Benchmarking Is the Only Content Format That Still Gets Cited by AI

Ahrefs tracked 1885 pages that added JSON-LD schema markup between August 2025 and March 2026, then compared them against 4000 control pages that didn't touch a thing. The result fits in 1 line: nothing happened. No meaningful lift in AI Mode citations, none in ChatGPT. Google AI Overviews showed a drop, and even that one can't be pinned on the schema with any confidence.

This flips 2 years of SEO gospel on its head. The confusion came from a correlation that looked airtight on paper. Across 6 million URLs, pages cited by AI were 3 times more likely to carry JSON-LD than pages that weren't cited. Correlation, not causation, and Ahrefs built the study specifically to settle which one it was. Verdict: schema markup doesn't cause citations. It just tags along with pages that are already doing something else right.

So what does move the needle? Not a technical tag buried in the head of your HTML. The format of the content itself. I tested that theory against a comparison article I published 2 days ago, and I want to be upfront about the limits here. That article is too fresh to have measurable citation data yet. What I can show you is the method, not the result. Not yet.

The Schema Markup Myth, Officially Dead

Here's the part that stings if you've spent the last 18 months adding structured data to every page on your site: it wasn't wasted, exactly, but it also wasn't the lever you thought it was. Ahrefs' difference-in-differences design is the kind of study that actually isolates cause from noise, and the noise is where schema markup lives now.

Add the JSON-LD, follow every step in the guide, ship the update. And the citation graph doesn't move an inch, no explanation offered, straight to the Dark Souls death screen: You Died.

A related finding softens the blow a little. A 2026 report from The Digital Bloom found that 28.3% of the pages most cited by ChatGPT have zero organic visibility on Google. Read that twice. Nearly a third of what gets cited never ranked in the first place. If your comparison article isn't sitting on page 1 for its target keyword, that's not automatically a death sentence for AI visibility. It might even be irrelevant.

There's a companion myth worth flagging while we're here: the title tag problem killing AI search visibility gets treated the same way schema does, a checkbox that feels productive without moving the number that matters. Same pattern, different tag.

The 2.7% That Captures 8.4% of Citations

A Growth Memo analysis of Ahrefs data puts a sharper point on this. Pages representing genuine primary research make up 2.7% of the corpus analyzed, and they capture 8.4% of all AI citations. Out of 90 citations tracked back to primary content, 75 trace to 1 single format: the data benchmark.

I read that number 5 minutes before a client call I hadn't prepped for, and it landed weird. Not because it was surprising in the abstract, more because I'd been treating my comparison articles as a conversion tool first and a citation asset second. The data says flip that. The listicle format ("best X for Y") also dominates Ahrefs' analysis of 750 ChatGPT prompts, which tells you the AI engines aren't randomly sampling the web. They're pattern-matching toward structured comparison, and a benchmark is the purest version of that pattern.

None of this fixes the actual bug I had open in another tab while writing this, a Convex query that decided pagination was optional again (the dashboard still won't tell you why). Different problem, same afternoon, zero connection to benchmarks. Just how the day goes sometimes.

There's a grim footnote buried in the same dataset: 17.5% of cited URLs, 64 out of 365 tracked, were dead or redirected by the time anyone checked. That's 203 citations evaporating because nobody kept the page alive. I think that number bothers me more than the schema result did, honestly, because it's a problem with an obvious fix that almost nobody bothers to apply. 😑

The 5 Levers That Move Citations

Princeton and Georgia Tech's GEO research (Generative Engine Optimization, the emerging discipline of getting cited by AI answer engines rather than ranked by traditional search) breaks down what actually correlates with citation rate. 3 levers stand out, and each one reads like an RPG stat buff, +40% here, +22% there, except no character sheet tells you which build actually ships:

• Inline citations pointing to primary sources: +40% citation rate
• Specific statistics added to the page: +37%
• Quotes attributed to named experts: +22%

The other 2 levers aren't measured with a clean percentage but show up constantly in what actually gets cited: a written methodology section, and a comparison table rendered as text rather than baked into an image (AI crawlers can't parse pixels). Add a URL that doesn't move once published, and periodic updates that refresh the content without changing that URL. Fivetran's benchmark from 2022 is still getting cited 4 years later for exactly this reason. Nobody moved the goalposts, they just kept the numbers current.

