“For AI, Start By Adding Structured Data”
That sentence sits at the top of a lot of agency proposals, and it sounds sensible enough that nobody in the room asks what it’s worth.
I didn’t ask either. I took the deck away instead.
What I did have was the number they’d put underneath it: pages that get cited by AI are about three times as likely to carry structured data. Say 3x in a meeting and the objections stop arriving.
But a ratio counted at one moment in time and an effect you get by adding something are two different animals, and the deck was quietly treating them as one.
Somebody has since run the second version of that question with a control group.
Does Adding Structured Data Get You Cited More?
It doesn’t. Across 1,885 pages that newly added , citations moved +2.4% in Google’s AI Mode and +2.2% in ChatGPT — both indistinguishable from zero (Ahrefs, 11 May 2026, source).
In Google’s AI Overviews the figure was −4.6%, and that one was statistically significant.
JSON-LD is the block of tags you drop into a page to spell out what’s on it in a form machines read: this is the product, this is the author, this is the price. It uses schema.org, a shared vocabulary, and the long-standing pitch is that search engines and AI systems are then less likely to get the contents wrong.
The study is Ahrefs’, written by Louise Linehan. Ahrefs sells AI visibility measurement tools, so this result points away from the direction that would sell more software.
Where The 3x Actually Comes From
From an observational count across 6 million URLs.
53% of the pages that got cited had structured data, about three times the share among pages that didn’t get cited. That count isn’t wrong.
It’s a ratio between two groups measured at the same moment, though. Nothing in that arithmetic says the number moves when you add the tags.
Sites that keep their structured data tidy usually keep everything else tidy too. They have people doing the updates, they publish often, they get mentioned elsewhere.
Read the 3x as the mark left by that whole bundle of upkeep and it stops being mysterious.
It’s the same shape as a health survey finding that people who eat breakfast get better grades. Handing out toast is not the intervention.
Put a correlation in an internal deck with no label on it and this swap happens by itself, because the number looks like it’s already been checked.
What The Control Group Made Visible
Ahrefs built a before-and-after comparison where the pages that added tags are matched against pages that didn’t.
- Treatment: 1,885 pages that newly added JSON-LD between August 2025 and March 2026
- Control: three pages per treated page from different domains, matched on prior citation levels, with no tags added (about 4,000 pages)
- Measure: citation counts in the 30 days before and the 30 days after the tags went in
Four different statistical approaches were tried on top of that, and they all land the same way.
The control is the part doing the work here. If AI citations happened to be rising across the board in the month the tags went in, a plain before-and-after would have read that rise as the tags earning their keep, and the deck would have been vindicated by nothing at all.
Matching against other pages over the same window is what separates the tag from the weather. So how far does the finding stretch?
What This Test Can And Can’t Tell You
The pages studied were already heavily cited by AI.
The population was pages that had picked up 100 or more citations from Google’s AI Overviews as of February 2025, which means the study says nothing about a page that has never been cited once.
Then there’s confounding — the situation where the thing you’re watching arrives alongside other changes, and you can’t tell which of them did the work. Pages that get JSON-LD added tend to get links, copy and technical fixes touched in the same stretch.
The remaining caveats matter most at the point where you try to apply this to your own site.
- Schema types (Article, FAQ, Product, HowTo, Organization and the rest) are pooled together, so per-type effects were never tested
- The observation window is 30 days after treatment only; 60 and 90 days weren’t looked at
- Only JSON-LD written directly into the HTML counts. Microdata, RDFa and formats injected later by JavaScript are out of scope
- The 4.6% drop in AI Overviews is real, but the publisher says plainly that they can’t explain why it happened
I’d take that last one at face value. If nobody knows why the number fell, “strip the tags out and it’ll come back” isn’t a conclusion you can draw from it either.
So Where Should Those Hours Go Instead?
Into numbers and sources inside the copy, and into third-party reviews outside it. Both have been measured for effect size in this same citation context, which is more than the tags can say.
One tactic has been ruled out, which frees the hours rather than losing them.
On the copy side, attaching figures and sources to your claims raised how often pages got cited (the trick to getting cited by AI is boring: add numbers and sources).
Off-site, products with third-party reviews behind them have swung from a 1% citation rate to 75% (reviews take your AI citation rate from 1% to 75%).
Both of those move something a reader would notice: what the page says, and how much the rest of the internet says about you.
Does that mean structured data comes out of the plan? No.
Rich results — the price and star ratings that show up inside a search listing — were never in scope for this test, and schema still does that job as well as it ever did. The decision isn’t whether to do it. It’s which goal it gets filed under.
Keep Technical Tags And AI Citations In Separate Ledgers
There’s one thing to settle today, and it’s the filing.
Structured data goes in for the search-side use, and gets judged on search-side metrics. It doesn’t get attached to a target for AI citations.
Keep both in one ledger and the month the citations don’t move, you’ll have no idea which half of the ledger to interrogate — and that’s the month somebody asks you whether the whole AI visibility line item is worth funding.
Then take the freed hours to the page contents and the off-site surface, and measure your AI citations before and after so you can see the difference. Order the work by what you measured yourself, not by a 3x someone counted sideways.
Do that and it stops mattering what the first line of the next proposal says. You can ask for the effect size.
Dull conclusion, I know. The dull ones are the ones that reproduce.
Sources
- Louise Linehan (Ahrefs), “We Tracked 1,885 Pages Adding Schema. AI Citations Didn’t Follow”, Ahrefs Blog, 2026-05-11, https://ahrefs.com/blog/schema-ai-citations/
