Writing For Retrieval: The Passage Is The Unit

Quick answer

Retrieval systems index passages, not pages. Your article is cut into fragments before anything touches it, and each fragment competes alone, stripped of the argument that surrounded it. This makes one writing habit expensive: the backward reference. Every “as we saw above,” every “this approach,” every pronoun pointing at a noun three paragraphs back creates a chunk that cannot be understood in isolation and therefore cannot be selected. The fix is not shorter content. It is modular content — sections that carry their own subject, state their own conclusion, and survive being read alone.

I want to make a craft argument rather than a technical one. The mechanics of chunking and retrieval I have covered elsewhere. This is about what those mechanics do to how you should write, and it is a bigger change than the industry has absorbed.

For twenty years, good web writing meant a coherent journey. Establish context, build an argument, refer back, develop. Editors trained writers to do that. Style guides encode it. It produces genuinely better reading — and it produces fragments that fail retrieval.

What your page looks like from the inside

Take a well-written section from the middle of a typical guide and look at it the way a retrieval system does — with nothing before it and nothing after.

CHUNK AS RETRIEVED:

  "This is where the second approach breaks down. Because
  it relies on the assumption we discussed earlier, any
  change to that variable propagates through the whole
  model. In practice this means most teams find the third
  option more robust, though it costs more to set up."

What a reranker sees:
  - Second approach to WHAT?
  - Which assumption?
  - Which variable?
  - Third option out of what set?
  - The subject of this passage is never named.

Score: near zero for any query.

The paragraph is well written. It is also unusable
as evidence, because it is entirely composed of
references to things outside itself.

That passage might be the best paragraph on the page. Its quality is invisible, because quality is judged against a query and this fragment does not declare what it is about.

The four dependencies that break a chunk

Dependency	Example	Fix
Pronominal	“This means teams should…”	Name the antecedent: “Metric drift means teams should…”
Definitional	Uses a term defined three sections earlier	Re-state it in one clause, every section that uses it
Enumerative	“The third option” with no list present	Name the option: “Hybrid retrieval, the third option…”
Positional	“As discussed above” / “we will see below”	Delete, or replace with the actual claim

All four are marks of careful writing in a linear document and all four are fatal in an extracted fragment.

Definitional dependency is the one writers resist most, because re-stating a term feels like insulting the reader. In a linear read it does. In a retrieved fragment it is the difference between a usable passage and noise.

My position

I have come to think mild redundancy is now a feature rather than a flaw, and that is a genuine reversal for me. Editors spend careers removing repetition, and they are right about the reading experience. But a reader encountering your section as a retrieved fragment has not read the rest, will not read the rest, and needs the term defined here. The cost is a slightly heavier read for the linear reader. The benefit is existing at all for everyone else. I take that trade now, deliberately, and I did not five years ago.

The modular section, specified

What a section needs in order to survive extraction. Five properties, and a section meeting all five will be retrievable under every chunking strategy in common use.

A heading that names the subject, not the position. “Why cosine similarity cannot see depth” rather than “The second problem.” Headings frequently become chunk boundaries and chunk labels.
The subject named in the first sentence. Not referred to. Named. If someone read this paragraph on its own they would know what it concerns.
A conclusion stated, not deferred. The answer appears in the section, ideally early, rather than being resolved three sections later.
Any specialised term used is glossed in place, in a clause, without a full digression.
Length in the retrievable band. Roughly 150 to 400 words. Shorter fragments carry too little to score well; much longer ones consume a disproportionate share of a finite context budget and get dropped in favour of two tighter alternatives.

Before and after

BEFORE — good linear prose, poor chunk

  ## The second problem

  This one is subtler. As we established, the system
  can't distinguish between them, and that means the
  approach above tends to fail at scale. Most teams
  work around it by doing what we'll cover next.


AFTER — same content, survives extraction

  ## Cosine similarity cannot distinguish depth

  Cosine similarity measures the angle between two
  embedding vectors and normalises away magnitude.
  Because magnitude correlates with text length, a
  400-word overview and a 4,000-word treatment of the
  same subject sit at nearly the same angle. The
  retrieval layer therefore cannot see depth at all.
  Depth is judged later, at reranking, by a different
  model with different inputs.

Word count went up. Reading time went up slightly.
The chunk is now independently useful, correctly
labelled, and answers a specific question.

Note what changed. Nothing was simplified. The technical content increased. What went away was every reference to material outside the section.

Where modularity is wrong

I should be honest that this is not universally correct advice, because a great deal of writing exists for reasons other than retrieval.

