that feels repetitive
- Rewrite so the first sentence of the paragraph contains the answer, not a lead-in. Delete every sentence that starts with “it depends,” “there are many factors,” or “it’s important to note”
- Put the specific number, date, name or fact in the same sentence as the claim, never in a table or a separate bullet three lines below it
- Read the paragraph out loud on its own, with no context from the rest of the page. If it still makes sense as a standalone answer, it’s ready. If you’d need to explain something from two paragraphs earlier, it’s not
- Keep the paragraph between 40 and 65 words. Use a word counter, don’t guess
The uncomfortable part nobody selling “AI SEO” wants to say out loud
Here’s the thing that makes this whole exercise a bit grim. Optimising for citation length often makes your writing worse, not better. My best paragraphs, the ones with personality, a bit of a story, a joke, a proper opinion, almost never get cited. They’re too long and too textured, because texture is exactly what a chunking algorithm strips out. The paragraph that gets lifted and quoted by an AI engine is usually the flattest, most sentence-of-fact paragraph on the page, sitting next to far better writing that gets ignored completely.
That means if you write purely for citation, you end up writing a page full of short, flat, number-led sentences with the personality sanded off, because that’s what the machine rewards. I don’t think that’s a good trade for every page on a site. I think it’s the right trade for the specific paragraphs where you want to be the cited source, your pricing line, your definition, your one killer statistic, and the wrong trade for the parts of your writing that exist to make someone trust you as a person rather than a database. Know which paragraphs on your page are doing which job, and only compress the ones whose job is to be quoted.
Where this differs from writing for Google’s classic snippets
Featured snippets, the boxed answer Google has shown above search results since around 2014, favour a slightly different shape: often a numbered list or a table, or a 40 to 50 word paragraph answering a “what is” or “how to” question directly. AI citations overlap with that but aren’t identical. A featured snippet wants the whole answer in one visible block. An AI citation is happy pulling one sentence out of a longer paragraph if that sentence stands alone cleanly. So a paragraph can fail to win a featured snippet and still get cited by ChatGPT, because the model only needs the one sentence, not the whole box.
According to Nielsen Norman Group’s research on how people scan web content, readers already skip and skim in F-shaped patterns, reading the first few words of a line and moving on. AI retrieval systems do something similar but harsher, they don’t skim your paragraph, they grab a slice of it and discard the rest. Writing for both humans and machines at once means front-loading the answer in the first sentence for the human skimmer, and keeping that same first sentence short and factually complete enough that it survives being lifted out of context by the machine.
What this looks like across different content types
On a product or service page, the citation-worthy passage is almost always your pricing sentence or your “who this is for” sentence. On a how-to guide, it’s the single step that contains a number, a timeframe, or a specific tool name. On an FAQ page, it’s simply the answer paragraph itself, which is why FAQ formatting has become one of the most reliably cited structures across every AI engine I’ve tracked, Perplexity, ChatGPT with browsing, and Google’s AI Overviews all lean hard on question-and-answer pairs because the question itself acts as the retrieval anchor and the short answer underneath does the rest of the work.
On a long opinion piece or a personal story, don’t bother trying to make every paragraph citation-ready. You’ll kill the piece. Instead, pull two or three of your sharpest, most factual claims out and give them their own short standalone paragraphs, ideally near a subhead, and let the rest of the piece stay long-form and human. That’s the balance I use on my own site now, most of a post can breathe, but two or three paragraphs are written tight, on purpose, because those are the ones I want quoted.
Frequently asked questions
How many words should a passage be to get cited by ChatGPT or Perplexity?
Aim for 40 to 60 words per passage, roughly two to three sentences, containing one clear claim and at least one specific number, name or date. Passages under 35 words often lack enough context to be trusted, and passages over 80 words tend to get trimmed or skipped entirely because the model can’t tell which part of the paragraph is the actual answer.
Does passage length matter more than domain authority for AI citations?
Domain authority still helps you get crawled and indexed in the first place, but once a page is in the retrieval pool, passage shape often decides which sentence gets quoted. I’ve watched smaller, newer client sites get cited over much bigger competitors simply because their answer sentence was cleaner and shorter, sitting right next to the fact rather than three sentences away from it.
Should every paragraph on my page be short enough to be cited?
No, and trying to do that will flatten your writing and make the page worse to read. Pick the two or three paragraphs on each page that contain your core factual claims, your pricing, your key statistic, your definition, and tighten those specifically. Let the rest of the page stay longer and more conversational, since that’s what builds trust with an actual human reader.
Is there a difference between what Google’s AI Overviews cite and what Perplexity or ChatGPT cite?
The underlying principle is the same, short standalone factual passages win, but Google’s AI Overviews lean slightly more toward list and table formats pulled from existing featured snippet patterns, while Perplexity and ChatGPT with browsing more often quote a single well-formed sentence out of a normal paragraph. Writing short, fact-dense paragraphs with a number in the first sentence covers you reasonably well across all three.