1 nuance before this turns into a checklist you follow blindly: domain authority still matters on its own axis. SE Ranking found that sites with 32,000+ referring domains are 3.5 times more likely to get cited by ChatGPT than sites sitting under 200. Format doesn't erase authority. It's a second lever on the same machine, not a replacement for the first one.

I Ran My Own Comparison Through This

I published a comparison piece on 2 social media posting APIs 2 days ago, Upload-Post versus a newer challenger called Zernio, built from 8 months of actually running both in production. It's not a weak article. It has hard numbers, a decision grid, an FAQ section, cited sources. My first pass through the checklist above, I expected to find it mostly clean.

2 gaps instead. The comparison table with the actual pricing and performance numbers exists, but only as an infographic image. Looks great, reads well, completely invisible to anything trying to extract structured data from the page, cloaked like it's wearing a Predator suit in front of a thermal scope. Second gap: there's no methodology section anywhere. I never wrote down why I picked the 8 criteria I compared on, or how the 8-month test period was structured. A human reader trusts the number because the writing sounds confident. An AI model has nothing to point at.

A benchmark stuck in a PNG is a benchmark no AI can read.

Both fixes are mechanical rather than creative, which is its own small relief. Convert the infographic's numbers into an actual markdown table sitting in the text. Add 3 or 4 sentences up top explaining the test window and the selection criteria for what got compared. Neither change touches the voice of the piece or the conclusions in it. I'm just giving the AI something to grab onto that isn't locked inside an image file.

1 lever from the framework doesn't apply here and I'd rather say so than pretend otherwise. There are no named expert quotes in that comparison, because it's a first-person account of running 2 tools myself for 8 months. That's fine. Not every article needs every lever, and forcing an expert quote into a personal test log would read as exactly the kind of padding this whole piece argues against. If you want to see what happens when you test the citation question directly instead of inferring it from format, I already ran asking 3 AI models to recommend my SaaS, which is the blunt version of this same question.

The Prompt That Rebuilds Any Comparison

Here's the prompt I'm using to run any existing comparison article through the same 2 fixes, methodology and table format, without touching the parts that already work.

You are restructuring an existing comparison article to improve
its citability by AI answer engines (ChatGPT, Google AI Overviews,
Perplexity), without changing its voice, conclusions, or examples.

INPUT: [paste your existing comparison article here]

Apply these changes only:

1. Extract every numeric comparison currently shown as prose or
   as an image into a plain markdown table. Each row is 1
   comparison criterion, columns are the items compared.

2. Write a short "Methodology" section (3-5 sentences) near the
   top explaining: what was tested, over what time period, and
   why these specific criteria were chosen over others.

3. If any statistic in the article is approximate or unsourced,
   flag it in brackets [VERIFY] rather than inventing a source.

4. Do not add expert quotes, invented statistics, or new claims.
   Only restructure what already exists in the input.

5. Keep the original URL slug unchanged if this is an update to
   a live page.

Return the restructured article in full.

Swap the bracketed input for your own article and run it. The output won't be dramatically different from your original, and that's the point. It's the same argument, wearing a shape that a language model can actually parse. 1 caveat that the prompt can't fix for you: none of this is a one-time job. A benchmark cited 4 years later got there by staying current, not by getting lucky once.

Schema markup was never the lever. Go find your comparison articles, turn the images into tables, and write down why you tested what you tested. 📊

Sources

• Ahrefs, "We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved." (May 2026).
• GEO research on citation levers via Princeton and Georgia Tech, covered by AuthorityTech.
• The Digital Bloom, "2026 AI Citation Position and Revenue Report."
• Omnibound, "AI SEO Statistics (2026)," citing SE Ranking domain authority data.
• Growth Memo analysis of Ahrefs primary-research citation data.

This post may contain affiliate links. If you click them, I might earn a small commission (costs you nothing, and helps me keep shipping quality articles every day for your reading pleasure).

---

75 out of 90 primary research citations trace to one format: the data benchmark. If you're shipping comparison content without production monitoring, you're watching citations die from 404s and redirects. The demo-vs-product checklist in the kit covers the visibility layer most builders skip.

→ Get the welcome kit