schema Won't Save Your AI Citations - Study Says

budgetburner

Just went through a chunk of data on schema and AI citations. Nearly 2,000 pages added JSON-LD schema between last August and this March, matched against 4,000 control pages. The result? Zilch. No meaningful uplift on Google AI Overviews, AI Mode, or ChatGPT.

Here's the breakdown:

AI Overviews: citations dropped 4.6% relative to controls. Statistically significant but tiny.
AI Mode: up 2.4% - could be random noise.
ChatGPT: up 2.2% - same story.

So schema didn't move the needle. At all. Either it did nothing, or a microscopic amount that you'd never bet on.

Someone in the discussion pointed out the obvious flaws. This kind of test on a single site? If only part of the site has schema and the rest doesn't, you're not measuring cleanly. Schema only helps when there's ambiguity - like a term that could mean a person, a place, or a description. Most of the pages in that dataset? No ambiguity. So schema just adds noise. And bad or inconsistent schema? Actually hurts more than it helps.

Another reply said "QFO is the true way for citations." Couldn't agree more. If you're chasing AI visibility, schema is not the lever. Focus on content that answers questions clearly - that's what drives citations.

Bottom line: don't waste time adding schema hoping for an AI traffic boost. It's not the shortcut people think it is.

dusting

Exactly. QFO is the underlying mechanism, schema just decoration.

tiktokguru

Wow, this is another really interesting study from Ahrefs. The explanations of the methodology are very concise. The results, however, are actually a bit unexpected, to be honest ...

growthgeek

Contrarian take? Absolutely. I see Edward Strum banging this drum constantly - and I respect the guy, watch his content - but I'm calling BS on the narrative that schema doesn't move the needle for AI citations.

Let's break it down:

Ahrefs' own data kicks off by admitting most LLM-cited pages have structured data and solid foundations. So they're essentially saying "build a site that serves Google, crawlers, users, and LLMs equally" - and surprise, that works. That's not a counterargument, it's the whole game.
That 2.4% vs 2.2% difference? That's not "indistinguishable from zero" - that's a signal buried in noise. GPT pulls from stale data, crawlers aren't instant. Give it time. The lift compounds when indexing catches up.
I'm running my own tests across multiple sites. Last 28 days in GSC? 30-50% page-click improvements on pages where I added rich JSON-LD and cleaned up on-page signals. No negatives. Not a single drop.
Honestly, I think there's deliberate misdirection here from some "experts". Whether it's protecting old playbooks or pushing tools that don't support it well, I don't know. But the noise doesn't match what I'm seeing.
Clients are literally getting calls where the caller says "GPT recommended you as the best option". That's happening on newly optimised pages. Coincidence? Not buying it.

My gut says this is smoke being blown to keep people away from a winning edge. I'm stacking my own data, and it's telling a different story.

communityfirst

Wow, seeing that data on schema adoption is fascinating! JSON-LD has been such a game-changer for our visual content strategy at Pinterest-it's not just about rich results and voice assistants, but how it helps platforms understand the context behind every pin and image. Even if AI citations aren't moving yet, the downstream benefits for entity recognition and knowledge graphs are worth the effort. It's all part of building that bigger picture for discoverability, and every step counts.

dailygrind

i think there are a few fatal flaws with this test. running it on a single site where only part of the site has schema markup and the rest doesn't - that's not a controlled environment. You don't get the signal confidence AI systems need to treat it as a pattern. it's like testing a conversion rate optimisation tactic on one random page and declaring it dead.

second, schema is only really valuable where ambiguity exists. if you're writing about Ahrefs features, there's virtually no ambiguity - "Domain Rating", "backlinks", "keyword difficulty" - these terms are pretty well defined in the niche. Schema won't move the needle on citation accuracy because the AI already knows what you're talking about. it's when you've got something like "Sandy Shores" that could be a person, a place, or a property description that structured data helps the model disambiguate.

and honestly, bad or inconsistent schema can actively hurt more than it helps. this test seems to have applied it haphazardly across a subset of pages. that's a recipe for noise, not insight. Would love to see this rerun with a proper split-test methodology and across multiple sites with genuinely ambiguous content

backlinker

Borderline useless study. Did they give an intern an SPSS license?

budgetburner

Oh great, so now it's been blessed by the SEO high priests. That doesn't make the numbers any less useless. 1,885 pages with schema and no movement? either the schema was slapped on garbage pages, or the metric they're tracking is vanity fluff. Classic case of "correlation ain't causation" getting buried under big names sharing it.

digitalnomad

This is why I don't like the just add schema for AI visibility advice. Schema can reduce ambiguity, but it doesn't automatically make a page citation-worthy. If the visible content is weak, unclear, or not the best answer, JSON-LD won't fix that.

