The paragraph lottery: how AI models pick what to quote
You publish a detailed, well-researched page. ChatGPT or Perplexity cites your site. You feel a small rush of excitement. Then you look at what they actually quoted, and it is one sentence from a paragraph buried halfway down the page, surrounded by everything else you wrote being completely ignored.
This is not random. It feels random, but it follows a pattern. Understanding that pattern is the difference between a page that gets quoted once in a blue moon and one that gets cited consistently, across multiple queries, for multiple sections.
Large language models do not read your page the way a human does. They do not start at the top, work through logically, and absorb everything with equal attention. They retrieve chunks. A chunk is typically a few sentences to a short paragraph, extracted during indexing and stored in a vector database. When a user asks a question, the model retrieves the chunks that score highest for relevance to that specific query, and it answers from those chunks. The rest of your page does not exist for that query.
This is why you see one paragraph quoted and nothing else. It is not that the AI disliked your other content. It is that your other content did not score highly enough for retrieval on that particular prompt. And in many cases, structural problems with the page made the situation far worse than it needed to be.
Why some chunks score high and others score zero
Relevance scoring in retrieval systems is built on semantic similarity. The chunk that best matches the meaning of the user's query wins. So the first question to ask is: are your paragraphs written as self-contained, answerable units, or are they written as part of a flowing narrative that only makes sense in context?
Narrative writing is great for human readers. It is punishing for AI retrieval. A paragraph that begins "As we mentioned above, this is why..." scores terribly in isolation because "this" has no referent. The chunk is meaningless without the surrounding context, and the retrieval system has no patience for that. It will skip to a paragraph that stands on its own two feet.
The paragraph that gets quoted is almost always one that:
- Opens with a clear, specific statement rather than a transitional phrase
- Contains a concrete fact, number, definition, or direct answer
- Could be read by someone who has never seen your page before and still make complete sense
- Uses language that closely matches how someone would phrase a search query or a question
Everything else on the page, written in the traditional long-form blog style of building arguments slowly and linking ideas together, tends to score poorly. It gets retrieved rarely, if ever.
The structural problems that make it worse
Even when your content is good, page structure can suppress retrieval across most of it. There are a few common culprits.
Walls of text with no internal signposting
A page with five 400-word paragraphs and no subheadings or internal structure gives the retrieval system very little to work with in terms of topical boundaries. The model struggles to identify where one idea ends and another begins. As a result, chunking becomes arbitrary, and most chunks end up blending multiple ideas in a way that reduces their relevance score for any single query.
Key answers buried inside longer paragraphs
This is extremely common. A paragraph starts with three sentences of context and framing, then delivers the actual answer in sentence four. The chunk that gets indexed may or may not include sentence four, depending on where the chunking boundary falls. Even if it does, the relevance score is diluted by the surrounding filler. Leading with the answer, then adding context underneath, is far more retrieval-friendly.
No schema to signal what the content is about
Schema markup gives AI crawlers explicit signals about the type of content on a page and how it is structured. A page with no schema is interpreted entirely from raw HTML, which means the model has to infer structure. A page with FAQPage schema, HowTo schema, or Article schema with clearly marked sections tells the crawler exactly what each block of content represents. That clarity directly improves the chances that specific sections get retrieved for the right queries.
This is a core part of what FlinnSchema works on with e-commerce and content brands: making sure the right schema types are applied so AI systems can parse and retrieve the right sections, not just whichever paragraph happens to score best by accident. You can see the kinds of approaches involved on the What We Do Differently page.
Entity density and why it matters for citation frequency
There is another layer beyond chunk structure: entity density. AI models weight content more heavily when it contains clear, named entities that match the topic of the query. Entities are things like product names, place names, people, standards, concepts with established definitions, and so on.
A paragraph that says "our solution helps businesses improve performance" has low entity density. Nothing in that sentence anchors it to a specific domain, product category, or query type. A paragraph that says "Shopify merchants using JSON-LD product schema saw a 34% increase in citation frequency across Perplexity queries in our March 2025 client data" has high entity density. It is packed with named things, a platform, a schema type, a metric, a source, a timeframe. That kind of paragraph gets retrieved constantly.
This does not mean stuffing your content with jargon. It means being specific. Named things, real numbers, real examples, real comparisons. Vague, generalised prose is retrieval-invisible regardless of how well-written it is.
How to rewrite pages so AI cites more of them
The good news is that most of this is fixable without rebuilding your entire site. Here is a practical approach.
