Most GEO advice about content structure is directional: use headings, keep paragraphs short, front-load the answer. A new academic framework puts an actual number on how much structure alone is worth, tested at the level of individual structural decisions rather than vague, hand-wavy best practices.
The framework
GEO-SFE, published in March 2026 by a team of four researchers, proposes that content structure isn't one variable, it's three independent ones stacked on top of each other, each operating at a different scale of the document. The framework's core move is separating structure from semantics entirely: it optimizes how content is organized and formatted while deliberately preserving what the content actually says, then measures whether citation behavior changes anyway.
Tested across six generative engines, the structural interventions alone produced a 17.3% relative improvement in citation rate and an 18.5% improvement in subjective quality ratings, both measured against the same underlying content before and after restructuring. That's the headline that should reframe how a lot of GEO content briefs get written: a meaningful share of citation performance is available from format decisions alone, independent of whether you've upgraded the actual argument, sourcing, or data in the piece.
Testing across six different engines rather than one is what makes the result worth taking seriously instead of filing under "works for one model's quirks." Generative engines don't share a single retrieval and summarization pipeline; a finding that only held up on one system would be a fact about that system's specific implementation, not a fact about structure as a general lever. A 17.3% lift that holds across six independently-built engines is a much stronger claim, and it's the reason this framework reads as more durable than the usual single-platform GEO tip.
The full paper, published on arXiv under the title "Structural Feature Engineering for Generative Engine Optimization," is worth reading directly for teams that want the methodology behind the headline number rather than a summary of it: the GEO-SFE paper lays out exactly how each of the three levels was isolated and measured.
The three levels, explained
The reason this three-level split matters more than a single "structure" checklist is that the levels behave somewhat independently. A page can have excellent macro-structure, a clean outline with logical section order, and still perform poorly on meso-structure if its actual claims are buried three sentences deep inside long paragraphs. Optimizing one level without the others leaves real citation-rate improvement on the table.
It also explains a pattern a lot of content teams have noticed anecdotally without having a name for it: two pages that look equally well-organized at a glance, both with clean headings and a logical outline, can perform completely differently in AI citation tracking. The visible, skimmable structure, the macro level, was identical. The invisible structure underneath, how cleanly individual claims chunk apart from their neighbors, was not, and that's usually where the actual performance gap was hiding the whole time, invisible to anyone doing a quick visual scan of the two pages side by side.
What moved and what didn't
| STRUCTURAL LEVEL | TYPICAL INTERVENTION | WHAT IT TARGETS |
|---|---|---|
| Macro | Reorder sections, move the direct answer up | Whether the engine finds the right passage at all |
| Meso | Split compound paragraphs, pair claim with support | Whether a passage extracts cleanly without context loss |
| Micro | Bold key numbers, convert prose lists to bullets | Whether a passage reads as quotable at a glance |
The paper's authors frame this as distinct from token-level editing approaches, the kind of optimization that tweaks individual word choices or sentence phrasing to nudge a model's scoring. Structural intervention, they found, substantially outperformed those token-level baselines, and citation behavior tracked more strongly with document-level content properties than with isolated lexical edits. In plain terms: rewording a sentence to sound more citable moves the needle less than restructuring where and how that sentence sits relative to everything else around it on the page.
How to apply it without a rewrite
Start with an audit at all three levels on your highest-value existing pages, the ones already ranking or already getting occasional citations, rather than new content. Macro: does the direct answer to the page's core question sit in the first two hundred words, or is it buried under introduction and context? Meso: pick three random paragraphs and check whether each one contains a single self-contained claim with its support attached, or whether the claim and the evidence for it are split across separate paragraphs a model would have to stitch together. Micro: scan for numbers and key facts sitting unformatted inside dense prose that could be pulled into a bolded phrase, a short list, or a table cell instead.
Run that audit on ten pages and a pattern usually shows up fast: most pages fail at exactly one level, not all three evenly. A well-outlined blog archive tends to fail at meso, long paragraphs that bury the actual claim inside throat-clearing sentences. A page written by someone used to dense, academic-style prose tends to fail at micro, real data sitting unformatted in the middle of a sentence instead of set apart where it can be lifted cleanly. Knowing which level is failing on which content type turns this from a vague "make it more scannable" instruction into a specific, assignable editing task.
That audit is worth running as a standing checklist rather than a one-off exercise, since new content ships every week and each new page carries its own fresh chance to get one of the three levels wrong. This is deliberately restructuring work, not rewriting work, which is why it's a cheaper lift than most GEO content initiatives. A comparison page with strong underlying research can often pick up a meaningful share of this 17.3% just by moving its verdict paragraph earlier, splitting its longest paragraphs at the claim boundary, and converting two or three buried statistics into a table. None of that touches the actual argument or sourcing.
The order matters too, and it maps to the order these levels are listed in. Fix macro-structure first, since a well-chunked paragraph inside a badly-ordered document still won't get found by a model scanning for the answer to the query. Fix meso-structure second, since that's where the actual extraction quality lives. Treat micro-structure as the final pass, the cheapest and fastest of the three, applied once the underlying organization and chunking are already sound. Doing it in reverse order, polishing bold text and bullets on a document whose core answer is still buried on paragraph six, wastes the easiest wins on content that the other two levels haven't earned yet.
There's a temptation, especially on a large content library, to treat this as a job for automation: run every page through a tool that reflows paragraphs and adds bold tags at scale. Resist that instinct, at least for the meso level. Splitting a paragraph at the wrong point, one that separates a claim from a piece of context it actually needs to make sense, can make a passage read as more extractable to a naive structural check while making it less accurate or more misleading once it's actually lifted out of context. The macro and micro levels tolerate more automation reasonably well: reordering sections and bolding numbers are close to mechanical tasks. The meso level, splitting information at its true logical boundaries rather than an arbitrary sentence count, still needs a human editor who understands what the claim actually depends on.
Budget the work accordingly. A macro-structure pass on an existing page, moving a buried answer up, tightening a heading hierarchy, is often a fifteen-minute edit. A micro-structure pass, scanning for unformatted numbers and converting a couple of dense sentences into a table or bullet list, is similarly fast once you know what you're looking for. The meso-structure pass is the one that takes real editorial judgment and real time, because it means rereading every paragraph and asking whether the claim and its support are sitting together or scattered, and rewriting the ones that aren't. If a content team only has budget for one of the three passes this quarter, meso is the one worth prioritizing, both because it's the hardest to retrofit later and because it's the level most existing content briefs never mention at all.
None of this replaces the deeper work of building genuinely authoritative, well-sourced content that deserves to be cited in the first place, the kind our own guide to earning AI citations walks through end to end. What structural optimization does is make sure that content isn't losing citations it already earned on substance, purely because of how it's organized on the page. A broader GEO program should treat this three-level audit as a recurring pass, not a one-time launch task, since new pages ship with new structural mistakes on a rolling basis.
The reference example of house voice on this exact question, the piece where we first published our own citation-signal research, is worth a re-read with the macro/meso/micro split in mind: notice how the direct-answer callout sits near the top, how each data point gets its own short block instead of living inside a long paragraph, and how the numbers that matter most are pulled into a statband instead of buried in prose. That's the three-level framework applied without a single sentence of it being named on the page.
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.