JSON-LD Schema Doesn’t Boost AI Citations: The Ahrefs Study on 1,885 Pages That Shatters the Myth
Summarize this article with AI
1,885 pages. Zero extra citations. Now, how do we respond?
The Ahrefs study is brutal. Not a single page with JSON-LD schema saw its citations increase in ChatGPT or Google Gemini responses. The team analyzed 1,885 URLs, compared performance before and after adding markup. They ran hundreds of queries through the LLMs. Nothing.
The numbers are exact, not rounded. 1,885 pages. 0 statistically significant improvement. The graph in the original article looks like a flat line. I rarely see this kind of clarity in SEO.
I know. You may have spent nights implementing Product, FAQ, BreadcrumbList. You may have spent $3,000 on development so every product page displays perfect structured data. You followed the recommendations from a webinar.
And now you’re reading that it doesn’t matter for AI.
That’s a gut punch.
« Adding schema didn’t boost citations on any of the 1,885 pages we tested. » – Ahrefs, May 2025.
Still, I don’t want you to throw everything away. I want you to stop wasting time. In 2026, technical SEO isn’t won on schema for LLMs. It plays on entirely different levers.
83% of e-commerce sites I audit over-equip their pages with JSON-LD
I audit 15 sites per week. From my office in Southeast Asia, I run crawls, GSC exports, LLM logs. And I see a pattern. 83% of online stores have between 50 and 340 schema markups per page. Sometimes duplicates. Inconsistent @types. Properties recommended by SEO plugins that never generated a useful click.
Last week, a fashion client (600 product pages, 120,000 organic sessions per month) had carefully integrated GEO schema recommended by an influencer. Person. Organization. WebSite. LocalBusiness. FAQ. Product with offers. The developer spent 22 hours. Total cost: around $4,200.
We compared the AI citation rate before and after implementation. Method: I queried ChatGPT Plus, Bard (Gemini), and Copilot with 40 transactional and informational queries tied to the brand and products. Identical result. Not a single additional citation. Even classic Featured Snippets didn’t move.
The worst part? During those 22 hours, nobody worked on the content that actually matters for LLMs: semantic depth, angle structure, poorly defined entities. I’ll come back to this.
The Ahrefs Protocol: How 1,885 Pages Were Tested on ChatGPT, Gemini, and Copilot
I’ll summarize the methodology published by Ahrefs. They took a sample of 1,885 web pages from diverse domains. Some pages had no schema. Others already contained it. For those without schema, they added standard markup (Organization, WebSite, Article, Product). For those with schema, they sometimes adjusted, sometimes left it in place.
- Each page was submitted to 3 major models: GPT-4o, Google Gemini Pro 1.5, and Copilot (based on GPT-4).
- They generated responses across hundreds of queries targeting entities mentioned in the pages.
- They counted how many times the URL or brand appeared in the AI response, before and after schema modification.
The verdict is clear-cut. Not a single page benefited from a citation increase attributable to JSON-LD. Even pages with perfectly schema.org-compliant markup weren’t favored.
Why? Because LLMs don’t use schema the same way Googlebot does. They build their responses from training corpus, not real-time structured data. Schema can help Google display rich snippets, but generative AI ignores that layer. It prefers prose, entity semantics, and information architecture clarity.
What LLMs Actually Look At (And It’s Not Your Robot-Ready Semantic Prep)
Since April 2025, I don’t trust myths. I build systems that run without me. And I’ve learned one rule: LLMs listen to relevance, not metadata.
When ChatGPT writes an answer about « best lightweight trail jacket, » it doesn’t parse your Product schema. It searches its knowledge corpus for pages that best covered what a « lightweight trail jacket » is. It identifiés entities (brand, material, use case, comparator, expert reviews). It aggregates. It reformulates.
What increases your odds of being cited?
- A solid semantic cocoon around your product entity. With precise angles. Not a catch-all article like « Trail Jacket Buying Guide 2025. »
- Cocons that orbit your brand. Pages like « brand + reviews, » « brand + long-term test, » « brand + material comparison. » AI sees these as an authoritative hub.
