Short answer: the best causal evidence available says adding schema markup does not increase how often AI systems cite your pages. A May 2026 Ahrefs study of 1,885 pages that added JSON-LD found no meaningful citation lift on Google AI Overviews, AI Mode, or ChatGPT. Most AI engines never read your JSON-LD at all, for two separate technical reasons. Schema is still worth auditing, but for narrower reasons than the AEO pitch suggests, and the audit that matters looks different from the one most teams run.
If you came here for a list of schema types ranked by citation impact, that list does not survive contact with the data. Here is what does.
The study that should have ended the schema-for-citations pitch
On 11 May 2026, Ahrefs published a controlled test by Louise Linehan and Xibeijia Guan. They identified 1,885 pages that added JSON-LD between August 2025 and March 2026, matched them against roughly 4,000 control pages with similar prior citation levels, and measured citations for 30 days before and after the markup went live. A matched difference-in-differences design strips out platform-wide trends, so a lift in the treated group that the control group did not share would be attributable to the schema.
There was no lift. AI Mode moved +2.4% and ChatGPT +2.2%, both small enough that Ahrefs classes them as noise across thousands of URLs. Google AI Overviews moved down 4.6%, and that one is statistically significant, with roughly a 1-in-2,500 chance of appearing by luck. Ahrefs does not claim the decline was caused by schema, and neither will I. Both treated and control pages were already declining beforehand.
The one statistically significant result in the largest causal test of schema and AI citations points in the opposite direction from what the industry has been selling.
The scope limit matters more than the headline, though, and most write-ups skipped it. Every page in the dataset already had 100 or more AI Overview citations before schema was added. So the study answers a precise question: does bolting JSON-LD onto pages that AI systems already find and cite make them cite those pages more? No. It does not tell you what happens when a page starts from zero citations, or what happens over six months instead of 30 days. That distinction is doing a lot of work, and I will come back to it.
The correlation everyone quotes, and why Ahrefs went looking for causation
Ahrefs began that project with a broader pass over 6 million URLs, and found that AI-cited pages were almost three times more likely to carry JSON-LD than uncited pages. That figure has been recycled across conference decks and LinkedIn carousels as proof that schema drives citations.
Linehan and Guan explain the gap themselves, and the explanation is the obvious one: schema markup tends to live on better-maintained, more technically sophisticated sites, and those sites also publish stronger content, earn more links, and rank better in organic search. Schema markup is a symptom of a well-run site, not the cause of its citations. The markup rides the wave. It does not make it.
This is worth internalising because the correlational finding keeps reappearing in new clothes. A September 2025 preprint on a 16-pillar auditing framework audited 1,100 URLs against 1,702 citations from Brave, Google AI Overviews, and Perplexity, and reported structured data among the pillars most strongly associated with citation, behind metadata/freshness and semantic HTML. Association, again. It is a preprint, not peer-reviewed work, and its design cannot separate schema from the site quality that tends to accompany it.
Why most AI engines never see your JSON-LD
This is where the technical picture gets genuinely useful, and where the usual framing of the problem is wrong. Two independent mechanisms block structured data from reaching most models, and they are often conflated.
The first is extraction. JSON-LD lives inside a <script type="application/ld+json"> tag. The HTML-to-Markdown pipelines that AI search systems use to prepare a fetched page for a model routinely strip or deprioritise <script> tags, so the markup is discarded before the model ever sees the document. Otterly.ai tested this directly. In a GEO experiment published 23 March 2026, Rick Tousseyn implemented five schema types across the Otterly site on 7 December 2025, then asked seven AI platforms to fetch and return the raw markup from specific URLs. Six of seven could not. Gemini returned the correct BlogPosting and WebPage JSON-LD. Claude reported that no schema was present. ChatGPT could not detect it, including in deep research mode. Perplexity was blocked at the fetch. Copilot could reach readable text but not raw HTML. Google AI Mode was the worst outcome: it confidently described schema types including a Service schema that had never been implemented on the page.
