Open any guide on writing for AI search and you get the same checklist. Answer the question in the first two sentences. Keep your answer chunks to forty or sixty words. Use question headings. Add schema. Put the important part in the first third of the page.
All of that is reasonable. None of it is enough anymore.
It is reasonable because formatting makes a passage easier for an engine to lift. It is not enough because everyone now has the checklist. Following it perfectly gets you to the starting line. It does not get you cited, because the page next to yours did all the same things.
Here is the part the checklists skip.
What actually gets a page cited
AI engines cite passages, not pages, so content gets cited when it contains a specific, self-contained claim worth quoting. Formatting makes that claim easy to extract. Original evidence makes it worth extracting. Most teams perfect the formatting and skip the evidence, which is why their well-structured pages still do not get cited.
For the full picture of what the engines weigh when they choose, see how AI engines actually decide what to cite.
Sit with the two verbs there: easy to extract, and worth extracting. They are different problems.
Formatting solves the first. A forty-word answer under a question heading is easy for an engine to lift cleanly. Good. But an engine lifting your tidy, well-formatted sentence still has to decide it is the sentence worth lifting over every other tidy sentence on the topic. That decision is not about format. It is about whether your sentence says something specific and true that the other sentences do not.
What the pages that won actually looked like
This autumn I ran 19 buying questions through Google AI Mode, the kind B2B buyers ask at every stage, and read every source behind every answer. Some vendors owned their answers. Others watched competitors write them. The pages that won had five things in common, and none of them was clever formatting.
One question per page. When a security reviewer asked about Calendly, 9 of the 11 sources were Calendly’s own pages, most of them separate help articles: one on SOC compliance, one on where data is stored, one on where its AI assistant’s data lives. The SOC compliance article is a single sentence saying how to get the report. It was still cited, because it answers exactly one question. Vanta’s page on audit timelines is titled with the buyer’s question, “How long does a SOC 2 audit take?”, and opens with the ranges. Its answer led with Vanta. More on the security test here, and the rollout test here.
The answer includes its limits. The security answer took Calendly’s own word on what it does not do: data stored only in the US, no HIPAA business associate agreements. Vanta’s timeline page measures the wait in months, not weeks. Neither is flattering. Both were quoted, and the vendor’s own words framed the bad news.
The name and the number sit in the same sentence. Ramp’s customers page pairs each named customer with one result, and the answer reproduced it as a table. Another vendor’s customers page put results and company names in separate blocks, and the answer credited one customer with three numbers that sit on neighbouring cards. The full proof test is here.
Specific examples beat averages. One vendor’s headline average was set against third-party benchmarks and treated as marketing. Its individual case studies, each with a named company and a number, were treated as evidence in the same answer.
The page exists. This is the one that cost vendors most. Where a vendor had not published a page on the question, the answer came from someone who had: a competitor’s post on why the product takes six months to roll out, the incumbent’s guide to switching away, a rival’s security comparison. You cannot be quoted on a question you never answered.
The pass-through trap
Here is the mistake that quietly wastes the most effort, and almost nobody warns you about it.
You read that statistics get cited, so you add statistics. You write “according to [an analyst firm], X percent of searches will move to AI by next year.” Clean, specific, sourced. And when an engine wants that fact, it cites the analyst firm. Not you.
When you quote someone else’s number, you are a pass-through. You did the formatting work and handed the citation to the original source. The engine traces the claim to its origin, and the origin is not you.
You get cited when you are the primary source of the fact. That is the uncomfortable part, because being the primary source means you have to actually have a number. You have to have run something, measured something, counted something that nobody else has published. A page full of other people’s statistics is well-researched and perfectly citable, for everyone except you.
For a software company this is easier than it sounds. Your rollout time, your data residency, your customer results, your contract terms: you are already the primary source for all of them. The pages that won in the tests above were mostly that, facts only the vendor could state, stated plainly.
Write the one sentence you want quoted back
Practical version of all of this: before you write a section, decide the single sentence you want to see quoted in an AI answer.
Not the paragraph. The sentence. Make it specific, make it true, make it yours, and put it where the engine will find it, which is first. Then write the rest of the section to defend and support that one sentence.
This inverts how most people write. They write the full argument and hope a quotable line falls out of it. Instead, choose the line, then build the case for it. A section organized around one liftable claim gets cited. A section that buries its best sentence in paragraph four does not, because the engine never reached paragraph four.
Run the test on your own draft. For each section, ask: if an engine quoted exactly one sentence from this, which would it be, and is that sentence specific enough to be worth quoting? If the honest answer is “any of them, and none of them say anything only we could say,” the section is formatted for citation and has nothing to cite.
The method, per question
You already have the inputs from the earlier steps. Your buyer query map gives you the questions. Your gap analysis tells you who owns each answer today, and your content plan tells you which ones to build for. Now, for each one:
- Read the current answer and its sources. Ask the question the way a buyer would, several times, and note who the engine quotes. Note which part of the answer each source supplies. That tells you what your page has to say better.
- Decide the claim. What is the one specific, true thing you can say in answer to this question that the cited source does not? If you do not have one, that is the work. Go get the number, run the comparison, count the thing. The research is the moat, not the writing.
- Give it its own page or its own section, titled with the question. One question, one place. A page that answers six questions loosely loses to six pages that each answer one precisely.
- Put it first, limits included. Lead with a self-contained answer of roughly forty to sixty words that states the claim plainly, including the part that is not flattering. This is the part an engine lifts.
- Defend it below. Method, evidence, the named example, the caveat. This is for the reader who clicked through and for the engine deciding whether to trust the claim.
Format makes the page extractable. Original evidence makes you the one extracted. Do both, in that order of importance.
What did not work
A few things I overdid before I understood the difference.
Over-formatting. Early on I chunked everything into perfect forty-word blocks, added FAQ schema everywhere, and front-loaded every page. Citations barely moved. The pages were beautifully extractable and said nothing a dozen other pages did not. I had optimized the easy half and skipped the hard half.
Borrowing stats. I padded pages with credible third-party statistics to hit the “be specific” advice. It made the content read as authoritative and sent the citations to the sources I quoted. This post used to open with two of them. They are gone, because I could not trace either one cleanly back to its source, and because the argument above says I should not need them.
Writing for the whole journey instead of the unit. I wrote long, flowing arguments that built to a point. Engines do not read to the point. They lift a passage. Once I started writing each section around one liftable sentence, the same information started getting cited, because it was finally in a shape an engine could take.
What this step is for
In our method this is the second half of T in CITE, Target: once you know which questions to build for, this is how you write the page that wins them. It only pays if the steps before it were done right. The map tells you which questions matter. The gap analysis tells you which ones writing can actually win. This step makes the winning page, and the last step, Earn, works on the sources you do not own.
The order is the whole point. A perfectly written page answering a question no buyer asks is invisible. A perfectly formatted page with nothing original to say gets passed over for the source that has a real number. Get the question right, confirm writing is the fix, then write the one sentence only you can write.
That is the part you cannot buy a subscription for, which is exactly why it is the part that works.
If you want this run end to end against your own category, it is part of the engagement, and the full method lives on the methodology page.