Article contents
A citable passage is a self-contained unit, not a fine paragraph
A citable passage is a block of text that a generative engine can extract from a page and reuse as is to answer a sub-query, without needing what comes before or after it. It is the second of the four conditions of a citation described in our method article: accessible, citable, identified, corroborated. This article deals with the second, at the scale of the paragraph.
The difference from a well-written paragraph comes down to one thing: autonomy. A text can be clear and accurate, and remain unextractable because its key sentence refers to “this point” or “as seen above”, which the model will not find in the fragment it keeps. Engines split pages into fragments of a few hundred characters before comparing them with the question; whatever is not in the fragment does not exist for them.
The structure we apply on our engagements has four parts, always in the same order: the question, the answer, the evidence, the caveat. It is not a style, it is a format constraint. What follows details it, then applies it to four types of page that are not rewritten the same way.
The four parts of a passage, and what each one prevents
Each part addresses a specific failure of extraction. Removing a single one is enough to drop the passage out of the selected sources, or to have it cited wrongly.
- 1QuestionThe heading, or the first sentence, restates the question in the user's words.Otherwise: the passage matches no sub-query.
- 2AnswerA complete, affirmative sentence that stands alone and is understood without the rest of the page.Otherwise: the model reconstructs the answer, and gets it wrong.
- 3EvidenceA figure, a date, a name, a method: something to cross-check the claim against.Otherwise: the claim is treated as an opinion.
- 4CaveatThe condition, exception or limit that says when the answer does not apply.Otherwise: the model generalises, or prefers a more cautious source.
The question does not need to be repeated in the paragraph if it already appears in the heading: the heading and the first paragraph together form the passage. When a paragraph deals with a question absent from the heading, its first sentence must contain it in affirmative form: “the cooling-off period for a contract signed at home is fourteen days” carries both the question and the answer.
The caveat is the part most often forgotten, and the one that sets an expert source apart from a generic one. A model comparing several passages on the same question readily keeps the one that specifies when the answer does not hold, because it will be able to use it without contradicting another source. In our audits, passages with a caveat are also the ones reproduced most faithfully.
What the model does with your paragraph
The journey of a paragraph from the page to the answer explains each of the rules above. It is described in broad outline by engine vendors, and is common to all retrieval-augmented generation systems.
- 01SplittingThe page is fragmented into blocks of a few hundred characters, most often at heading and paragraph boundaries.
- 02ComparisonEach fragment is matched against the sub-query; those that paraphrase it best rise to the top.
- 03SelectionThe fragment is kept if it answers on its own, with a verifiable element. A fragment that depends on context is discarded.
- 04ReuseThe model quotes literally or rephrases; the sharper the passage, the more faithful the rephrasing.
- 05AttributionThe claim is linked to your page, with your name.
Two consequences. The first: the boundary of the fragment is often that of the paragraph. An answer spread over three paragraphs is split into three incomplete fragments, and none is kept. The second: rephrasing is proportional to ambiguity. “Our lead times are short” will be rephrased at random; “delivery in five working days in mainland France” will be reproduced almost word for word.
Rewriting a service page
The service page is the most frequent case and the worst handled. It is written to convince a visitor already on the page, not to answer someone asking a question elsewhere. The result is positioning copy: no question, no figure, no condition.
The rewritten version answers the question “how much does an accountant cost for a small limited company”, gives a range, specifies the scope and flags the exception. The commercial tone has not disappeared, it has moved: precision has become the argument.
On a service page, we aim for one citable passage per service and per decision question: price, lead time, scope, guarantees. Each passage comes under its own heading, worded as the question it deals with.
Rewriting a product sheet
The product sheet poses the opposite problem: it contains many facts, but scattered between a written description, a specifications table and a “questions” tab. The model finds the figures, but not the sentence that links them to a use.
The second version does what the specifications table does not: it translates each figure into a condition of use. Power becomes a floor area, the sound level a siting rule, and the final caveat points to the right product. These are the sentences a model reproduces when asked “which heat pump for a 120 m² house”.
The table remains indispensable, in HTML rather than as an image, and preceded by this translating paragraph; its markup is covered in a dedicated article.
Rewriting an article and a local page
The in-depth article and the local page are two cases where the author has the material but delays it. The article opens with context, the local page with praise for the town. In both cases, the answer arrives in the third paragraph, too far for a fragment.
The in-depth article
The context is not deleted, it is moved after the answer. The human reader gains too: from the first sentence, they know whether the article covers their case.
The local page
The citable version names the districts, the hours, the prices and the surcharge. These are the elements a model cross-references with a question such as “emergency plumber Valencia price”, and the ones that third-party sources, directories and reviews, must confirm for the page to be corroborated. For businesses present in several languages, the requirement applies to each version, as shown in our article on multilingual content.
How long should a passage be?
The range of forty to eighty words is not a rule published by the engines, it is an observation. Below forty words, the evidence or the caveat is almost always missing; above eighty, the fragment spills over into the next paragraph or mixes two questions. The diagram below illustrates what we see on our panels.
Illustrative data. What to take from it: passages of forty to eighty words are over-represented among citations relative to their weight on pages; very short ones, the majority on sites, are rarely reproduced.