Audit your paragraphs for self-containment
Go through each paragraph and ask: if someone read only this, would it make complete sense? If the answer is no, rewrite the opening sentence so the paragraph introduces itself. Cut transitional phrases that depend on what came before. Make each unit of content work independently.
Move answers to the front
For any paragraph where you are answering a question or making a point, put the answer in the first sentence. The explanatory context can follow. This is sometimes called the inverted pyramid structure, borrowed from journalism. It is also exactly what AI retrieval rewards.
Add subheadings that match real queries
Subheadings do double duty. They help human readers scan, and they act as topical boundaries for chunking systems. A subheading that mirrors a real user query ("How long does schema implementation take?" rather than "Timeline") dramatically improves the relevance score of the content that follows it. Think of each subheading as a micro-query that labels the chunk below it.
Implement the right schema types
For a content page, Article schema with clear headline, description, and author fields is the baseline. If the page answers questions, add FAQPage schema. If it explains a process, consider HowTo. If it covers a product or service, Product or Service schema is appropriate. The schema should reflect the actual structure of the content, not be bolted on as an afterthought.
If you are not sure which schema types apply to your pages, a Free AI Visibility Audit is a good starting point. It will show you where your content is being ignored and why.
Create a clear page-level summary
Near the top of the page, include a short, dense paragraph that summarises what the entire page covers. Think of it as an abstract. This single paragraph will often become the most-cited chunk on the page because it packs the highest concentration of relevant entities and answers into one place. Make it count.
What to do when AI cites you but traffic does not follow
There is a related problem worth naming here. Sometimes AI systems do cite your content, but you see no traffic from it. This happens for a few reasons: the citation appears without a clickable link, the answer is so complete that the user has no reason to visit your site, or the page being cited is not the page best suited to convert that visitor.
The solution is not to hide information in hopes of forcing a click. That strategy tends to backfire because AI systems simply cite a competitor who answered more fully. The better approach is to make your pages worth visiting beyond the information they contain: interactive tools, downloadable resources, calculators, booking flows, and so on. Give people a reason to arrive. Our post on why AI cites your business but sends no traffic goes deeper on this if it is a problem you are already experiencing.
You should also be tracking which AI platforms are sending traffic, because it varies more than most people expect. The guide to tracking ChatGPT and Perplexity traffic in Google Analytics will help you get that data into your reporting so you can see what is actually happening.
The bigger picture: writing for two audiences at once
The underlying shift here is that content now needs to work for two very different types of reader: humans who scroll and scan and appreciate narrative, and AI systems that retrieve in chunks and reward specificity and self-containment. These goals are not incompatible, but they require deliberate structure.
A page written entirely for human flow, with long paragraphs that build on each other, will be largely invisible to AI retrieval systems. A page written entirely as a set of isolated fact nuggets will feel robotic and off-putting to human visitors. The sweet spot is modular writing: each section self-contained and answer-first, with enough narrative connective tissue that the page still reads naturally.
Getting this right consistently across a site of any meaningful size takes real effort and a clear method. It is not a one-off fix. It is an ongoing approach to how content is structured, how schema is applied, and how pages are maintained as AI retrieval systems continue to evolve. That is the kind of systematic work FlinnSchema was built to help e-commerce and content brands do.
Frequently Asked Questions
Why does AI only quote one part of my page even when the rest is relevant?
AI retrieval systems break pages into chunks and rank each chunk separately against a user's query. Only the highest-scoring chunks get retrieved. If the rest of your content is written in a narrative, context-dependent style, those chunks will score poorly in isolation and will not be retrieved, even if a human would consider them relevant.
Does adding schema markup actually help AI quote more of my page?
Yes, because schema gives AI crawlers explicit signals about how your content is structured and what each section covers. Without schema, the crawler has to infer structure from raw HTML, which is less accurate. Appropriate schema types like FAQPage or HowTo help the system understand topical boundaries and retrieve the right sections for the right queries.
How long should each paragraph be for AI retrieval?
There is no single perfect length, but paragraphs of roughly 60 to 120 words tend to work well. They are long enough to contain a complete, meaningful answer but short enough that the chunk does not blend too many ideas. The more important rule is that each paragraph should open with its core point rather than building to it.
If AI is quoting me accurately, why am I not getting more traffic?
Citations without clicks are common when the AI answer is complete enough that users feel no need to visit the source. The fix is not to withhold information but to give visitors a compelling reason to arrive: tools, interactive features, personalised recommendations, or consultations. Making your schema clear also helps ensure the citation includes a link rather than appearing as an uncredited paraphrase.