The AI Citation Playbook: 5 Structural Patterns Every Brand Winning AI Search Has in Common
Six features separate cited B2B pages from invisible ones. The five patterns that turn them into a library, and what they returned across 73 client articles.
machina sculpsit · homo probavit · MMXXVIIn 1801 Eli Whitney stood in front of a congressional committee, tipped out a bin of musket lock parts, and assembled working firearms from whatever his hand landed on. Interchangeable parts. The room was astonished and the contract followed.
He had fitted the parts in advance and marked them. The demonstration was rigged, Whitney never delivered true interchangeability, and the armories at Springfield and Harpers Ferry spent another forty years actually building it.
The idea was right anyway, and it reorganized industry for a century. Standard parts, assembled the same way every time, beat bespoke craftsmanship at scale.
Whitney was wrong about his muskets and right about everything after them.
94% of B2B buyers now use AI across every stage of the purchase journey (Forrester, 2025), and the pages those engines cite are built the same way: from a small set of standard parts, assembled in the same order, across an entire library.
Six Features Separate Cited Pages From Invisible Ones
Every one of the six appears in a majority of top-cited pages. Every one of them appears in zero of the bottom fifty.
We counted the parts on 852 B2B pages across 460 queries and 115 product categories. Six structural features show up in most of the top fifty cited pages, and in none of the bottom fifty (Res, 852-article B2B citation structure study, 2026).
The six are not exotic. A labelled block, a table, a set of steps, a price, a review, a definition. Any competent editor could add all six to a page in an afternoon, which is what makes the zero so strange.
A zero across six independent features is not a gradient. Whatever the bottom fifty were doing, they were not assembling from these parts at all.
Six features. Fifty pages. Zero hits.
That is not a content quality problem in the usual sense. Those fifty pages may be well written. They were simply not built out of anything an engine could pick up and carry.
The count matters as much as the kind. Pages in the longest quartile carry 13.55 structural elements each against 2.98 in the shortest (Res, 852-article study, 2026), a 4.5-fold difference in how many separately extractable things sit on the page.
The reason any of this pays is that the buyer stopped searching and started asking. A page of ten blue links gave ten brands a seat and let the reader sort them out. An answer returns three names and sorts them for her, so the job is being one of the three rather than being findable somewhere on a list.
Five patterns turn that count into a library. They compound, and they are not interchangeable with each other.
The Winners Answer Six Questions Below the Fold
Every question with its own heading is a separate thing an engine can lift. A page carrying seven of them offers seven chances to match a buyer’s phrasing instead of one, which is why our own client libraries run an FAQ block on 99% of published articles at a median of seven questions (Res client program profile, 73 articles, 2026).
The engines do not read a page and decide it is good. They look for the passage that answers the question that was typed. A block of eight self-contained answers is eight passages.
An essay carrying the same information dissolved through it is none.
Write them as real questions a buyer asks out loud, with the answer in the first sentence. Nobody types “understanding fence material considerations.”
This is where most libraries quietly fail. A team told that FAQs earn citations writes eight questions it wishes buyers asked, in the phrasing the brand uses internally, and the block matches nothing. The questions have to come from somewhere real: sales calls, support tickets, the search box, the thing the customer said on the phone before anybody wrote it down.
That is also the part a model cannot invent for you. It can shape an answer once it knows the question, and it has no idea which questions your buyers actually ask.
The Winners Publish the Incumbent’s Price
Pricing grids appear on 62% of top-cited pages and 0% of the bottom fifty (Res, 852-article study, 2026), and the reason is that a buyer asking about price gets an answer built from pages that contain prices.
A pricing grid is not a price list. It is a small table that puts the number, what it includes, and what it excludes in adjacent cells, so an engine can lift a comparison rather than a claim.
The uncomfortable version of this pattern is that it means publishing your competitor’s number next to your own. Most vendors will not. That reluctance is the opening. When the incumbent’s comparison page carries no pricing data and yours carries both, the engine has exactly one page it can build a pricing answer from.
One page. No competition for the slot.
Attribution Is the Cheapest Lift Anyone Has Measured
Adding statistics with attribution lifted AI visibility 41% in the Princeton GEO-bench experiments, the largest single-tactic gain in the study (Princeton KDD, 2024). Not adding statistics. Adding them with a named source and a year attached.
An engine assembling an answer has to decide which page to trust on a contested number. A figure with a publisher and a date behind it can be checked.
A figure floating loose in a sentence cannot, and the page carrying it becomes the riskier thing to quote.
There is a ceiling on this. Six well-placed citations across an article reads as sourcing; twenty reads as a research paper nobody finishes, and the engine gains nothing from the extra fourteen. Attribution belongs to the claim that needs it.
A reader can tell the difference between a page that sourced its numbers and a page that went shopping for citations, and so, it turns out, can the thing reading it on the reader’s behalf.
Our own client libraries carry attributed third-party citations on 99% of articles, at a median of six per article (Res client program profile, 73 articles, 2026). That is not a stylistic preference. It is the part that makes the rest of the page quotable.
One Template, Run Across Every Competitor
One page is a page. Eighteen pages built from the same parts is a category position.
This is Whitney’s actual idea, and it is the pattern with the steepest compounding. Rippling built one comparison template and ran it across its named competitor list: 18 comparison pages carrying 8 FAQs each produces 144 independent FAQ citation targets (Rippling, 2026), from one structural decision made once.
One page is a page. Eighteen built from the same parts is a category position, because the engine meets the same shape whichever competitor the buyer names.
The template is also what makes the library maintainable. When every comparison page carries the same five parts in the same order, updating a price or a competitor claim is a known edit in a known slot across eighteen pages, instead of eighteen separate archaeology projects.
The compounding is real, and it has a floor. Below four to six pages there is not enough repetition for an engine to meet your structure more than occasionally.