Tousseyn ran a third test that I find more persuasive than either citation-count comparison. He placed a fact inside FAQ schema that existed nowhere else on the site or the wider web, then asked the platforms a question only that markup could answer. None of them used it, even when a prompt pointed at the specific page.
If a fact lives only in your JSON-LD, most AI engines cannot retrieve it, which puts a hard ceiling on any theory that schema feeds AI answers directly.
The second mechanism is rendering, and this is the one the "AI crawlers don't run your JavaScript" line is really about. Vercel and MERJ's crawler analysis, published 17 December 2024 and repeatedly re-confirmed since, covered 569 million GPTBot requests and 370 million from Claude across Vercel's network in a single month, and found no evidence of JavaScript execution by either. GPTBot downloads JavaScript files in roughly 11.5% of requests and ClaudeBot in about 23.84%, and neither runs them. PerplexityBot behaves the same way. The exception is Gemini, which inherits Googlebot's rendering infrastructure, which also explains why Gemini was the one platform in Otterly's test that could read raw markup.
These two problems need different fixes, and treating them as one thing is how teams end up doing the wrong work. Script-stripping means schema will not reach most models no matter how you serve it. Client-side rendering means your actual content, the prose, will not reach them either. The second is far more damaging than the first, and it is the one worth panicking about. A page can rank well on Google, which renders JavaScript, while being an empty shell to every non-Google AI system.
Where structured data demonstrably still matters
None of the above makes schema pointless, and the strongest counter-evidence comes from Google itself.
At Search Central Live Toronto on 21 April 2026, Google structured data engineer Ryan Levering said that schema is used as context served to models during fan-out, according to conference recaps from attendees. Query fan-out is documented in Google's own guidance as the set of concurrent related queries AI Mode and AI Overviews generate to gather information. Per the session reports, Levering's reasoning was pragmatic: parsing structured data is cheaper for Google than repeatedly inferring meaning from raw text, markup can carry identifiers and metadata that never appear in visible content, and precise markup keeps machines from pulling in irrelevant context. He also said to be more precise with schema and that spamming it will not help.
Microsoft has been more direct. Fabrice Canel, a principal product manager at Bing, said at SMX Munich in March 2025 that schema markup helps Microsoft's LLMs understand content, as reported by Search Engine Land. For B2B and enterprise audiences skewing toward Copilot, that is the one platform where a vendor has confirmed its models consume your markup.
So the honest reconciliation is a split by consumer, not a single verdict. Google's own AI surfaces and Bing/Copilot sit downstream of indexes that parse structured data. Third-party engines fetching your page in real time mostly do not see it. Same markup, different fate depending on who reads it.
Google's guidance is also blunt about how much weight to put on this. Its generative AI optimisation guide, last updated 10 July 2026, lists "overfocusing on structured data" under things you do not need to do, stating plainly that structured data is not required for generative AI search and there is no special schema.org markup to add, while recommending you keep using it for rich result eligibility. The same page carries a requirement that is far more consequential and far less discussed: to be eligible for generative AI features, a page must be indexed and eligible to appear with a snippet, and the site must be included in Search generative AI features in Search Console.
Eligibility gating in Search Console is a binary prerequisite for AI Overview appearance, which makes it a higher-priority audit item than any schema type you could add.
The FAQPage and HowTo problem
If your audit template still treats FAQPage and HowTo as high-value AI plays, it is out of date.
Google added a deprecation notice to its FAQ structured data documentation on 7 May 2026, and FAQ rich results stopped appearing in Search on that date, per Search Engine Journal's coverage. This removed the last remaining eligibility, including for the government and health sites Google had preserved when it restricted the feature in August 2023. HowTo rich results went the same way on desktop back in 2023. Google published no blog post and gave no reason.
