Tactical

How do I write a first paragraph that AI will lift verbatim?

Last updated: September 14, 2026

Write the opener as an answer that survives being copied out of the page. Three rules do almost all the work. One, name the subject in the sentence, no "it," "this," "our platform," or "the company," spell out the actual name. Two, state the claim with a concrete number in it, because a specific figure is what a model quotes and what makes the passage look trustworthy. Three, make the paragraph stand alone, zero back-reference to a heading, a prior paragraph, or a "see above," so a model can lift it and it still reads. If a stranger who pasted only that paragraph into a chatbot would be confused, it won't get lifted. That's the whole test. Everything below is how to pass it.

1st ¶where the liftable answer goes. Not the intro, not paragraph four, the first thing on the page

Names + numbersthe two ingredients a model reaches for when it quotes a passage

Standalonethe passage must make sense with the rest of the page deleted

Why "be clear" is useless advice here

Every GEO post tells you to "write clearly" and "be concise." Fine, but that's not the mechanism. A model quoting your page isn't grading your prose. It's doing something closer to extractive question answering: given a query, find the smallest chunk of text that fully answers it on its own, and surface that. So the real question isn't "is my writing good." It's "can a machine cut one paragraph out of my page and have it stand up as a complete answer." Those are different targets, and optimizing for the first one gets you nice paragraphs that never get lifted.

The unit you're writing is not a paragraph. It's a quotable span. Write for that.

The three rules, and why each one is load-bearing

Rule 1: Name the entity in the sentence. Models resolve pronouns worse when a passage is torn out of context, and when they extract your paragraph, the context is gone. "It handles payroll for teams under 50" is dead weight, "it" who? "Gusto handles payroll for teams under 50" is liftable. The single most common reason a good answer doesn't get quoted is that it opens with a pronoun or "our product" instead of the name.

Rule 2: Put a number in the claim. A concrete figure does two jobs. It makes the passage specific enough to answer a real question ("under 50 employees," "in about 4 minutes," "for $39 a month"), and specificity reads as credibility to a model choosing between competing passages. Vague is skippable. "Fast setup" gets passed over; "setup takes about four minutes" gets quoted.

Rule 3: No back-references. "As mentioned above," "this approach," "the second option," anything that points at text the model won't carry along. When the passage is extracted, everything it points at vanishes, and the quote breaks. Write each candidate paragraph as if it's the only thing anyone will ever read from the page. Because for a lifted answer, it is.

Graded examples: watch a real opener get fixed

Same business, a payroll tool called Gusto, three drafts. Read them as a model would: could I quote this paragraph alone and answer "what is a good payroll tool for a small team"?

Grade: F, un-liftable

Welcome! We're passionate about making your back office simpler. In today's world, running a business is hard enough without wrestling with spreadsheets, which is why we built something better.

Names nothing (who is "we"?). No claim, no number, no subject a query could match. It answers no question. A model skims past it entirely. This is the default opener on most sites, and it's invisible.

Grade: C, better, still not quotable

It's an easy-to-use payroll platform that helps small businesses save time and stay compliant. Setup is quick and support is great.

Now there's a claim, but "it" has no antecedent once lifted, and "quick," "easy," "great" are unquantified. A model can't tell if this beats the next passage, because there's nothing specific to grab. Names + numbers are both missing.

Grade: A, liftable as-is

Gusto is a payroll and benefits platform built for small businesses, and it runs full-service payroll, including tax filing in all 50 states, in about four minutes per pay run for teams under 50 employees. Plans start at $49 a month plus $6 per person.

Subject named (Gusto). Category stated (payroll and benefits, small business). Concrete numbers a query can match (50 states, four minutes, under 50 employees, $49 + $6). Paste it into a chat cold and it answers the question with no page around it. That's the target.

Notice the A version isn't "better writing" in a literary sense, it's arguably flatter. It's built to be cut out and quoted, which is a different craft. Numbers must be true, obviously; a made-up figure is a fast way to get contradicted and dropped. Use real ones.

Where this comes from, so it isn't just my opinion

How we got here

The three rules aren't a style preference, they fall out of how extraction actually works. Research on passage segmentation for extractive question answering (arXiv, 2025) frames the goal as splitting documents into "self-contained" chunks, and its annotation guidance defines the answer span as the narrowest self-contained sub-string that fully answers the question, with the questions themselves written as standalone queries that "name their subject explicitly." Read that back into writing advice and you get exactly the three rules: self-contained (rule 3), fully answers on its own (rule 2, a specific claim), names the subject explicitly (rule 1). The people building the systems that quote your page are optimizing their extractors for precisely the passage shape these rules produce. You're just meeting the machine where it already is. One honest limit: production models are messier than a research benchmark, so treat this as writing for the shape extraction rewards, not a guarantee of a quote.

A 30-second checklist for any opener

Check	Pass looks like
Subject named?	The actual name is in the first sentence, not "it" or "we"
Claim present?	It asserts one specific thing, not a mood
Number present?	At least one concrete, true figure
Stands alone?	No "above," "this," "as mentioned," nothing pointing off-passage
Answers the query?	Pasted cold into a chat, it answers the obvious question

Run every important page's first paragraph through those five. The ones that fail are the ones getting skipped, no matter how well they rank.

The catch worth saying out loud

A liftable opener gets you quoted once a model is already reading your page. It does nothing to get the model to your page, that's a different job, done by being mentioned on the third-party sites the retrieval step reads. Answer-first writing and off-site mentions are two halves of the same win: mentions get you retrieved, a liftable passage gets you quoted. Nail the writing and skip the mentions and you'll have a beautifully quotable page that no model ever opens.

Want to see which of your pages AI can actually quote, and whether it's finding them at all? Pressfront's free AI Visibility Report checks the live engines against your category so you can go get on the sites AI reads.

Related entries

Pairs with What is co-occurrence, and why do mentions beat backlinks for AI?
Explains Why is my brand invisible to AI even though I rank on Google?
Prerequisite Retrieval vs training: which one decides whether AI mentions you?
Related How do I get mentioned on the third-party sites AI reads, without a backlink?
Compare with How consistent are AI recommendations, really?

Changelog

September 14, 2026: First published. Craft rules grounded in extractive-QA passage-segmentation research; three graded before-and-after openers and a five-point checklist.

Sources (primary): "Passage Segmentation of Documents for Extractive Question Answering", arXiv:2501.09940, 2025 (self-contained span definition and standalone-query framing). Extraction behavior in production models is messier than a benchmark, so this is writing for the shape extraction rewards, not a guaranteed quote. Engine behavior shifts; this entry is dated and revisited.