Length is counted in useful words: a sixty-word passage of which twenty are adverbs and hooks is worth a forty-word passage. Conversely, a longer passage remains extractable as a list or a table, because the splitting then follows the boundaries of the items.
The mistakes that make a passage unextractable
The faults below come up in almost every audit. None is a matter of writing quality: very well-written pages accumulate them, because they were designed to be read from beginning to end.
- A pronoun or a demonstrative in the first sentence (“This solution allows…”) that points outside the fragment.
- A conditional answer without the condition (“The lead time depends on several factors”): that is an absence of answer.
- A figure without a unit, a scope or a date (“30% savings”).
- Two questions dealt with in the same paragraph, so that neither has a complete answer.
- The evidence before the answer: three sentences of justification, then the conclusion, which the fragment boundary cuts off.
- A superlative in place of evidence (“the best value for money on the market”).
- The caveat relegated to a footnote or a tab, outside the fragment.
A citable passage is also a passage that can be held against you. Every published figure must be defensible, dated and revised when it changes. An honest range is worth more than a precise figure you cannot justify: the model will reproduce it just as readily.
A five-question review checklist
Before publication, each passage goes through the five questions below, in under a minute per paragraph, without tools, preferably by a reviewer who does not know the subject: they will read the passage as the model does, without the context.
| Review question | Test | Fix if the test fails |
|---|---|---|
| Does it stand alone? | Copy the paragraph into an empty document and read it without its heading. | Replace pronouns and demonstratives with the full subject. |
| Does it answer from the first sentence? | Read only the first sentence. | Move the answer sentence up; put the context after it. |
| Does it contain evidence? | Underline a figure, a date, a name or a method. | Add the verifiable element or, if none exists, rephrase as a qualitative observation. |
| Does it say when it does not apply? | Look for a condition or an exception. | Add the caveat sentence at the end of the passage. |
| Is it between forty and eighty words? | Count. | Cut the filler; split it if it deals with two questions. |
Table: scroll horizontally.
On a medium-sized site, we apply this checklist in three waves, starting with the pages that carry a purchase decision.
- Week 1Inventory
- List of decision pages: services, flagship products, local pages
- Questions actually asked: forms, customer service, panel
- Identification of existing passages and their faults
- Weeks 2 to 4Ten priority pages
- One passage per decision question
- Review with the checklist, by a reviewer from outside the subject
- Honest update date on every modified page
- Weeks 5 to 8Extension and control
- Articles and secondary pages
- Measurement of citation share on the panel, per engine
- Rework of passages rewritten but still absent
The third wave is the most instructive. Passages that have been rewritten but are still absent from answers almost always reveal a problem outside the text, an access block or a lack of corroboration. That is the moment to go back to the four conditions rather than rewrite a third time. The order of a page's sections, which decides where the model looks for the answer, is the subject of the next article.
What to remember
- A citable passage is a self-contained unit: it is understood outside the page, with no pronoun or cross-reference.
- Four parts, in order: the question, the answer, the evidence, the caveat. Removing a single one drops the passage.
- The fragment boundary is the paragraph boundary: an answer spread over three paragraphs is never kept.
- Each page type has its own fault: the service page positions, the product sheet scatters, the article and the local page delay.
- A five-question checklist is enough to validate a passage before publication.
Frequently asked questions
Should every paragraph of a page be turned into a citable passage?
No. Only the paragraphs that answer a question people actually ask need this structure, generally one per sub-question of the page. Transition, context or argument paragraphs remain useful to the human reader and do not need to be rewritten. Aiming for one citable passage per decision question, then checking with the review checklist, is enough in most cases.
Must a citable passage necessarily contain a figure?
No, but it must contain a verifiable element. A figure is the simplest; a date, a name, a condition or a described method play the same role. When no honest figure exists, a dated qualitative observation (“in the cases handled in 2025, the most frequent situation is…”) is worth more than an invented figure, which the model would reproduce and which you would then have to correct.
Does the question, answer, evidence, caveat structure harm human reading?
Experience shows the opposite. A reader who finds the answer in the first sentence knows immediately whether the page covers their case, and reads more of what follows. What harms reading is a page made entirely of identical blocks, with no transitions. The solution is to alternate: a citable passage under each question heading, then development text written normally.
How can I tell whether a rewritten passage is actually reproduced by the engines?
By asking the passage's question to ChatGPT, Perplexity, Claude and Gemini, several times, and noting whether your page is cited and whether the reproduction is faithful. A one-off check gives an indication; the real measurement is done on a panel of questions, every week, per engine, tracking citation share before and after the rewrite.
Do the same rules apply to an FAQ?
Yes, with one additional constraint: each FAQ answer must stand entirely on its own, without relying on the rest of the page, because it is often extracted in isolation and marked up as FAQPage structured data. The useful length runs from fifty to ninety words, direct answer first, caveat next, with no cross-reference to another section.
An SEO practitioner since 2001, Kamel Malek runs SEO360 (Alicante, Valencia, Madrid, Paris). He has published three books on visibility in generative engines, including Generative Engine Optimization and Rétablir les faits.
