Front-load the page, not the paragraph

Front-loading is the practice of putting definitions, findings and conclusions near the top of a page rather than building to them, and the largest public dataset on ChatGPT citations supports it. The trouble starts when teams apply the same rule one level down, to the paragraph, because at that granularity the same dataset says the opposite.

#What the page-level data shows

Kevin Indig's analysis of 3 million ChatGPT responses and 30 million citations, isolating 18,012 verified citations, used sentence-transformer embeddings to match each cited passage back to a specific sentence in the source document, then recorded where that sentence sat in the page. The distribution leans hard toward the top: 44.2% of citations come from the first 30% of content, 31.1% from the middle 30% to 70%, and 24.7% from the final third.

A companion analysis of roughly 98,000 citation rows drawn from about 1.2 million ChatGPT responses across seven verticals located the peak more precisely. The 10% to 20% section performed best across all industries, while the bottom 10% of a page earned just 2.4% to 4.4% of citations.

That part of the conventional advice survives contact with the evidence. The top of the document wins, the very bottom is close to dead weight, and burying your conclusion under 600 words of throat-clearing costs you.

#The paragraph-level finding runs the other way

Same study, one step down in granularity. Inside the paragraphs that did get cited, 53% of citations come from the middle of paragraphs, 24.5% from first sentences, and 22.5% from last sentences.

Put the two results side by side and the shape is not "earlier is always better." It is earlier in the page, later in the block. The cited material sits high in the document and in the body of the paragraph, not at its edges.

This matters because of what the standard remedy looks like in practice. A team reads "front-load," opens the CMS, and rewrites the article as a stack of two-sentence answer blocks. That page has a top and a bottom and almost no middle. It has been optimised for the sentence position that carries under a quarter of citations, by deleting the position that carries more than half.

#Why the middle of a paragraph is the part worth quoting

The reason is not mysterious once you look at what each sentence in a well-built paragraph does. The opening sentence is usually a topic sentence: a general claim, phrased in the same vocabulary as every competing page on the subject. It is interchangeable. The sentences after it carry the qualification, which is where the number, the named product, the version, the threshold and the exception live. That is the material that makes one page's answer different from another's, and it is the material a model needs when it is assembling a specific response rather than a generic one.

The linguistic profile of cited text points the same direction. Heavily cited passages averaged 20.6% proper nouns, against the 5% to 8% typical of English prose. Entity density that high does not come from a topic sentence. It comes from the sentence after the topic sentence, where you name who, what and under which conditions.

So does the tone measurement. Cited text clustered around a subjectivity score of 0.47, which Search Engine Land describes as reading like analyst commentary: fact plus interpretation. A bare assertion carries no interpretation. The interpretation is the follow-on sentence, the one that gets cut when you compress a paragraph to its claim.

Readability sits in the same range rather than at an extreme. Winning content averaged a Flesch-Kincaid grade level of 16, versus 19.1 for lower-performing content. Simpler than the material it beat, but nowhere near simple. Stripping prose down to headlines is not what the winning pages did.

#Short is not the same as extractable

The length data makes the point bluntly. Pages above 20,000 characters averaged 10.18 citations versus 2.39 for pages under 500 characters, with the biggest lift appearing between 5,000 and 10,000 characters.

Treat that as correlational, not causal. Long pages tend to sit on stronger domains, cover more subtopics, and therefore match more queries, and none of that is controlled for here. The honest reading is narrower: there is no evidence that cutting a page to fragments earns citations, and the pages that do earn them are substantial. Ranking still carries weight independently, since of pages ranking No. 1, 43.2% were cited by ChatGPT, 3.5x more often than pages beyond the top 20.

The retrieval mechanism explains why fragments underperform. Retrieval operates on chunks the engine cuts, not on your paragraph breaks, and chunk granularity is a live research question rather than a settled one. The Effect of Text Chunk Size on Retrieval-Augmented Generation Performance, submitted in June 2026, frames the trade-off directly: smaller chunks narrow the focus of each retrieval unit but "contain less information, which may limit the model's ability to generate coherent responses." A two-sentence block that gets chunked alone arrives at the model with nothing to support its own claim.

Selection pressure then does the rest. ChatGPT retrieved roughly 6x as many pages as it cited, and 85% of the retrieved pages were never cited. Getting retrieved is cheap. The cut happens at the point where the model decides which retrieved passage actually earns a place in the answer, and a passage with no substantiating middle has little to offer at that step.

#How to apply this

Move the answer up without moving the substance out. The definition, the finding and the recommendation belong in the first fifth of the page. The evidence for them belongs immediately after, in the same paragraph, not in a section three screens down.

Write paragraphs with three working parts: a claim, a substantiating middle that carries the specifics, and a short close. The middle is the citable part, so it should be the densest part. If a paragraph has no middle, it has nothing distinguishing to quote.

Drop "is the answer in sentence one" as your only structural check. It passes pages that have been hollowed out. A better check is whether the second and third sentences of each paragraph contain a named entity, a number with a source, or a stated condition.

Do not chase raw length. The length finding points at substance, not word count, and padding a thin page out to hit a character target adds noise the engine has to filter.

#Caveats worth stating

All of this measures ChatGPT, and citation behaviour differs by engine. It is observational rather than experimental, so it describes what cited pages look like, not what causes a citation. And the primary analysis is published behind a paywall on Growth Memo; the figures quoted here come from Search Engine Land's reporting of it, which is where anyone reproducing this should start.

The practical takeaway holds regardless. Position and compression are two different levers, and the advice to front-load only ever referred to the first one.