What is content chunking, and why does it decide whether you get cited?

The short answer
Content chunking is the process of splitting a page into smaller pieces so a machine can retrieve one of them.
Here is the part most articles get wrong. Chunking is something engines do to your page. It is not something you do to your page.
You do not choose the chunk size. Google does, ChatGPT does, and neither publishes the number. What you control is whether your writing still makes sense after somebody else cuts it up.
Key takeaways
- The engine chunks, not you. Advice about picking a chunk size is written for people building their own systems.
- No major engine publishes its chunking rules. Anyone quoting an exact size for ChatGPT or Google is guessing.
- Retrieval happens at passage level. ChatGPT retrieves around 33 pages per prompt and cites 49.98 percent of them, scoring passages rather than whole pages.
- So the test is simple. Can any few hundred words of your page be cut out and still answer something on their own?
- Back-references are the main thing that breaks. “As we saw above” turns a liftable passage into a dead one.
Where this comes from. We run organic for more than 250 enterprises at Pepper and track over 10 million prompts across every major engine. We read a lot of pages that rank and never get quoted, and the cause is usually structural. That work shapes the view here, along with the client reviews we sit in and the talks we have at the events we run.
What is content chunking?
Content chunking is splitting a document into smaller passages so a retrieval system can find and use one piece without reading the whole thing.
It matters because that is how answers get built. An engine does not hand a model your entire article. It hands over a fragment. What a results page contains explains where those fragments end up.
Who does the chunking
This is the distinction that makes the rest of the subject clear.
- If you build a retrieval system, you choose the chunk size, the overlap and the splitting method. Most published advice addresses this reader.
- If you publish web content, you choose none of those. The engine decides, differently on each platform, using rules it does not share.
So the useful question is not what size to chunk at. It is how to write something that survives being chunked by rules you cannot see.

See where you show up. Pepper’s GEO platform tracks Brand Visibility, Domain Prompt Presence and Share of Voice across ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews. Your team can log in, connect Search Console and GA4, manage the prompt set and run your own agents in the Agent Atlas. A growth team works the same account alongside you. Book a growth audit or see where you show up.
What we actually know about how engines chunk
Less than the confident articles suggest, and the gap is worth stating.
What is documented
Retrieval is passage-level. Ahrefs analysed 1.4 million ChatGPT prompts and found the engine retrieves roughly 33 pages per prompt, then cites 49.98 percent of them. Being retrieved is about half the job. The other half is being the cleanest answer among the candidates.
Google has confirmed that AI Overviews and AI Mode run on core Search ranking and quality systems, with no separate AI index.
What is not documented
No major engine publishes its chunk size, its overlap, or its splitting method. The figures circulating, typically 256 to 512 tokens with ten to twenty percent overlap, come from open-source retrieval benchmarks. Those describe what engineers choose when building their own systems.
That is useful context and poor evidence about Google. We are stating the range so you can recognise it, and we are not building advice on it.
Why it still matters
Even without the exact numbers, the shape is clear. A few hundred words is the working unit. If your argument only makes sense across two thousand words, it will not survive the cut.

The test that replaces chunk size
One question, and it needs no tooling.
Can a reader who lands on any single section of your page get a complete answer to something?
If yes, that section can be lifted and quoted. If no, it depends on surrounding text the engine may never pass along.
What breaks a chunk
Four things, and we see all four weekly.
- Back-references. “As we saw above”, “as mentioned earlier”, “this is why the previous point matters”. Each one makes the passage depend on text that may be gone.
- Buried answers. A heading followed by three paragraphs of context before the answer. The chunk that contains the heading may not contain the answer.
- Pronouns with distant antecedents. “It does this well” is useless when the thing “it” refers to sat four paragraphs up.
- Tables as images. A table rendered as a picture carries no text to chunk at all. The technical audit covers how to spot this with a raw fetch.
How to write for chunking
Five rules. All of them are editing, not engineering.
- Answer in the first two sentences under every heading. Not the first heading. Every one.
- Make each section self-contained. Someone should be able to read it cold and get something useful.
- Repeat the subject instead of pronouns. Say “the crawler” again rather than “it”, at least once per section.
- Use real HTML for tables and lists. An image of a table is invisible to retrieval, and getting cited in AI Overviews covers what else blocks a passage from being used.
- Put a specific in each section. A number, a date, a named method. It gives the passage something to be cited for, which the citation evidence shows is what separates a quoted page from a passed-over one.
The read-it-cold check
Take any section of your page, cut it out, and read it with no surrounding context.
If it still answers something, you are done. If it reads as a fragment, the problem is almost always a back-reference or a buried answer, and both take minutes to fix.