Narrative and persuasion. An argument that builds, a case study with a turn, anything whose effect depends on sequence. Modularising it destroys the thing that made it work. Publish it anyway and accept it will not be cited.
Genuinely sequential instruction. Step four cannot be self-contained if step four requires step three. Number them, state the prerequisite in the step, and accept partial retrievability.
Writing for an audience already in the room. A newsletter to people who read the last one does not need re-explaining. Retrieval is not the goal.
Anything where the reading experience is the product. If people come for the prose, do not sand it flat for a reranker.

The uncomfortable part

There is a real homogenising pressure here and I do not want to pretend otherwise. If every page converges on self-contained, answer-first, moderately redundant sections, the web gets flatter. The distinctive voice, the argument that unfolds, the essay that rewards reading in order — those are penalised by this machinery. I write modular sections because I want the work to be found, and I am aware that is an accommodation to an indexing system rather than an improvement in craft. Anyone telling you these two things fully coincide is selling optimism.

A test you can run on any page in two minutes

The scissors test. Crude, fast, and it finds nearly everything.

Copy one section out of your page into an empty document.
Read it as though you have never seen the page.
Ask three questions: What is this about? What question does it answer? Is anything in it unresolvable?
If you cannot answer all three, that section is not retrievable. Repeat for every section.

Most well-regarded long-form content fails this on a majority of its sections. That is not an indictment of the writing — it is a mismatch between how it was built and how it is now consumed.

What this means for existing content

You are not going to rewrite an archive. So triage.

Page type	Action
Ranks well, never cited	Worth the rewrite. This is the pattern modularity fixes
Reference, glossary, spec table	Leave alone. Already modular. Do not “improve” it
Narrative or persuasive essay	Leave alone. Different job
New content	Write modular from the start. It costs nothing at draft time
Cornerstone guides	Fix headings and opening sentences only. Cheapest meaningful improvement

Row five is the highest return for the effort: renaming positional headings and rewriting each section’s first sentence catches most of the damage without a full rewrite.

The version of this I believe

Write each section as though it is the only part of your page anyone will ever read. Because for retrieval systems, that is literally true — and increasingly true for humans arriving at an anchor link from a generated answer.

That instruction is not a hack and it is not a trick. It is the discipline of making every part of your work independently defensible, which is a decent standard regardless of what indexes it. The uncomfortable part is that I arrived at it by reverse-engineering a machine rather than by getting better at writing, and I am not going to dress that up as craft.

The chunking and reranking mechanics behind this are in vector databases and embeddings and how retrieval selects your page. Why this creates two different competitions is in ranking and citation are two different processes.

Frequently asked questions

Why is the passage the unit of retrieval rather than the page?

Because documents are cut into chunks before embedding, and retrieval returns chunks. Your page as a coherent whole does not exist in the pipeline after that first step. Each fragment competes alone, stripped of the argument around it, and is scored on whether it answers a query by itself.

What makes a section fail retrieval?

Four dependencies on material outside itself. Pronominal, where this or it refers backward. Definitional, using a term defined in an earlier section. Enumerative, referencing the third option with no list present. And positional, phrases like as discussed above. All four are marks of careful linear writing and all four are fatal in an extracted fragment.

How long should a section be for retrieval?

Roughly 150 to 400 words. Shorter fragments carry too little content to score well against a query. Much longer ones consume a disproportionate share of a finite context budget and get dropped in favour of two tighter passages that between them cover more ground.

Is repeating definitions bad writing?

In a linear read, mildly. In a retrieved fragment it is the difference between usable and unusable. A reader encountering your section as an extracted passage has not read the rest and will not, so the term needs glossing there. Mild redundancy is now a feature, which is a genuine reversal of conventional editing advice.

What is the scissors test?

Copy one section out of your page into an empty document, read it as a stranger, and ask three questions: what is this about, what question does it answer, and is anything in it unresolvable. If you cannot answer all three, that section is not retrievable. Most well-regarded long-form content fails on a majority of sections.

When should I not write modular sections?

When sequence is the point. Narrative and persuasive writing whose effect depends on order, genuinely sequential instructions where step four requires step three, writing for an audience already following along, and anything where the reading experience is the product. Publish those as they should be and accept they will not be cited.

Which existing pages are worth rewriting for retrieval?

Pages that rank well but are never cited — that is precisely the pattern modularity fixes. Leave reference pages, glossaries and spec tables alone since they are already modular. Leave narrative essays alone. For cornerstone guides, fixing only the headings and each section’s opening sentence catches most of the damage cheaply.

Does writing for retrieval make the web more homogeneous?

Yes, and it is worth being honest about that. If every page converges on self-contained, answer-first, mildly redundant sections, distinctive voice and arguments that unfold across a piece are penalised. Modular writing is an accommodation to an indexing system rather than an improvement in craft, and claims that the two fully coincide are optimistic.