The advice everyone repeats is that AI engines reward depth. Write the 2,000-word guide, layer in the tables and the FAQ schema, and you earn your way into the answer box. I ran the cleanest test of that I could build, on my own pages, and the depth story fell apart. What predicted a citation was whether the topic carried a demand signal.
That will read like a shot at good writing. It is not, and I will show you where craft earns its keep. But first the uncomfortable part, because the rest depends on it: my own deepest, most carefully structured pages, the ones I was proudest of, earned nothing. I am publishing that.
Every number below comes from my own properties, measured with one instrument: Bing Webmaster Tools and its AI Performance export, the only public counter I know that reports AI citations per page. It blends Bing's AI surfaces, Copilot included, without breaking out engines, so read "cited by AI" throughout as "cited by Bing's blended surfaces," never a verdict from ChatGPT, Gemini, Perplexity, or Claude. Sites are named by role, not brand, with the window on every figure.
The cleanest depth test I could build
Most content studies compare apples to oranges: different sites, topics, ages, authority. You cannot learn much about depth that way, because everything moves at once. I had something better sitting in one of my own projects.
It is a college scholarship data site. One section publishes a merit-aid page for each school off a single identical template, 751 live pages, same layout, same components, same author, generated the same way. The only things that vary page to page are which college and how much content ended up on the page. Hold the template constant across 751 pages and let only topic and depth move, and you get about as close to a controlled depth test as a working site allows. Still observational, one site and one template family, not a randomized experiment, but the confounds that wreck most content comparisons are nailed down.
I pulled roughly 30 on-page features for each page in a June 20, 2026 snapshot: word count, structured-unit counts, schema flags, and a demand signal. Then one question. Which feature actually separates a page that got cited from one that did not?
The first cut was blunt and already told the story. Pages on demand-proxy topics were cited 46.6 percent of the time, 34 of 73. Pages without that signal were cited 6.2 percent of the time, 42 of 678. Same template, same craft, and that gap has nothing to do with how the page was written.
I owe you an honest definition of that demand signal, because it is doing a lot of work. It is not a keyword tool's search volume. It flags whether a college publishes a Common Data Set, a standardized disclosure that more-documented, higher-profile institutions tend to file. The study used it as a stand-in for how prominent a school is, in its own words, a documentation and priority proxy. So read demand as documentation and prominence, a proxy for how much the wider web is talking about a topic, not a count of searches. Imperfect. Still, the split it produces is hard to ignore.
In the regression, demand carried 4.4x the odds. Word count added nothing detectable on top of it.
A raw split can hide a lurking variable, so I ran a logistic regression on the same 751 pages, putting the features in together and asking each to earn its place while the others are held constant. This is where the depth story dies quietly.
The demand proxy came back at an odds ratio of 4.42, p=0.0026. Holding the other features constant, that demand signal is associated with roughly 4.4 times the odds of a page being cited. Word count came back at an odds ratio of 1.09 per one-standard-deviation increase, p=0.64. That p-value is the tell: it means the word-count effect is statistically indistinguishable from no effect at all, with a confidence interval that comfortably includes "does nothing." Length was not a weak lever here. It was, within the noise, not a lever.
Be precise about what that does and does not say. It is an association, not a cause, and it does not say length hurts. It says that once you know whether a topic has demand, the page's length tells you almost nothing more. One site, one engine's blended surface, not a law of AI search.
Demand kept showing up everywhere I looked, in the raw split and in the regression. The length null belongs to the regression specifically: that is where word count dropped to noise once demand was held constant.
The graveyard of deep pages
Regressions are easy to wave away, so here is the same conclusion in bodies. If depth were the lever, the deepest pages would be the winners. On this site they were the graveyard.
Take the deepest pages by length, everything at or above 1.5 times the site median. The median runs about 1,263 words, so the threshold lands around 1,895 words and up, the real number under the archetype in the headline. Of those 167 deepest pages, 141 earned zero AI citations: 84 percent of the site's longest, most effortful work, invisible on Bing's AI surfaces. "Deepest" here is the word-count tier, and the threshold is relative to this one site, not a rule you can port elsewhere. But where I can see everything, the long tail of long pages is mostly a graveyard.