Our methodology: how we weight chunking work
| Factor | Weight | Why it carries this much |
|---|---|---|
| Answer placement under each heading | 35% | The single thing that decides whether the right chunk carries the answer |
| Section independence | 30% | Back-references are the most common cause of an unusable passage |
| Machine-readable markup | 20% | A table as an image cannot be chunked at all |
| A specific in every section | 15% | Gives the passage a reason to be quoted rather than paraphrased |

Chunking approaches at a glance
| Approach | Who it is for | Typical 2026 cost | What it changes | Where it falls short |
|---|---|---|---|---|
| Answer-first editing | Publishers | A day per page | Whether the right chunk holds the answer | Reads slightly repetitive to a linear reader |
| Removing back-references | Publishers | Minutes per page | Whether a passage survives being lifted | Easy to reintroduce without a style rule |
| Real HTML tables | Publishers | A developer hour | Whether tabular content exists to retrieve | Needs doing once, then it stays fixed |
| Schema markup | Publishers | Low | Machine understanding of page type | A controlled study found minus 4.6% on AI Overviews |
| Choosing a chunk size | Engineers building RAG | Engineering time | Retrieval quality in your own system | You cannot apply it to someone else’s engine |
| Platform plus a growth team | Teams without capacity | Custom | Tracking, plus the editing it points at | Still gated by your publishing pace |
How to evaluate your own pages
The check costs an afternoon and needs nothing but a text editor.
The scorecard
Score a page out of 100.
| Factor | Weight | How to score it honestly |
|---|---|---|
| Every heading answered in two sentences | 30% | Someone has checked under all of them, not just the first |
| Sections survive being cut out | 30% | You have actually cut three out and read them cold |
| No back-references | 20% | Search the page for “above”, “earlier” and “previously” |
| Tables and lists are real HTML | 10% | Someone has viewed the page source and confirmed it, rather than assuming the CMS did the right thing |
| Each section carries a specific | 10% | A number, a date or a named method appears in every section, giving the passage a reason to be quoted |
The live test, over 90 days
Fix 30 prompts in your buyers’ words. Real ones, such as “what is content chunking”, “best data observability tool for a small team”, “how much does sales enablement software cost” and “how do I structure content for AI search”.
Run them on Google, ChatGPT and Perplexity. Record whether you were cited, and which passage was quoted.
Rerun the same set monthly for 90 days. Three readings is the shortest honest window. One is a snapshot, two could be noise, three shows direction. Keep the set fixed, because changing it resets the comparison.
That quoted-passage column is the whole point. It shows you empirically where the engine is cutting, which is the closest thing to seeing its chunking rules that you will get.
Two things to watch in it. First, the quoted passage is rarely the one anyone predicted, and that gap is the most useful feedback available. Second, compare the passage across engines. Where Google and ChatGPT quote different sections of the same page, you are seeing their splitting rules diverge, and it tells you which sections are doing work you did not know about. The free method for running this sits in how to measure AI search visibility without expensive tools.
Weak approach versus strong approach
The weaker approach: read that the optimal chunk size is 512 tokens, try to write sections of exactly that length, and change nothing about how the sections work. Engines ignore your intended boundaries entirely.
The stronger approach: forget size, and make every section independently useful. Then check what actually got quoted and edit toward that.
One optimises for a number you do not control. The other optimises for a property you do.
Red flags
- A stated optimal chunk size for Google or ChatGPT. Neither publishes one.
- Advice to insert artificial chunk breaks. The engine splits on its own rules, not your markers.
- Chunking sold as a technical service. On a published site this is editing work.
- Schema offered as a chunking fix. A controlled study found minus 4.6 percent on AI Overviews.
- Word count targets per section. No published evidence supports one for web content.
- Guaranteed citation from restructuring. Nobody controls generated output, and Google warns against guaranteed rankings.
Five questions worth asking
- “Can this section be read cold?” The fastest quality check available.
- “Where is the answer under this heading?” If it is in paragraph four, move it.
- “How many back-references does this page have?” A search for three words answers it.
- “Which passage actually got quoted?” Empirical, and more useful than any theory.
- “Is this table an image?” Still surprisingly common on older pages.
Reduced to one principle: write so any section can be read alone.
One closing note that costs us something. This is editing work on pages you already own. It is a smaller engagement than a content programme and it often removes the need for one, ours included in that general point.
What does this cost?
| What you are buying | Typical 2026 cost | What it covers |
|---|---|---|
| The read-it-cold check | Free, an afternoon | Which sections survive being lifted |
| Removing back-references | Minutes per page | The most common cause of a dead passage |
| Answer-first restructuring | A day per page | Moving answers to the top of each section |
| Converting image tables to HTML | A developer hour | Making tabular content retrievable at all |
| Tracking the quoted passage | Free, two hours a month | Empirical evidence of where engines cut |
| Platform plus a growth team | Custom | Pepper: tracking, plus the people and agents doing the work |
