Most advice about writing for AI search stops at "write good content", which is the same advice as before and just as useless. The question is different now: not what makes a page worth reading end to end, but what makes one paragraph of it worth lifting out and repeating to somebody who will never see the rest.
The clearest public answer comes from someone who does it at scale. Aja Frost, senior director of global growth at HubSpot, walked through her team's on page checklist at GROW Europe 2025 (HubSpot Live). Seven points, and the last one is genuinely arguable.
What makes an AI quote your page rather than someone else's?
Two things, and both are mechanical. The engine has to find a passage on your page that directly answers a sub question it is working on, and that passage has to survive being cut away from everything around it. Almost every item on Frost's checklist makes one of those two more likely.
Key takeaways
- Answer engines split one question into many sub questions, so a page that answers only the headline question is covering a fraction of the work.
- They read in passages, not pages. A section that depends on the three paragraphs above it is a poor candidate for extraction.
- Put the complete answer in the first sentence of a section, then go deeper underneath it.
- Original data is the scarcest signal, and you probably already have some in your CRM and your sales calls.
- Frost's final point, tying every paragraph back to your product, is a real tradeoff rather than a free win, and we deliberately do it less than she recommends.
What is query fan out, and why does it change how you write?
When someone asks an answer engine a question, the system does not run one search. It breaks the question into related sub questions, answers each separately, then synthesises those partial answers into one response. Frost calls this query fan out, and the consequence is that your page competes to supply pieces of an answer rather than the whole answer.
So a page about choosing an accountant is not really competing on "how do I choose an accountant". It is competing on what qualifications matter, what the work costs, how long it takes, what happens if you switch mid year, and a dozen others nobody typed. Frost points practitioners at free fan out tools including Qforia, built by Mike King, to see what a given question expands into. Breadth of coverage inside one page now beats elegance of argument, because each sub question is a separate chance to be the source.
Why does an answer engine read passages instead of whole pages?
Because that is how retrieval works. The system chunks documents into passages, scores those passages for relevance to each sub question, and feeds a small number of the winners to the model that writes the answer. The model very often never sees your page, only a fragment of it.
That is the mechanism behind almost every recommendation below, and it is why a beautifully argued essay can lose to a plainer page with better internal packaging. We covered the retrieval funnel in why your business does not show up in ChatGPT.
What belongs in the first sentence of a section?
The complete answer, with nothing in front of it. Frost's first two points work as a pair: open with a sentence that fully answers the question in the heading, then go one click deeper with two or three short paragraphs of context, definitions or method. Engines want fast confirmation they are in the right place, and the depth underneath signals the answer is credible and complete rather than thin.
Business owners resist this because it feels like giving away the ending. It is exactly that, and it is correct. The reader who wanted only the answer got it and thinks better of you, and the retrieval system takes the one chunk that answers the question. A useful test: if the first sentence of a section is scene setting, the section starts in the wrong place.
Where do you find original data if you have never run a survey?
Inside your own business. Frost's third point is that engines hunt for net new information to cite, and that supplying it does not require commissioning research. Her two suggested sources are mining your own CRM for patterns and interviewing your sales and customer success teams for what they actually see.
This is the item most small businesses skip and the one with the least competition. A plumbing company that publishes what share of its emergency callouts trace to one specific failure, drawn from two years of its own job records, has something no competitor can copy. The bar for original is much lower than the word suggests.
Worth admitting this cuts at us. Of the seven points here, it is the one our own articles do least well, because we mostly cite other people's research rather than publishing our own.
Does a separate FAQ section still matter?
Yes, and for a different reason than it used to. Frost recommends at least three sub questions at the foot of the page, each under its own subheading with a short direct answer beneath it. The old reason was a rich result in Google that no longer appears the way it did; the new reason is that an FAQ block is the cheapest way to cover the fan out.
Each question is a self contained passage answering a distinct sub question, which is precisely the unit the retrieval system shops for. It also helps the reader, who often arrives with one of those exact questions.
How much structure is too much?
More than you would use in a document a human reads carefully. Frost describes answer engines as lazy readers and prescribes heavy formatting to suit them: bullets, explicit headers, tables and lists throughout, rather than saving structure for the parts that obviously demand it.