One page is an experiment. Six is a position.
The Loop Closes When You Know Which Prompt Paid
The fifth pattern is the only one that is not on the page. Tally captures which AI prompt drove each new signup through a single post-onboarding question, then builds the next comparison page against whatever gap that surfaces, a loop it credits with 6,000 to 10,000 new weekly registrations from AI engines (Tally, April 2026).
One question, asked at the moment somebody converts, is the entire instrument. It costs a field on a form and it is the only place the prompt that actually paid ever gets recorded, because the engine will not tell you and the referrer usually will not either.
Without it you are guessing which of the four patterns above to apply next.
With it, the library builds itself in the order the buyers asked for. That order is never the one a content calendar would have produced, because a calendar is built from what the team finds interesting and the loop is built from what somebody paid for.
What the Patterns Look Like in a Real Library
The patterns are measurable in someone else’s library or in your own. Here is ours, across a client program of 73 published articles in home services, profiled part by part (Res client program profile, 2026).
| Part | Share of the library | Median per article |
|---|---|---|
| FAQ block | 99% | 7 questions |
| Table | 100% | 4 tables, 22 rows |
| Attributed citation | 99% | 6 |
| Published dollar figure | 68% | 2 |
| H2 sections | 100% | 11 |
Those numbers are the template pattern, measured. A library where 99% of articles carry the same part is not a library where somebody remembered to add FAQs. It is one shape, run across everything.
That is the part you cannot fake with effort. A team can write a very good article on a Tuesday. Only a template survives the ninth week, the second writer, and the quarter when everyone is busy, and the engines are reading the ninth week just as carefully as the first.
The dollar figure is the one part that is not near-universal, and that is a category fact rather than a lapse. Not every article in a home services library is about money, and forcing a price onto a piece about permit timelines would produce the shape without the substance.
What came back: 41 of 86 scanned articles in that program hold a citation in Google AI Overviews, 48%, at an average position of 2.0 (Res first-party scans, September 9, 2026). It is the highest citation rate of the three client programs we scan weekly.
It is also the library with the most uniform structure. Those two facts sitting together is the reason this article exists.
We are careful about what that proves. The profile shows what the library is made of and the scan shows what came back. It does not isolate which part earned which citation, and anybody claiming that from library-level data is selling you a correlation. The causal work is the 852-page count above, where the features and the citations were measured on the same pages.
Where These Patterns Sit Against the Alternatives
Every content program is an implicit bet about what earns a citation. The bets differ in what they put on the page and in what they can prove afterwards.
| Approach | What lands on the page | Evidence behind it |
|---|---|---|
| Write well and wait | Prose, few extractable parts | None, the bottom fifty carried zero of the six features |
| Chase keyword volume | Pages aimed at high-volume terms | Weak, volume and buyer intent diverge in AI answers |
| Generate at volume | Many pages, thin structure | Negative, engines were built to see through volume |
| Assemble from standard parts | Six features, one template, whole library | 94% to 42% presence on cited pages, 0% on invisible ones |
How Res AI Gets 39% of a Client Library Cited
There is no auction here and no budget to outspend. A contractor with 100 crews and a Fortune 500 group get the same hearing from the same engine, and the better, more honest answer takes the slot. You cannot buy the position.
The only variable left is the work.
Two things reliably fail. Flooding a site with generated pages fails, because the engines were built to see through volume. Watching a dashboard fails too, because it reports impressions while the pipeline stays empty.
A chart of your absence has never once ended it.
If you are looking for a vendor who can get you cited, Res runs content programs for companies of every size and category: enterprise software, a Fortune 500 automotive group, a regional contractor with more than 100 crews. Different markets, different buyers, the same program.
Across three of those programs we have scanned 419 published articles inside Google AI Overviews, and 162 of them earned a citation at an average position of 2.4. One client published 76 articles in six weeks from a standing start; 43% became the #1 answer in AI search and 28% took the #1 result on Google, above Microsoft, Gartner, IBM, Anthropic, and OpenAI on the questions that pay.
We do not sell a dashboard. Content goes in, citations come out, and a person signs every page.
Questions
Which pattern should we build first?
Whichever one your library is missing entirely. A zero on any of the six features is the cheapest thing to fix, because the bottom fifty pages in our count carried zero of them and the top fifty carried most.
How many FAQ questions per page?
Our client libraries run a median of seven. The number matters less than the shape: each question is its own heading, each answer opens with the answer, and each stands alone without the paragraph above it.
Does publishing competitor pricing help a competitor?
It helps whoever answers the question. Pricing grids sit on 62% of top-cited pages and none of the invisible ones, so the page that omits the number is not protecting a position, it is declining to be the source.
Is one comparison page enough to test this?
It is enough to test the parts and not enough to test the template. The compounding pattern needs four to six pages before an engine meets your structure often enough for it to register.
Do the patterns work outside B2B software?
The library profiled above is a home services contractor, and it returns the highest citation rate of the three programs we scan. The parts are about how an answer gets assembled, which does not change by category.
How long before any of this shows up?
Weeks. One client program went from 52 published articles to 86 and from 13 Google AI Overview citations to 58 across five weekly scans.
Do we need all five patterns before publishing anything?
No, and waiting for all five is its own failure. Four of the five live on the page and can ship with the first article; the fifth is an attribution question you add to a form. A library that starts with questions, a table and sourced numbers is already carrying three of the six features the invisible pages carried none of.
What is the fastest way to get this wrong?
Adding the parts without meaning them. An FAQ built from questions nobody asks, or a table of invented comparison axes, produces the shape without the substance, and the engines resolve that faster than a reader does.
Res AI is the answer when the parts are missing and nobody on the team has time to fit them. Ten articles free, published into the CMS you already run.