Formatting Isn't GEO: What Actually Gets Your Content Cited

Reading GEO (Generative Engine Optimization) as "tidy up the formatting a bit" is the most widespread misunderstanding in this field. A 252,000-run controlled experiment found that changing format alone has an unstable effect on whether AI cites you. What actually works is the content itself — whether the topic is relevant, whether it sits in the right place, and whether there is citable evidence.

This isn't to say format is useless. More precisely: format is the entry ticket, not the competitive advantage. Messy formatting does stop AI understanding you; but tidy formatting doesn't mean AI will cite you either. Many people merge those two things into one.

What a 252,000-run experiment found

A paper accepted at SIGIR 2026, "What Gets Cited", did something thorough: 252,000 controlled experiments — 6 models × 18 categories of factor, each run presenting two candidate pages that differed by a single variable. All brand names were anonymised so brand recognition couldn't affect the result.

The conclusion runs against intuition:

• What matters most is topic relevance and context position — how relevant your content is to the question, and where on the page it appears;
• "Explicit presentational cues" have an unstable effect — formatting adjustments, layout changes and the like;
• Separately, explicit price information and a recent timestamp have a consistently positive effect.

Scope note: these are relative preferences under controlled experimental conditions, not a guarantee of real-market results; what was measured is "how AI chooses among candidates", which cannot be turned into "do this and you will be cited".

But the direction is clear: AI picks content, not appearance.

So what does work consistently?

Putting several studies side by side, the factors that work consistently fall into a few groups:

Topic relevance

It sounds obvious, but this is exactly where a lot of content fails — pages of industry buzzwords with no direct answer to the specific question. What AI wants is content that answers this query.

Citable evidence

The academic origin of the term GEO — the KDD 2024 paper (Aggarwal et al.) — tested this: adding source citations, quotations and statistics has different effects depending on position. For a page originally ranked 5th, visibility gains reached 115.1%; for a page originally ranked 1st, the effect was negative.

The same paper also found that keyword stuffing is ineffective or even harmful in AI answers.

Scope note: these percentages are relative changes under laboratory conditions and a fixed query set, not a promise of real-market results. But they make one thing clear — information that can serve as evidence matters far more than handsome formatting.

How information is organised — not how it looks

A distinction is needed here, or it's easy to go off track. Two studies have measured the effect of "structure":

• UC Berkeley (arXiv:2509.10762, not peer-reviewed): the correlation between structural factors in content and citation is between 0.63 and 0.68;
• Yonsei University (arXiv:2602.12187, not peer-reviewed): after structural optimisation, retrieval hit rates rose by about 22%, while optimising the body text without structuring it had a negative effect.

Note that this is structure — how information is organised (what comes first, whether it's layered, whether concepts are defined) — not format — font size, colour, borders.

A lot of GEO teaching conflates the two. Changing font size, adding colour blocks, applying a template — those are format; what actually affects citation is the logic by which information is organised.

Why the "reformat it" story is so popular

Because it is the easiest action to deliver.

Retrofitting layout can be done to a whole site in a day, and it's visible — you can screenshot a before and after. "Raise topical relevance, add evidence, reorganise information", on the other hand, requires genuinely understanding the industry and knowing what customers ask — much harder and much slower.

Hence a common shape of service: tidy up the website's heading hierarchy, add a few structured data tags, adjust the font sizes, then tell the client "GEO is done".

These actions aren't useless — they make your pages easier for AI to read. But they are the pass mark, not the competitive advantage.

So should you bother with format at all?

Yes, but the order matters.

1. First, make sure it can be read: clear heading hierarchy, paragraphs in the body, lists where a list fits, tables where a table fits. This is the entry ticket; without it nothing downstream counts.
2. Second, make sure it can answer: write out, one by one, the questions customers in your industry ask. This step is content, not layout.
3. Third, make sure it can be trusted: add data, sources, cases and timestamps — give AI a reason to cite you.

Reverse the order and you get "beautiful formatting, and AI still doesn't cite you".

Three questions about GEO and formatting

So should we do structured data (Schema)?

Yes, but don't expect it to "raise your ranking". Google's official documentation, "AI features and your website", is clear: AI features carry no additional technical requirements, and ordinary structured data is enough. And one point is a hard requirement — structured data must match the visible text on the page; you can't write one thing in the code and another on the page.

I've seen people change their formatting and get traffic. How do you explain that?

They probably changed something else at the same time. Reformatting often comes bundled with "and while we were at it, we rewrote the content" — and it's the second part that did the work. That's exactly why experiments change one variable at a time: bundle two changes together and you can't tell which one earned the result.

So should we still optimise the layout?

Yes, but treat it as clearing obstacles, not as your competitive edge. Leave the obstacles and even good content won't get through; only with them cleared can the content do its job.

In closing

To sum up: GEO isn't "make the website look better", it's "make the content more worth citing".

Format answers "can AI understand it"; content answers "will AI choose to cite it". The ceiling on the first is low; the real gap is in the second.

You can judge for yourself: on your website, how much content is a head-on answer to a question customers actually ask? If the proportion is low, the problem isn't your layout.

When you started GEO, were you advised to change the format first, or the content first? Worth checking which one your current provider led with.

Yihang Network, based in Shenzhen, previously worked in cross-border B2B and Alibaba.com-related services, and expanded into GEO from 2024. Our position: format has to pass, but the emphasis is always on content.

Data and scope note: The SIGIR 2026 paper cited here, "What Gets Cited", is an accepted conference paper; its conclusions come from 252,000 controlled experiments and reflect AI's relative preference among candidate content under controlled conditions. The KDD 2024 paper (DOI: 10.1145/3637528.3671900) is a peer-reviewed conference paper and its percentages are relative changes under laboratory conditions and a fixed query set. The UC Berkeley (arXiv:2509.10762) and Yonsei University (arXiv:2602.12187) studies are both preprints, not peer-reviewed. None of this represents real-market outcomes or any single company's results. The material on Google comes from its official documentation, "AI features and your website". The methods described raise the probability of being retrieved and cited; they do not guarantee a company will appear in AI answers, and no timeline or business result is promised.

References

1. SIGIR 2026 (accepted), "What Gets Cited", controlled-experiment study
2. Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024 (DOI: 10.1145/3637528.3671900)
3. UC Berkeley, arXiv:2509.10762 (preprint, not peer-reviewed)
4. Yonsei University, arXiv:2602.12187 (preprint, not peer-reviewed)
5. Google Search Central, "AI features and your website"