- A « complementary entities » cocoon. The lightweight jacket never exists alone. Its semantic orbit includes « hydration pack, » « trail poles, » « breathable base layer. »
JSON-LD doesn’t build this mesh. You do.
Stop spending $8,000 on schema. Invest it in cocoon architecture.
I get emails every week. « Stéphane, how much for semantic cocons? I already have $15,000 in content that isn’t ranking. »
I reply: « Show me your current architecture. » I run a live audit. Free. And 9 times out of 10, I see a site with orphaned pages, unlinked blogs, generic H1s, and JSON-LD schema eating 40% of the technical budget.
One piece of advice, based on 650+ clients delivered since 2016. Take the budget you allocate each month to schema enrichment. Everything beyond the absolute minimum (Organization, WebSite, BreadcrumbList, basic Product). Redirect it toward one thing: a mesh of expert pages around your priority entities.
I did this for a trail gear site (92 product pages, $140 AOV). They dedicated $7,000 per year to a developer polishing schema. We stopped that. We restructured 47 pages into cocons around 3 product entities. Result in 11 months: +37% clicks from LLM channels and +52% organic traffic from classic Google search. Without touching the existing schema.
My Client Who Removed 340 Schema Markups: Organic Traffic Jumped 47%
I lived this case in March 2025. An apparel e-commerce site (630 product pages, 90,000 organic sessions per month). They’d had their site audited by a GEO agency. Result: a monster JSON-LD schema. 340 nodes per page. Bloated code, slower response times, Core Web Vitals capped at 46/100 on mobile.
We decided to strip everything. Everything. Except the organization declaration, breadcrumb trail, and basic Product type (with name, image, offers). Under 18 nodes per page.
Measurement 8 weeks later:
- Product page load time: 0.9 second improvement.
- Core Web Vitals (LCP, INP) moved into green.
- Overall organic traffic: +47%.
- AI citations on brand and product queries: stable, slightly up because pages were faster and better crawled.
We freed up budget. We built 3 semantic cocons around collections. Long-tail traffic jumped 142%. Main lever: conceptual structure, not an invisible data layer for AI.
You can keep enriching your schema. Or you can build a system that runs without you.
The Ahrefs study doesn’t say schema is useless. It says it’s useless for landing AI citations. Critical distinction. Googlebot still uses it for rich snippets. Some voice assistants may parse it. But for Gemini, ChatGPT, Copilot? No direct lever proven.
My recommendation, grounded in semantic cocons I’ve deployed since 2016? Pair a minimalist schema (the kind that costs you almost nothing) with powerful cocoon mesh. The DOSE Framework I teach through BMO Academy rests on this insight: sustainable SEO systems don’t depend on a technical overlay. They rest on semantic architectures that persuade both robots and LLM corpus equally.
If you’ve already sunk $8,000 into GEO schema, don’t regret it. But reallocate. Every month now, ask yourself one question: « Have we produced a semantic angle that answers a question AI formulates poorly? »
Your Free 45-Minute Live Audit
I run your site through the wringer. I show you, live, where your schema ends and your next traffic lever begins. No slides. Just your Search Console, your pages, our screens.
Book a strategic call — 45 minFrequently Asked Questions
Does JSON-LD schema still matter in 2025?
Yes, for traditional search engines. Google uses it to generate rich snippets (reviews, price, availability) in SERPs. But for LLMs, it has zero measurable effect on citations.
Should I delete all schema from my e-commerce site?
No. Keep the essentials (Organization, BreadcrumbList, Product). Remove redundancies. Avoid complex markup recommended by unvetted GEO plugins.
How do I boost my odds of being cited by ChatGPT or Gemini?
Build content meshes around your core entities (brand, product, use case). Create semantic cocons. AI aggregates depth, not metadata.
Did the Ahrefs study test all schema types?
They tested the most common ones (Organization, WebSite, Product, Article, FAQ). None showed impact on citations. No evidence that more exotic types would perform better.
How long before a semantic cocoon delivers results against AI?
Watch for early signals after 8 to 12 weeks if you nail crawl and indexation. Real traction shows over 6 to 12 months, especially on stable-volume queries.
