AI SEO glossary · Being quotable

What Is Chunking in AI Search?

Also called: passage retrieval, passage chunking

Written by Mike Kohl, Founder, Health Biz Scale·Updated October 28, 2026·SMART stage SCALE: get found and named

Chunking is the way an AI system splits a web page into short passages, called chunks, and then picks the passage that best answers a question. For a health practice, it means a tool often reads one slice of your service page and not the whole page. That slice is judged without the headings and paragraphs around it.

Key takeaways

  • ✓Chunking is the way many AI systems split a page into short passages and pick the one that best answers a question.
  • ✓A passage is judged on its own, so a pronoun like 'it' or a phrase like 'as mentioned above' means nothing once the surrounding text is gone.
  • ✓Google says snippets are created automatically from page content and are designed to preview the part of a page that best relates to a specific search.
  • ✓The most serious chunking problem on a practice site is two passages that disagree, such as a different visit length in the pricing section and the FAQ.

How chunking works

Picture a patient asking an AI tool how long a first functional medicine visit takes. The tool does not read your whole website the way a person would. In many systems, the page is cut into short sections ahead of time. When a question arrives, the system finds the sections that look most relevant and writes its answer from those.

A research paper from 2020 shows the basic idea. Lewis and colleagues paired a language model with an index of Wikipedia passages. The system retrieved passages first and wrote from them, and the authors reported more specific, diverse and factual language than a model working from memory alone. It is a research setup that describes the method, so read it as background on how retrieval works. The same approach sits behind the tools described under retrieval-augmented generation.

Search engines work with pieces of pages too. Google says snippets are automatically created from page content and are designed to preview the content that best relates to a specific search. Google also says that to appear in its AI features, a page needs to be indexed and eligible to show with a snippet. Vendors do not publish exactly how their AI tools cut pages, so treat any exact chunk size you read online as a guess.

What you can count on is the effect. A passage stands alone. Four things go wrong when it does.

  • The split fact. The claim sits in one passage and its proof, such as a price, credential or timeframe, sits in the next. The tool takes one and misses the other.
  • The orphaned reference. Phrases like "it," "this program" and "as mentioned above" mean nothing once the passage is lifted out.
  • The buried answer. The section opens with background and reaches the answer in the last line. The tool grabs the background.
  • The disagreement. Two passages on the same page give different facts, such as a visit length of 60 minutes in the services section and 90 minutes in the FAQ. A tool can quote the wrong one in a confident voice, which is the kind of error described under hallucination.

The fourth problem is the one practices miss most. A person reading the page top to bottom sees both numbers and shrugs. A tool sees one chunk.

This is why answer-first writing matters. It is the habit that makes each chunk usable, and the quality it produces is called extractability, meaning how easily a tool can lift a passage out.

Example: chunking at a health practice

This example is illustrative, not a transcript. A chiropractor in a mid-size city has a page on sciatica care. The page opens with a story about the founder, explains the anatomy in the middle, and mentions the first visit near the bottom.

Here is what to expect. It usually takes a while, and we will explain our approach as mentioned above.

Read alone, that paragraph says nothing. "It" has no meaning and "above" points to text the tool did not take. A competitor's plainer page has this section instead.

The first sciatica visit at [Practice Name] takes about 45 minutes. [Provider Name] reviews your history, tests movement and explains the findings before any treatment plan is set.

That block names the practice, answers in the first sentence and keeps the number next to what it measures. It is the easier passage to lift. It also keeps its limits in the same block, because the plan is set after the findings. A clinical qualifier that sits in a different section from the claim it limits can be lost.

Now the quiet failure. The chiropractor's FAQ says the first visit takes an hour. Two chunks, two answers. No one notices until a patient arrives expecting something different.

Different searches can pull different slices of the same page, so a section that works alone is ready for any of them.

It helps to picture the page as a stack of index cards. Each card gets shuffled into a pile with cards from every other site. A card that says who, what and how long wins its turn. A card that says "see above" waits at the bottom. Moving the key fact onto the card that needs it takes a few minutes per section, and most pages have only a handful of cards that matter, such as the first visit, pricing and the main condition or treatment.

It also helps the practice. The same edit that makes a chunk easy for a machine makes the section faster for a patient reading on a phone in a waiting room.

How to check yours in 60 seconds

  1. Open your best-ranking service page.
  2. Copy one section, from a heading through its paragraph, and paste it alone into a blank note.
  3. Read it as a stranger. Does it name the practice or provider? Is the answer in the first sentence? Does every number sit beside what it describes?
  4. Use your browser's find tool to search the page for "$", "minutes" and "weeks."
  5. Confirm each price, duration and count shows the same value everywhere it appears.

Any section that fails step 3, or any number that fails step 5, is your first fix.

What to do about it

  1. Put the answer in the first sentence under each heading, and make the heading a plain question or label a patient would use.
  2. Name the subject in the section. Write "The first visit at [Practice]" and not "it" or "we."
  3. Keep a claim and its evidence in one paragraph. A price, credential or timeframe belongs next to the thing it supports.
  4. Keep clinical limits with the claim. If a treatment helps some patients, keep that wording in the same block. Do not trim a hedge to make a passage shorter.
  5. Reconcile repeated facts. Make a list of every price, visit length and provider credential, then match every page to it.
  6. Fix small things first. Moving or rewording one sentence is often enough. See fact density for how many specifics a good passage carries.
  7. Recheck the section test after every edit, and again when prices or providers change.

Frequently asked questions

What is chunking in AI search?

Chunking is the way an AI system splits a web page into short passages, then picks the passages that best answer a question. Each passage is judged on its own. For a practice, that means the section about your first visit has to make sense without the heading and paragraphs around it.

How long is a chunk?

Vendors do not publish one size, and it differs between systems, so I would not plan around a number. Think of a chunk as a short section of a few sentences under one heading. If a section answers one question completely and names the practice, it works as a chunk in nearly any system.

Do I need to rewrite my pages for chunking?

Usually not. Most pages need small edits, not a rewrite. Put the answer in the first sentence under each heading, name the practice or provider instead of using pronouns, keep each number beside what it describes, and make repeated facts match. Clinical qualifiers should stay inside the same block as the claim they limit.

Is chunking only about AI tools?

No. Google says snippets, the short excerpts under a search result, are created automatically from page content and are designed to preview the part of a page that best relates to a search. A section that works alone helps in both places. The same habit helps a human skimming on a phone.

The bottom line

Chunking means an AI tool reads a short passage of your page, not the whole page, and judges that passage on its own. For a health practice, a plain section that names the practice, answers in the first sentence and keeps its numbers consistent can beat a better-looking page that buries the answer. The one thing to do is pick your best service page this week and test one section as a stranger would read it.

Health Biz Scale does this

We rebuild service and condition pages so each section stands alone as a clear passage, with one consistent set of facts across the site.

See how we do AI SEO →

Part of AI SEO for Health Practices: The Complete Guide. All terms