Two credible studies on content length and AI citations landed a few months apart in 2026, and they point in opposite directions. Neither one is wrong. They are measuring different things, and the gap between them is the actual playbook for content length AI citations rewards.
The apparent contradiction
| SOURCE | FORMAT THAT WINS | THE NUMBER |
|---|---|---|
| Growth Memo (815K query-page pairs, ChatGPT-specific) | Short, focused pages | Outperform comprehensive guides on ChatGPT citations |
| ConvertMate GEO Benchmark 2026 | Long pages, 20,000+ characters | 4.3x more citations than pages under 500 characters |
| SparkToro (January 2026) | The first 30% of any page | Pulls 44.2% of all LLM citations, regardless of total length |
Growth Memo's April study ran 815,000 query-page pairs and found shorter, tightly focused pages winning ChatGPT citations over comprehensive guides. ConvertMate's benchmark, covering a broader set of engines, found the opposite pattern at the top end: pages over 20,000 characters pulled 4.3 times more citations than pages under 500 characters.
Put those two studies side by side on a content calendar and you get two teams arguing past each other: one says cut the guide down, the other says the guide isn't long enough. Both are looking at real data. Neither is looking at the whole picture.
What the content length AI citations numbers actually say
Read them as two different questions and the conflict disappears. Growth Memo is asking which page ChatGPT specifically chooses to cite when several pages could answer the query. ConvertMate is asking, across engines, whether a page has enough substance to get retrieved and cited at all. A five-hundred-character page rarely has enough material to be the best source for anything. A twenty-thousand-character page has more surface area to be relevant to more sub-queries, the same sub-query behavior we cover in how AI agents actually search your site.
It's worth being honest about how these two studies were built before treating either as gospel. Growth Memo's ChatGPT-specific number comes from observing real query-page pairs at massive scale, 815,000 of them, which makes it a strong read on ChatGPT's actual behavior but says nothing about Perplexity, Claude, or AI Mode. ConvertMate's benchmark pools across engines, which makes it a better read on the aggregate market but blurs any single engine's specific preference. Neither methodology is wrong. They're just answering different-sized questions, and conflating them is where most of the 'short vs. long' arguments online go sideways.
Why SparkToro's number resolves the fight
SparkToro's number is the piece that resolves it. Across the citations SparkToro tracked in January, 44.2% pulled from the first 30% of the page, no matter how long the page ran. Length gets you retrieved. The opening third gets you cited. A long page with a padded, generic intro loses to ChatGPT's preference for focus, because the part doing the work, the first 30%, is not actually focused.
This is the same failure mode we see across comparison pages that win citations: the pages that underperform almost never fail because they're too short or too long in the aggregate. They fail because the opening section reads like a warm-up instead of an answer.
Growth Memo's broader body of work backs this up from a different angle. Its March study on how AI picks its sources, covering 21,000-plus citations, found content depth and focus interacting in ways that reward pages doing one job well over pages trying to cover everything. And its May research on 'reasoning lift' found that high-reasoning prompts, the multi-step, comparison-heavy questions B2B buyers actually ask, generate 4.6 times more searches per query and cite noticeably different content than simple factual prompts do. A shallow, short page can win a simple factual citation. It has no shot at surviving the kind of multi-hop research pass a real buying decision triggers.
There's a coverage angle worth folding in here too. Ahrefs' 75,000-brand study found YouTube mentions carrying the strongest off-site correlation with AI visibility of any signal measured, at 0.737, ahead of every other external factor including backlinks. A page that anchors a written asset with a genuinely useful video walkthrough is not padding length for its own sake. It's adding a second, independently-weighted signal that reinforces the same claim the written page is making, which is a different kind of depth than just adding more paragraphs.
How to build one page for both
Write long enough to cover the sub-queries a buyer's research pass will actually generate: comparisons, edge cases, objections. Then front-load the first 30% of that page with the single most extractable, direct answer to the primary question, not a scene-setting intro. That structure satisfies ConvertMate's substance threshold and Growth Memo's focus preference in the same document, because the 'focused page' ChatGPT rewards is really just describing a well-written first section, not a short total page. It's the same architecture we build into hub-and-spoke content for long B2B sales cycles: a focused entry point, backed by enough depth to survive the follow-up questions.
This matters more once you factor in how buyers actually behave once an AI Mode answer gets them partway there. Growth Memo's April usability research, a study of 185 documented high-stakes purchase tasks, found AI Mode users far more willing to accept an LLM-generated shortlist than classic Google users, who tend to build their own comparison set by hand. A shortlist-accepting buyer is not going to click through five separate short posts to build a picture. They're going to lean on whichever single page gave them a complete, confidently structured answer on the first pass, which is exactly the page shaped by the first 30 percent rule above.
A content length AI citations checklist
One more wrinkle worth planning for: the 'first 30%' rule isn't a fixed word count, it's a proportion, which means it scales with how long you decide the full page needs to be. A 1,500-word page's opening 30% is roughly 450 words, room for a real, sourced, multi-sentence answer. A 20,000-character page's opening 30% is closer to 900 to 1,000 words, enough room to include a comparison table or a short data block before the deeper sections begin. Don't treat the opening as a fixed-length snippet bolted onto whatever comes after. Size it deliberately as a fraction of the whole.
There's a testing implication too, and it's the one most content teams skip. Run the same page through a per-engine citation check, not just a blended one, the same discipline we argue for in why your mention-rate dashboard is lying to you. A page that's winning ChatGPT citations because of a sharp opening section might be losing ConvertMate-style aggregate citations because the rest of the page never got built out. You won't see that gap in a single combined score. You'll only see it by checking the two studies' predictions against your own page, separately, on a schedule.
Before you publish, check the page against both studies at once: does the first 30% stand alone as a complete, extractable answer, and does the full page run long enough to earn retrieval in the first place. If either answer is no, you're optimizing for one engine's preference at the expense of the other's threshold. This is the exact review we run inside our content marketing engagements, and it's the structural lens behind the retrospective in our content-restructuring case study, where consolidating thin pages into fewer, deeper ones lifted both citations and rankings at once.
In practice, this changes how an editorial calendar gets built, not just how a single page gets outlined. Instead of assigning 'short posts' and 'long guides' as separate content types, assign a single owner per topic cluster who is responsible for both the extractable opening and the depth underneath it. Splitting those two jobs across different writers is exactly how you end up with a long page that never gets a properly focused lead, or a short page that never earns enough retrieval to be in the running for a citation at all.
Whichever team owns this, make the review recurring, not a one-time launch checklist. Content that passed both tests at publish can drift out of compliance as an engine's preferences shift, the same way Google and ChatGPT's citation-placement tests are shifting the ground under GEO right now, so a page that worked in Q1 is worth a second look in Q3.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.