The trap is real, though. Content that is all structure and no substance reads like something a machine generated, and readers feel it immediately. The workable balance is to keep prose in the explanatory passages and give the extractable facts, sequences and comparisons a structural home of their own.
Can each section survive on its own?
That is the question to ask of every section before you publish. Because engines read in chunks, a section that relies on the ones around it will either be skipped or, worse, quoted in a form that misrepresents you. Frost's version is what she calls the Taco Bell test: if a human dropped into the middle of your page would need surrounding context to decode a reference, an engine certainly will.
In practice this means repeating things you were taught not to repeat. Name the subject again instead of writing "it". Restate the year, the product, the location. It reads slightly redundant to someone consuming the page top to bottom, and almost nobody does.
Should every paragraph mention your product?
This is Frost's seventh point and the one worth arguing about. Her recommendation is a product reference every paragraph or every other paragraph, and her reasoning is sound: content that never connects the insight back to what you sell makes the reader smarter without teaching anyone, human or machine, why your brand is relevant to the concept. If an engine learns your topic thoroughly and holds no association between that topic and you, you have donated training material.
The counterargument is equally real, and it is a matter of taste rather than fact. Content that ties every point back to the product reads like sales collateral, and readers discount sales collateral heavily. The credibility you spend buying that association is spent on exactly the audience most likely to buy something considered.
We have deliberately landed on the other side. These articles carry one mention of our own tool each, placed where it answers the question at hand, and that trades recommendation strength for trust. HubSpot's team believes the trade runs the other way, they have far more data than we do, and they may be right. What we will not do is pretend there is no tension, because a reader who notices a product mention in every paragraph notices what it is for. The version we endorse without reservation is narrower: make sure the concept and your name appear in the same passage somewhere, and make sure that passage is useful on its own.
What does this look like as a checklist?
Frost's seven points, in her order, applied to one page:
- Open every section with a sentence that fully answers its question.
- Follow with two or three short paragraphs of context, definition or method.
- Include original data you own, not only cited research.
- Add an FAQ of at least three sub questions, each with its own subheading.
- Use heavy structure: bullets, headers, tables, lists.
- Make every section survive being read in isolation.
- Connect the concept back to your product, and decide for yourself how hard to push it.
Two figures from the same talk explain why this is worth the effort. Frost told GROW Europe that AI Overviews now appear in nearly six in ten searches and halve click through rate when they do, and that more than half of buyers say they use AI to make purchasing decisions. She also reported, as HubSpot's own first party figure with no control group, that buyers arriving from AI answers convert at more than three times the rate of traditional visitors. Treat that multiple as a company reporting its own results, not independent research. Her explanation is the more useful part: an AI informed buyer has already built a shortlist and read reviews, so the visit is late in the process rather than exploratory.
Structure alone will not get you quoted if nobody outside your site ever mentions you, which is the other half of the problem and the subject of authority beats optimization. To see whether your current pages are extractable before you rewrite anything, SignalCheck scores a page across the technical and answer engine layers and shows what is missing.
Frequently asked questions
How long should the direct answer under a heading be? Roughly two to three sentences, or about 30 to 60 words. Long enough to fully answer the question without a follow up, short enough to be lifted whole into a synthesised response.
Do I need to rewrite my whole website? No, and it is a poor use of a first month. Pick the two or three pages that actually drive enquiries, rewrite those properly, then judge whether the rest is worth the effort.
Is an FAQ section still worth adding now that Google removed the rich result? Yes, for a different reason. The visual reward in classic Google search is gone, but the format still produces exactly the self contained question and answer passages retrieval systems select, and it covers sub questions your main sections do not.
What counts as original data if I am a small business? Anything true about your own operation that nobody else could publish. Job records, quote to close rates, seasonal patterns, the five questions your team answers most often. Frost's point is that mining your CRM and interviewing your own staff is enough.
Will writing this way hurt readability for humans? Slightly, in that it is more repetitive and more heavily formatted than an essay would be. It also helps the majority of readers who arrived with one question and wanted it answered before committing to anything longer.