Writing to Be Cited: How to Get Your Texts Quoted by AI

What Makes a Passage Citable to an AI

When ChatGPT, Perplexity, or Gemini answers a question, it assembles the reply out of passages pulled from several different pages. A passage is citable when a model can lift it almost verbatim, attach your link, and still be saying something true. That is a test of extractability, not of style: if a reader saw only this paragraph, with no heading above it and nothing after it, would they get a correct and complete statement out of it?

The gap between this and classic SEO is wider than it looks. Ahrefs found that only about six to eight percent of the URLs ChatGPT cites also rank in Google's top ten for the same query, and roughly 80 percent do not rank in the top 100 at all. Ranking earns you a slot in a list. Citation earns you a sentence inside the answer, and the two are selected on different grounds.

So write every passage to survive being torn out. Picture it pasted into a stranger's document with no heading, no paragraph before it, no graphic beside it. Does it still say something exact? Then it is citable. If it leans on 'as described above' or drifts into 'that depends on your situation' without ever landing, there is nothing to quote, and the model moves on to a page that landed.

Answer First: Build Question-and-Answer Blocks

The pattern is blunt: make the heading the question a real person types, then answer it completely in the first sentence underneath. Reasoning, example, and exceptions come after. Putting the result before the derivation is what people mean by answer-first structure, and it matches what a language model looks for: one statement it can use immediately, with the context underneath in case it needs more.

Take a flooring contractor. Instead of 'in this section we look at the question of screed drying times', write 'cement screed is normally ready to tile about four weeks after pouring, and we measure residual moisture before we start'. The first sentence is the quotable answer. The second says how the number gets checked, which is what stops it reading like a slogan.

Make sure the answer sentence leaves no precondition dangling. 'It depends on the individual case' is an evasion, not an answer. Name the normal case first and in full, then put the deviations in the paragraph below. A model will happily quote a standard case that is clearly marked as the standard case. So will a reader in a hurry.

Figures, Dates, Sources: Checkable Beats Persuasive

Checkable statements travel further than well-turned ones. A number, a date, a percentage, or a named source does more for you than any amount of phrasing. 'Lots of people search with AI now' is unusable. 'Pew Research Center, tracking the real browsing of 900 US adults in 2025, found people clicked a traditional result link in 8 percent of visits where Google showed an AI summary, against 15 percent where it did not' is citable, because it carries who, when, how many, and how much.

Put the source inside the sentence, not in a footnote or a caption. Running text is what gets chunked and quoted; the furniture around it often is not. Write 'according to Pew Research Center' or 'per the vendor datasheet' into the claim itself. It also protects you, because a model that carries your attribution along is likelier to carry your link with it.

Be honest about how solid each number is. If it is an estimate, write 'estimated'. If it moves, write 'as of mid-2026'. That reads less promotional, and it keeps the sentence usable after the figure changes, because anyone quoting it can see exactly what it was true of. Undated absolutes age into liabilities.

Structure a Machine Can Cut Cleanly

A retrieval system chops your page into chunks before it understands any of it, and what it quotes is a chunk, not a page. Real heading levels beat bolded sentences, one idea per paragraph beats a wall of text, and properly marked-up lists and tables keep their meaning once the styling is stripped. Alongside that sits structured data such as FAQ or Product markup, written in the shared vocabulary of Schema.org: a way to tell a machine which string is a price, which is an address, and which is an answer. Keep your expectations calibrated, though. Google says its AI features need no special markup, and schema has not been a direct ranking factor since 2018. What markup buys you is fewer garbled facts about your hours and prices, not a citation.

Keep each paragraph to one thought, short enough that a model could quote the whole thing. Long blocks bury the claim you wanted lifted, and a chunk cut mid-argument comes back sounding like something you never said. If a paragraph carries two claims, split it, or you will get quoted as half of each.

Cut references that only work on screen. 'See the chart on the right' or 'in the box below' means nothing to a system reading text alone, and nothing to someone hearing the page read aloud. Say the number in the sentence, then let the graphic repeat it.

Say It Plainly: Language That Survives Being Quoted

Irony, wordplay, and rhetorical questions work on readers who can hear your tone. A machine takes them at face value and can quote you saying the exact opposite of what you meant, with your name attached. Write literally what you mean, and be strictest in health, legal, and financial copy, where a sarcastic aside lifted straight turns into advice you would never give.

Define each term the first time you use it, in half a sentence, inside the sentence. 'Customer lifetime value, meaning the total revenue one customer brings over the whole relationship, is the number we optimize for.' Now the passage explains itself to a reader who does not know the term and to a model that needs a definition, which is one of the easiest ways to become the source that gets quoted.