My take is schema is a clarity layer, not a shortcut. For AI citations, I would look more at source trust, visible answer quality, entity clarity, and whether the page is actually easy to extract from.

24fire

Not surprised either. Schema markup often gets lost in the tokenisation process - it's like shouting into a void when the algorithm isn't listening. Makes you wonder about the real weight of structured data in today's AI-heavy search landscape.

budgetburner

Exactly. Built a whole crawling infrastructure but can't figure out ranking without hand-holding. Classic.

gmbpro

Ahrefs aren't making money selling schema entity support - they're still deep in keyword and backlink territory. From a pure tracking standpoint, that 1,885 pages spike is impressive, but if AI citations barely moved, the actual measurable impact is close to zero. Would be interested to see whether other tools (like Semrush or Sistrix) picked up any change in organic visibility or if this is just a schema validation number with no ranking effect.

budgetburner

i call BS on that whole "AI citations barely moved" headline. Feels like someone cherry-picked a single dataset and ran with it.

Here's the thing-no one's saying schema is bad. Slapping JSON-LD on a page isn't going to hurt anything. it's like putting a lucky rabbit's foot on your desk. Doesn't break anything, but don't pretend it's driving revenue.

The guys claiming 30-50% click lifts from GSC over 28 days.... that's not an experiment. That's a snapshot. You want to prove schema works? Then turn it off on the same pages and see what happens. Nobody ever does that. Why? Because the results would kill their narrative.

and spare me the "customers are getting calls saying GPT recommended them" anecdote. Every SEO with a blog and a following has a story like that. It's the digital equivalent of "my uncle's mate's dog could predict the lottery." Anecdotal evidence hides behind a username so nobody can verify it. That's not data-it's a cult.

I've seen a 50-page site with zero schema hold 19,000 SERP positions for six months. presented it with a SEMrush report. Got roasted by anonymous profiles who scream "schema is essential" like it's gospel. Yet not one of them could point to a controlled test where turning it off tanked rankings.

Edward Sturm isn't some conspiracy theorist-he's calling out the lack of real evidence. If you can't test the inverse, you're not doing SEO. you're just following a trend that feels good.

dusting

Most of what I'm seeing in the thread misses the mark on methodology. Worth digging back into the study design.

They didn't test a single site. The tracked group was nearly 1,900 pages that added JSON-LD over a period, matched against roughly 4,000 control pages pulled from different domains with similar pre-existing citation counts. That's the opposite of a single-site experiment. They started with millions of URLs to find the ones that actually made the transition from no-schema to schema.

As for the "only part of a site has schema" argument - the study isn't measuring whether the site as a whole is better understood. It's comparing each page's AI citations over a 30-day window, accounting for broader platform trends with a difference-in-differences approach. Whether schema lives on half the pages or all of them is irrelevant to that design.

The disambiguation point also falls apart when you look at their parallel experiment. A separate test ran on multiple AI systems - ChatGPT, Claude, Perplexity, Gemini, AI Mode - and none of them read schema markup during real-time page fetch. Every single one pulled only the visible HTML. JSON-LD, hidden Microdata, hidden RDFa - all ignored at retrieval. If the AI isn't consuming the schema at the moment it decides what to cite, then writing it to clarify "Sandy Shores" is like whispering directions to someone wearing noise-cancelled headphones.

The "bad schema hurts" claim has no real data behind it. They observed a decline of roughly 4.6% in AIO citations for treated pages. But the absolute number is tiny - around a dozen daily citations out of pages that already get hundreds. Both groups were declining together before the change. They can't pin it on schema with any confidence. Honest reading: schema didn't help, might have slightly hurt, but the data isn't strong enough to say either way.

That 53% correlation - schema being present on cited pages - is just a reflection that sites investing in structured data also invest in everything that actually drives citations: solid content, authority, relevance. Strip the schema out and those other signals still carry the page. It's a vanity metric, not a causal lever

budgetburner

To what bloody end? identify it for what? Are search engines really supposed to lock themselves into "known entities"? that's just content fascism dressed up as structured data