How to structure a page so AI quotes it verbatim

Getting cited by an AI engine is not only about being right. It is about being liftable: putting a clear, self-contained answer where the engine will actually find it. Two forces decide this. Engines read the top of a page far better than the middle, and they quote short, complete passages rather than whole articles. This is a practical guide to structuring a page so the answer you want is the answer the engine can lift verbatim.

What "quoted verbatim" actually means

When an AI engine names a source, it rarely paraphrases a whole article. It lifts a short passage, often a sentence or two, that answers the question directly, and it attaches your page as the citation for that passage. So the unit that gets cited is not your page as a whole. It is a specific, quotable block on it. If that block is clear, complete and easy to find, you get named. If your answer is smeared across three paragraphs the engine has to stitch together, you usually do not.

That reframes the whole task. You are not writing to impress a reader who will scroll patiently from top to bottom. You are writing so a machine can locate one clean chunk, decide it answers the query, and paste it into an answer with your name on it. Two things govern whether that happens: where the chunk sits on the page, and whether it stands on its own. The rest of this guide works through both.

Where the answer goes: the top of the page wins

The single strongest pattern in the field data is positional. In an analysis of 1.2 million AI answers and 18,012 verified citations, published by Kevin Indig in Search Engine Land, the first third of a page produced 44.2 percent of ChatGPT citations, the middle third 31.1 percent, and the final third 24.7 percent. Indig described the shape as a ski ramp: citation likelihood is highest at the top and slides steadily downward. Where you place an answer materially changes the odds it gets quoted.

This is not a quirk of one dataset. It lines up with how language models handle long inputs. In "Lost in the Middle: How Language Models Use Long Contexts," published in the Transactions of the ACL in 2024, Liu et al. found that models use information at the start and end of their context most reliably, and that accuracy degrades for information positioned in the middle, a U-shaped curve. The field citation data and the lab finding point the same way: content near the top of a page is simply more likely to be used. Lead with the answer, and do not make the engine dig for it.

"The engine cites a chunk, not a page. Put the liftable answer in the first third, write it so it stands alone, and you have done most of the work."

Want to know which of your passages AI engines actually quote today? Get a free visibility audit and see exactly how ChatGPT, Gemini and Perplexity describe you.

How long the answer should be: short and self-contained

AI answers are short, and they are shaped like quotes. The Pew Research Center, in research published on 22 July 2025, found that the median Google AI Overview summary ran about 67 words, and that 88 percent of the summaries it examined cited three or more sources. An engine assembling a 67-word answer from several sources is not going to lift a 300-word paragraph from any one of them. It wants a compact, complete statement it can drop in whole.

The practical implication, and this is practitioner best-practice rather than a measured law, is to write the answer you want quoted as a self-contained block of roughly 40 to 70 words. That is long enough to state a fact with its important qualifiers, and short enough to be lifted intact. Treat the range as a heuristic, not a hard cutoff: there is no verified "ideal word count," and anyone quoting an exact number as a rule is guessing. The point is to make each answer a complete thought that survives being pasted somewhere else.

Why retrieval is the gate you have to pass first

Here is the part most page-structure advice skips. Before any model writes a word, a retrieval stage runs and selects a handful of passages, and the generation model typically only sees the chunks that retrieval surfaced. It does not read your whole page and reason over it. It reasons over the fragments it was handed. And most retrieved pages are never cited: the exact figure varies by system, but the pattern is load-bearing, so treat retrieval as a heavy filter rather than a formality.

That changes the order of operations. If retrieval does not select your passage, the model never sees it, and it cannot quote what it never received. So the first job of your page structure is not to persuade the model, it is to get the right passage picked up by retrieval. Clean chunk boundaries, a clear answer near the top, and headings that match the query all help the retriever find and select the passage that carries your answer.

Be honest about the evidence here, because retrieval behaviour is genuinely contested. Retrievers do show a primacy or position bias of their own, tending to favour the start of a document, as documented in work on position bias in modern information retrieval presented at EMNLP 2025 Findings. But a companion paper argues that the end-to-end effect on a full retrieval-augmented generation pipeline is smaller than the isolated bias suggests. The sensible reading is that position helps at the retrieval step, but it is one factor among several, not a magic dial.

The structure that gets quoted

Pull the findings together and a concrete page pattern falls out. None of these moves game an engine. They make the answer you already have easy to retrieve, easy to read in isolation, and easy to lift. This part is consensus craft among practitioners, so hold it as best-practice, not as something a study proved line by line.

• Lead each section with the answer. Open with a direct 30 to 50 word answer to the section's question, then follow with the supporting sentences. Do not warm up for three sentences before you say the thing; the engine, like the reader, wants the answer first.
• Make each section self-contained. A section should be readable on its own, without the paragraphs around it. Avoid unresolved pronouns like "this," "it" or "they" that only make sense in context, because the engine may quote the passage with none of that context attached.
• Keep paragraphs short. Two to four sentences gives the retriever clean chunk boundaries to work with. Walls of text blur where one idea ends and the next begins, which makes a single clean quote harder to extract.
• Use lists and tables. Structured formats are naturally easy to chunk and lift. A tidy list of facts or a small comparison table hands the engine discrete, quotable units instead of prose it has to parse.
• Write question-shaped headings. Phrase headings the way a user would phrase the query, so the retriever can match the heading to the question. A heading like "How long should the answer be?" mirrors real intent better than "On length."

The anti-patterns that keep you out of the answer

Most pages that fail to get quoted fail in the same few ways. They are readable enough for a human who is already committed to scrolling, but they give the engine nothing clean to lift. If you recognise your own pages here, these are the highest-leverage things to fix first.

• Burying the answer in the middle. A great answer three-quarters of the way down a long page sits in exactly the region the field data and the "lost in the middle" finding both show gets used least. Move it up.
• Pronoun-dependent sentences. A sentence that opens with "This means" or "They also" cannot be quoted alone, because the referent lives in a sentence the engine did not take. Name the subject explicitly in any sentence you want lifted.
• Marketing fluff the engine must interpret. Vague, adjective-heavy prose forces the engine to decide what you actually meant. A plain statement of fact gives it something it can quote without having to interpret you first.
• Walls of text with no chunk boundaries. One enormous paragraph offers the retriever no clean seams. Break ideas apart so each has an obvious start and end, and each becomes a candidate chunk.

A note on what this cannot promise

Structure improves your odds; it does not guarantee a citation, and it is worth being straight about the limits. There is no verified rule that engines extract "the first one or two sentences of every section," no reliable public number for the exact share of retrieved pages that get cited, and no evidence that adding FAQPage or any other schema makes an engine quote you. If you see those claims stated as measured fact, treat them with suspicion. The honest version is narrower: engines favour the top of the page, they quote short self-contained passages, and retrieval decides what the model ever sees.

Within those limits, though, the guidance is unusually actionable. You are not chasing an algorithm you cannot observe. You are writing clear, front-loaded, standalone answers, which is good writing for humans too, and arranging them where the engine reads best. That overlap is why this work compounds: the same structure that earns the citation also serves the reader who lands on the page directly.

The takeaway

To get quoted verbatim, stop thinking about pages and start thinking about liftable chunks. Put the answer you want in the first third of the page, where the field data shows 44 percent of ChatGPT citations come from and where language models read most reliably. Write it as a short, self-contained block of roughly 40 to 70 words, matching the compact, quote-shaped answers engines actually produce. Give the retrieval stage clean chunk boundaries and question-shaped headings so it selects your passage in the first place. Do that, and the answer the engine quotes is the one you chose.