The 2026 Reflex: Marking Up Everything in JSON-LD to "Get Cited by ChatGPT"
Since Google rolled out AI Overviews and AI Mode in France, the reflex has become widespread: add Organization, Article, FAQPage schema on every page possible, hoping it tips the balance toward citations in ChatGPT, Perplexity, or Google's generated summaries. The reasoning seems logical — more structure, more clarity for the machine, so more chances of being picked up.
Except a study published by Ahrefs in 2026 just measured this seriously, on a large sample, with a control group. And the result breaks a good chunk of the myth.
What the Ahrefs Study on 1,885 Pages Reveals
Between August 2025 and March 2026, Ahrefs tracked 1,885 pages that had just added JSON-LD markup, compared against a control group of 4,000 unchanged pages. The goal: use a difference-in-differences analysis to measure whether adding schema moved the needle on citations across three AI engines.
| Engine | Citation Change | Interpretation |
|---|---|---|
| Google AI Overviews | -4.6% | Significant decline, still unexplained |
| Google AI Mode | +2.4% | Indistinguishable from statistical noise |
| ChatGPT | +2.2% | Indistinguishable from statistical noise |
The researchers' conclusion: adding schema produced no major citation gain on any of the three engines tested. AI Overviews even declined slightly — a result the study itself can't fully explain, and one that calls for follow-up research.
An important limit to keep in mind: the study only looked at pages already receiving more than 100 AI citations before the test. For a page that's never been cited, or is still poorly indexed, schema might play a different role — removing an ambiguity that was blocking the AI crawler's understanding of the content in the first place, before citation even enters the picture.
Why the 53% Correlation Doesn't Mean Causation
One number keeps coming up to defend schema markup: roughly 53% of pages cited by generative AI engines use structured data. That's true — and it proves nothing about causation. A well-marked-up page is usually a well-built page: clear content, logical structure, serious technical maintenance. Schema correlates with overall site quality; it isn't the cause of the citation.
It's the same bias as saying "sites with an SSL certificate sell more": statistically true, useless as standalone advice, because SSL isn't what drives sales — it's a technical prerequisite that sites which sell well for other reasons also happen to have.
What Schema Markup Actually Does (and What It Never Does)
- It disambiguates: it tells an AI engine that this block of text is an article, that this author is an identified person, that this price belongs to this specific product.
- It eases extraction: an engine that has to guess a page's structure wastes time and runs a higher risk of misinterpretation.
- It never forces citation: no markup compensates for thin content, content that duplicates ten other pages, or content that doesn't actually answer the question asked.
- It doesn't replace authority: E-E-A-T signals (identifiable author, cited sources, editorial consistency) weigh far more heavily on citation decisions than a well-formed JSON-LD block.
For the technical implementation — syntax, validation, prioritization by page type — our full Schema.org guide for SEO and GEO walks through each schema step by step. The goal here is different: understand what markup can and can't do before you spend hours on it.
The Schemas That Still Matter in 2026, and the Order to Deploy Them
| Schema | Why Prioritize It | Effort |
|---|---|---|
Organization | Identifies who you are — the baseline disambiguation for any AI | Low |
Article / BlogPosting | Structures date, author, topic — the minimum condition to be picked up cleanly | Low |
FAQPage | Q&A format directly usable by conversational engines | Medium |
Product + Review | Useful for AI comparison engines on transactional queries | Medium |
BreadcrumbList | Helps place a page within the site hierarchy | Low |
What these five schemas have in common: none of them "sells" content on your behalf. They make what already exists legible. If your content is thin, schema will just make it... thin, but neatly labeled.
How to Measure the Real Effect on Your Own Site
Rather than trusting a general study — even a rigorous one like Ahrefs' — the only number that really matters is your own site's. The method we use in audits comes down to three steps:
- Build a panel of 20 to 30 prompts representative of the queries where you'd want to be cited (not your brand name — the actual questions your audience asks).
- Run this panel before any markup change, and note for each prompt whether you're cited, by which AI, and with what exact excerpt.
- Rerun the same panel 4 to 8 weeks after deploying the schema (enough time for AI engines to recrawl and reindex), then compare.
To put a concrete order of magnitude on it: an ecommerce site with a panel of 25 product-page-related prompts, starting at 3 citations before markup, and reaching 4 after adding structured Product and Review schema to 40 product pages, fits the scenario Ahrefs measured — a real but modest improvement, not a miracle effect. And in that kind of case, the real gain rarely comes from schema alone: it usually comes from the product description rewrite that tends to accompany the markup rollout.
This method has one advantage the Ahrefs study doesn't: it isolates your context, your niche, your direct competitors — variables no general study can capture on your behalf. To go further on the metrics worth tracking over time, our guide on measuring AI visibility and GEO KPIs covers the full dashboard.
FAQ
Should I stop adding schema markup in 2026? No. Schema remains a solid technical practice that disambiguates your content and eases extraction by AI engines. What changes is the expectation: stop treating it as a standalone citation lever, and start treating it as a technical hygiene prerequisite, the same way you'd treat a clean sitemap.
Which schema gives the best return for the time invested?
Organization and Article/BlogPosting first — low effort, immediate disambiguation benefit. FAQPage next if your content already answers concrete questions: the format will fit naturally onto the markup, with no rewrite needed.
Does schema act the same way on Google AI Overviews and on ChatGPT? No, and the Ahrefs study confirms it: the three engines tested reacted differently (slight decline on AI Overviews, neutral effect on AI Mode and ChatGPT). AI engines don't share the same crawl pipelines or source-selection criteria — measuring on one engine tells you nothing about the others.
Key Takeaways
Schema markup remains a solid technical practice — it disambiguates your content, eases extraction, and belongs in any solid SEO/GEO foundation. But in 2026, the available studies converge on one point: it doesn't, on its own, trigger a flood of citations in AI engines. The real lever is still the content itself — its clarity, its freshness, its actual authority.
Before spending hours marking up a hundred pages, get your score out of 100 with a SeAudit audit to find out where your content is actually blocking citation, or check a sample report to see the level of detail in a full SEO + GEO analysis.