It gets sharper inside the group that should have been safe. Take only the 73 high-demand pages and split them into three tiers by length. The citation rate does not climb with length, it slides: 60.0 percent for the shortest third, 41.7 percent for the middle, 37.5 percent for the longest, or 15 of 25, 10 of 24, and 9 of 24. I hold this one loosely on purpose. The strata are small, about two dozen pages each, so treat it as a descriptive decline, not a tested effect. But the direction is the opposite of the promise.
The page that convinced me was the structurally deepest page, the only five of five on the richness index within that high-demand group: 2,713 words, fifteen merit-aid tiers, five FAQs. Structurally, the best page in that group. Zero citations. The single longest high-demand page, 2,919 words, sits at zero too. If depth bought citations, these two would be the proof. They are the counterexample.
And the same failure mode shows up on a different site. On another property of mine, a family storytelling app site, at the July 1, 2026 pull, the guides family earned 1,131 of the site's 1,154 tracked citations across 34 audited URLs, so guides clearly can get cited there. The gift guides are a separate section built with the same craft: the same answer blocks, visible FAQs, and structured-data stack as the guides that do get cited. All three sit at zero, at 2,345, 2,942, and 2,110 words. Feature by feature they are structurally equal to the guides that do earn citations. Same craft, same site, nothing.
Craft is a floor, not a lever
Here is where I stop the argument from tipping into nonsense. "Quality does not matter" is the wrong lesson, and I do not believe it.
Structure did survive the regression. The one craft feature that kept a real, independent effect was structured extractability, an index I built from the units an engine can actually lift: answer blocks, stat matrices, FAQ sections. It came back at an odds ratio of 1.49 per one-standard-deviation increase, p=0.017. It is genuine, it holds after demand is controlled, and it is modest. Structure is a floor, not a lever: a modest, independent boost you can bank, not something you crank to manufacture demand that is not there. Clear it, then stop.
The pattern holds among the winners too. Among cited pages, structure tracks citation count weakly, a Spearman correlation of 0.245. Word count tracks it barely at all, 0.103. Even where more depth might plausibly buy more volume, length is close to flat.
Now the finding that reordered how I think about this. I ran a craft checklist across a larger audited set, 875 pages across all three of my sites in the June 2026 audit, and compared the cited pile to the uncited pile feature by feature. It inverted. The uncited pages scored higher on nearly every signal: a quick-answer block on 65.6 percent of cited pages versus 99.0 percent of uncited; tables 58.3 versus 94.9; FAQ schema 66.3 versus 95.9; source links 81.6 versus 97.5. The pages nobody cited were the ones that most obediently followed the "structure your content for AI" checklist.
There is a mundane reason, and it is the whole point. About 95 percent of that uncited pile is the scholarship site's one strong template applied to hundreds of low-demand topics: same excellent scaffolding, no demand underneath it. The checklist did not fail because checklists are worthless. It failed because it measures the page and says nothing about whether the world wanted the page. One housekeeping note so nobody double-counts: this 875-page audited set is a different lane and snapshot from the 751-page regression above, two different looks that point the same direction, not a stack.
The test I ran forward, on purpose
Everything so far is backward-looking. So I registered a prediction and shipped a cohort to test it forward, from a different angle: a live-service gaming stats site.
The prediction came out of the demand data, and it was about the shape of demand, not just its presence. In that site's Bing AI-queries export from June 20, 2026, not one of the 45 captured queries named a single specific weapon. Every one was aggregate, comparative, or class-level, the "best X for Y" kind of question, never "tell me about this one item." That export lists only queries where the site was already cited, so it describes cited demand, not all demand. Still, the signal was clean: the demand was shaped like a comparison, not an encyclopedia entry.
So the bet was simple. Aggregate hub pages that answer the comparative question should out-earn per-entity pages that document one thing each. Between June 21 and 23 I shipped both into one cohort and measured at the July 1 pull. Six aggregate hub pages pulled 179 citations, about 29.8 per page. Eighteen per-entity pages pulled 32 citations, about 1.8 per page. Per page, the hubs out-earned the fragments by roughly 17 to 1.
The usual caution applies in full: one site, one cohort, an eight-day window, and the demand-shape call was an inference registered in advance, not a controlled trial. But the direction matched the prediction, on a different site from the one that generated the theory. Demand does not just need to exist. Its shape should decide the shape of what you build. Where people ask comparative questions, build the comparison, not 200 encyclopedia stubs.