The first row is free and it tells you how much of the rest you need.
What nobody should promise you
An optimal chunk size for a production search engine. None of them publishes one.
Control over how your page is split. The engine decides, and it decides differently on each platform.
Citation from restructuring alone. Structure makes a passage liftable. The passage still has to say something worth lifting.
Guaranteed results. Nobody controls generated output, and Google warns against providers guaranteeing rankings.
When chunking is not your problem
The answer that costs us the sale. If your pages are not indexed, or they do not appear in search results at all, restructuring them will change nothing and we would tell you not to start here.
Retrieval comes first. A perfectly chunked page that nothing retrieves is cited zero times. Check crawler access and indexing first, which is free and covered in the technical audit. Come back to this once you are reliably in the candidate pool.
Frequently asked questions
What is content chunking?
Splitting a document into smaller passages so a retrieval system can find and use one piece without reading the whole thing. For publishers it is something engines do to your page, rather than something you configure yourself.
What is the best chunk size for AI search?
There is no answer you can act on, because no major engine publishes its chunk size. The figures in circulation, usually 256 to 512 tokens, come from open-source benchmarks for systems engineers build themselves.
Can I control how AI engines chunk my content?
No. The engine chooses the size, the overlap and the split points, and it does so differently on each platform. What you control is whether each section still makes sense once it has been cut out.
How do I know if my content chunks well?
Cut any section out of the page and read it with no surrounding context. If it still answers something useful, it will survive chunking. If it reads as a fragment, look for back-references and buried answers.
Why does chunking affect whether I get cited?
Because retrieval happens at passage level. ChatGPT retrieves around 33 pages per prompt and cites about half of them, scoring individual passages rather than whole documents against the question.
Do back-references really matter that much?
Yes, more than almost anything else structural. A phrase like “as we saw above” makes the passage depend on text the engine may never pass to the model, which turns an otherwise good answer into a fragment.
Should I add schema markup to help chunking?
Not for this. A controlled study of 1,885 pages found schema produced minus 4.6 percent on AI Overviews and no significant effect elsewhere. It remains worth having for rich results in classic search.
Is content chunking the same as topic clustering?
No. Chunking happens inside one page, splitting it into retrievable passages. Clustering happens across pages, grouping related articles around a subject. They solve different problems at different scales.
Where to go next
Open your most important page, cut out the third section, and read it with nothing around it.
If it still answers something, your structure is fine. If it reads as a fragment, you have found why a page that ranks never gets quoted, and the fix is an afternoon of editing.
What AI engines cite · Get cited in AI Overviews · Book a growth audit
Sources and further reading
Primary studies
- Ahrefs. Study of 1.4 million ChatGPT prompts, roughly 23.4 million cited and 23.4 million non-cited URLs. The engine retrieves about 33 pages per prompt and cites 49.98 percent of them, scoring passages against sub-questions rather than scoring whole pages. Link
- Ahrefs. Controlled study of 1,885 pages adding JSON-LD schema, matched against roughly 4,000 control pages using difference-in-differences. Google AI Overviews minus 4.6 percent and statistically significant, with AI Mode and ChatGPT indistinguishable from zero.
- Ahrefs. Analysis of 863,000 keyword SERPs and 4 million AI Overview URLs. 37.1 percent of cited pages ranked in the top 10 organic results, with 36.7 percent ranking beyond position 100, which is why passage quality matters more than position here. Link
Further reading
- Google Search Central. AI search optimisation guidance, 15 May 2026. Confirms AI Overviews and AI Mode run on core Search ranking and quality systems with no separate AI index, and warns against providers guaranteeing rankings.
- Pepper. What AI engines actually cite for the wider passage-level evidence, getting cited in AI Overviews for the Google surface specifically, and internal linking for how retrieval agents move between pages.
- Deliberately excluded: every chunk size, overlap figure and splitting method attributed to a production search engine, because none of them publishes these. The commonly quoted 256 to 512 token range comes from open-source retrieval benchmarks describing systems engineers build themselves, which is a different thing from how Google or ChatGPT split a web page.
Latest Blogs
Most advice on this is inference dressed as fact. OpenAI documents exactly one thing clearly, which is access. Here is what is documented, what is inferred, and how to verify the rest for yourself.
Being named in an AI answer and having your page used as its source are different outcomes with different causes and different fixes. The gap between them is the most useful diagnostic in AI search, and most reporting collapses it into one number.
Every platform reports share of voice and almost none of them define it the same way. Here is the definition worth using, the three choices that change the number, and how to measure it yourself without buying anything.
Get your hands on the latest news!
Similar Posts

Content
17 mins read
How to structure content for answer engines, when the structure is not the hard part

Content
12 mins read
How to write content that AI engines actually cite

Blogging
14 mins read