Trade vague quantifiers for figures wherever you have one. 'Often', 'some', and 'as a rule' are the spots where your page declines to say anything checkable. When you genuinely have no number, name the condition precisely instead: not 'usually fast', but 'same day for orders placed before noon'.

Resolve Contradictions Instead of Leaving Them Open

Technical writing is full of apparent opposites: 'cloud is cheaper, and also more expensive'. A retrieval system cannot tell which half applies, so it grabs one at random or skips the passage. Name the condition and both halves become quotable: 'Cloud hosting is cheaper under spiky, unpredictable load, and typically more expensive than owned hardware under constant high load.' Each clause now carries the case it is true in.

Same discipline when you correct a myth. Lead with the true statement and label the false one as false, rather than opening with the misconception and rebutting it three sentences later. 'A charger left plugged in draws measurable power, though only a few cents of electricity a year' is safe to lift. A myth sitting alone at the top of a paragraph is not.

The rule underneath all of it: leave no question open that you could have answered. Every ambiguity you resolve is one more passage a machine can use safely, and one less place where it has to guess.

How to Tell Whether It Is Working

The honest test is the direct one: ask ChatGPT, Perplexity, or Gemini the exact questions your page answers, and watch whether your content turns up as a quote, as a linked source, or at least in substance. Perplexity shows its sources openly, which makes it the easiest place to check. Note which of your sentences get adopted word for word and which get paraphrased into mush.

Track it over weeks, not in one sitting. The same question gives different answers on different days, and the answers are unreliable narrators about their own sources: Columbia Journalism Review's Tow Center ran 1,600 queries across eight AI search tools and found more than 60 percent of the source attributions wrong. Watch the trend across repeated checks of the same core questions, never one screenshot.

Feed what you learn back into the text. When a model truncates or twists your answer, the phrasing is usually the culprit: the claim was not in the first sentence, or a condition was left hanging. Rewrite that passage, then check it again. Writing to be cited is a loop of phrasing, checking, and re-sharpening, not a single pass.

• Does the first sentence answer the heading completely, on its own?
• Is there at least one figure, date, or named source in the paragraph?
• Would the paragraph still be correct pasted somewhere with no context?
• Is every technical term defined where it first appears?
• Is every contradiction resolved and every ironic aside gone?

One Paragraph, Rewritten: Before and After

Take an ordinary marketing sentence apart. Before: 'For many years we have been the leading supplier of sustainable packaging in the region.' Nothing here can be quoted and stood behind. 'Many years' has no date, 'leading' has no measure, 'the region' has no borders, and 'sustainable' has no definition. The sentence delivers a mood, and a mood cannot be cited.

After: 'We have made compostable packaging since 2009 and shipped to about 340 wholesale customers across Germany and Austria in 2025.' Three checkable details now: a start year, a product category with a specific property, and a customer count with a year and a territory. Those are the parts a model can lift into an answer without inventing anything to hold them together.

The rule behind that rewrite: swap every evaluative adjective for a number, a date, or a place. If a claim will not break down into checkable parts, it is an opinion, and opinions rarely survive into an answer that has to name a source.

Five Steps to a Citable Page

One: before writing, collect your checkable facts in one place, the figures, dates, names, and where each came from. Two: turn the three questions your customers actually ask into the headings. Three: answer each heading in the first sentence beneath it, before any explanation, so the quotable statement always sits at the top of the chunk.

Four: read each paragraph and check it holds a single thought. If it holds two, split it, because a short self-contained block survives being cut and a long nested one does not. Five: read the whole thing aloud and delete every word that carries no information. What is left is denser, and denser text is easier to quote.

This works on any page type, a product page, a guide, an FAQ, a technical article. The cost is one extra editing pass, not a new content strategy. The payoff shows up over weeks rather than days, as your sentences start appearing inside answers instead of merely near them.

Limits and Misunderstandings

Writing to be cited is not writing for machines instead of people. A page a model can use reads more clearly for humans too: concrete numbers, honest structure, answer first. Google warns explicitly against producing separate content 'for AI', and treats content mass-produced that way as scaled abuse. Stack keywords or chop sentences into machine feed and you lose both audiences at once.

Second misunderstanding: nothing guarantees your sentence gets picked. You are moving a probability, not flipping a switch. Profound's look at roughly 730,000 ChatGPT conversations found only about 18 percent trigger a web search at all; the rest answer from training data, with no citation on offer. What you control is being usable when retrieval does happen.

And structure is no substitute for substance. If you have nothing precise, new, or useful to say, no amount of formatting rescues the page. Citability is the packaging, not the goods. Both have to be there before a text makes it into an answer, and the strongest lever sits off your own pages anyway: across about 75,000 brands, Ahrefs found that how often a brand is mentioned around the web correlates with AI citation roughly three times as strongly as backlinks do.