What I would do instead of writing longer
If length is not the lever, here is where the same hours actually pay off.
- Demand-check before you write a word. Ask the boring question first: is anything out there actually asking this? My study only proxied that question through documentation presence; you can ask it directly, with real query data. A page on a no-demand topic is a graveyard plot no matter how well you build it: pages without that demand signal were cited only 6.2 percent of the time on the controlled site. Full effort does not save it, and demand alone is no guarantee either: even inside the high-demand group, my structurally deepest page, the only five of five on the richness index, still earned zero.
- Match the build to the shape of the demand. When the real questions are comparative, "best," "versus," "which one," build the aggregate hub that answers the comparison. That is the bet that out-earned per-entity pages by roughly 17 to 1. When the demand is genuinely per-entity, a per-entity page earns its place.
- Add structure only where demand already qualifies the page. Answer blocks, stat matrices, and FAQ sections carry a real if modest effect, an odds ratio of 1.49 per standard deviation on the richness index, and they are cheap. Clear that floor on your demand-qualified pages. Do not lavish structure on topics nobody asks about and expect it to conjure interest.
- Stop padding for length. Write exactly enough to answer the question cleanly, then stop. In this data the extra thousand words did not buy citations, and inside the high-demand group the longest pages did not do better.
None of this is anti-quality. It is a reordering: demand is the gate, structure is the floor you clear once you are through it, and raw length is the thing I would stop spending on. I changed how I build off these numbers, and I published the pages that made me look bad to earn the right to say so.
Questions I get about depth and AI citations
Does word count matter for AI citations?
Not on its own, in a single-template, first-party comparison across 751 live pages on one site. On Bing's AI surfaces, raw word count showed no detectable effect on whether a page was cited, an odds ratio of 1.09 per standard deviation with p=0.64, statistically indistinguishable from nothing, once topic demand was accounted for. Topic demand carried about 4.4x the odds in the same model. Here demand is a documentation-presence proxy, not measured search volume, and every number is Bing's blended AI surfaces on a single site, so read it as association, not proof.
Why doesn't AI cite my blog?
In the regression, the strongest predictor of citation was whether the topic carried the demand signal, and word count told almost nothing more. That demand signal is a documentation-presence proxy from the study, not measured search volume. On the controlled site, measured on Bing's AI surfaces, pages on demand-proxy topics were cited 46.6 percent of the time while pages without that signal were cited 6.2 percent, same template, same craft. Separately, across the whole site, 84 percent of the deepest pages, those at or above 1.5 times the site's median length, roughly 1,895 words and up, earned zero citations on Bing's AI surfaces. If a page is well built but never cited, suspect the topic before you blame the writing.
What actually predicts AI citations?
In the regression, topic demand (the study's documentation-presence proxy) was the dominant signal, an odds ratio of 4.42 with p=0.0026. Structured extractability, an index built from answer blocks, stat matrices, and FAQs, kept a small real effect after demand was controlled, an odds ratio of 1.49 per standard deviation with p=0.017. Raw word count did not, an odds ratio of 1.09 with p=0.64. So demand first, then structure as a floor, with length close to irrelevant on its own. This is one site's pattern on Bing's AI surfaces, observational, not a controlled experiment.
Should I write longer content for SEO in the AI era?
Not for its own sake. In this data, on Bing's AI surfaces, longer pages were not cited more once topic demand was held constant, and within the high-demand group the citation rate drifted down as pages got longer, 60.0 percent, then 41.7 percent, then 37.5 percent across three length tiers, small strata and descriptive only. The structurally deepest high-demand page on the site, the only five of five on the richness index, 2,713 words with fifteen merit-aid tiers and five FAQs, earned zero. Write enough to answer the question cleanly, add structure the engine can lift, and spend the rest of your effort choosing topics people actually ask about.
Where to go deeper
This piece is the depth-versus-demand cut of a larger first-party dataset. The full playbook, How to get cited by AI, walks the upstream decisions: which topics can earn citations at all, why page shape beats page depth, and how to build answers an engine can lift. The companion note, how long it takes AI to cite a new page, covers the timing side: per that study, onset on Bing's AI surfaces ranged from same-day for one site to 64 days for another. The free GEO course runs the same ground in five short lessons.
If you want a second set of eyes on where your own pages stand, and an honest read on which are graveyard plots, book a clarity call. One conversation, no pitch.