Three dates matter operationally. Rich results ended 7 May 2026. The FAQ search appearance filter, the rich result report, and Rich Results Test support were removed in June 2026. Search Console API support for FAQ rich result data ends in August 2026, which is this month. If you run automated dashboards or BigQuery exports that pull FAQ rich result data through the API, those calls will start returning empty values rather than failing loudly, which is the worst kind of reporting break.
The two schema types most often recommended as AI citation plays, FAQPage and HowTo, are both retired as rich results. FAQPage remains a valid schema.org type and Google has said unused structured data does not cause problems, so there is no cleanup emergency. But note what Google has not said: it has not confirmed that it still uses FAQ markup as a signal. Treat leaving it in place as harmless rather than as a hidden benefit.
The audit that actually matters
Reordered by expected impact, given everything above. The three highest-impact items in a structured data audit have nothing to do with which schema types you have. They are about whether machines can reach the page at all — everything else is decoration on top of that.
1. Check what a non-rendering crawler sees. Disable JavaScript and load your highest-value pages, or fetch them with curl and read the raw HTML. Whatever is missing is missing for ChatGPT, Claude, and Perplexity. If your main content, pricing, or comparison tables only appear after hydration, stop the audit and fix that first. Nothing else on this list matters by comparison.
2. Confirm generative AI eligibility in Search Console. Per Google's guidance, being indexed is necessary but not sufficient. Verify the site is included in Search generative AI features, and that your target pages are indexed and snippet-eligible rather than sitting in "Crawled – currently not indexed."
3. Audit crawler access separately from Googlebot. Training crawlers and search crawlers are different agents with different consequences. Blocking GPTBot affects future training runs; blocking OAI-SearchBot removes you from ChatGPT's search results. Check robots.txt against the current agent list rather than assuming one rule covers all of them.
4. Then check accuracy, not coverage. Google's long-standing rule is that structured data must match visible content, and Levering's "be more precise" framing points the same way. The failure mode worth hunting is markup that contradicts the page: Product schema on service pages, aggregate ratings with no visible reviews, author entities that do not resolve. Automated schema plugins are a common source here; it is not unusual for a plugin to add markup that introduces dozens of new validation errors.
5. Check dateModified is honest. This is the item most likely to be quietly broken. dateModified should update on real content edits, not on every rebuild or deploy, and not be hardcoded to match datePublished. A dateModified that bumps on every deploy is not a freshness signal, it is noise, and Ahrefs' own analysis of citation freshness found the picture is more complicated than "newer wins" anyway: most AI assistants cite fresher content than the organic SERP, but Google AI Overviews cites content slightly older than organic on average.
6. Fix the FAQ API dependency before the end of this month. Covered above.
7. Ship JSON-LD server-side. Cheap, and it removes one variable. Google can process structured data generated by JavaScript; nothing else reliably can.
What to drop from the audit: ranking schema types by "citation impact," adding markup types you do not have matching visible content for, and any retainer line item priced on schema volume. llms.txt is also on Google's own list of things it ignores.
Frequently asked questions
Does schema markup help with AI citations? Not directly, on current evidence. The largest causal test, Ahrefs' May 2026 difference-in-differences study of 1,885 pages, found no meaningful lift on AI Overviews, AI Mode, or ChatGPT. Google's own guidance says no special schema is required for its AI features. Schema helps machines parse your content accurately, which is a different and more modest claim.
Is schema markup a Google ranking factor? No, and Google has said so consistently for years. It supports rich result eligibility and content understanding.
Why do some studies show huge AI Overview gains from schema? Read their controls. Otterly.ai's experiment reported a large AI Overview increase, then found in its own analysis that competitor brands who made no schema changes moved in parallel over the same period, which points to an algorithmic shift rather than a schema effect. That experiment's numbers are also reported inconsistently, with its per-platform table showing +611% for AI Overviews and its summary citing +1,500%, so treat single-site percentage lifts with care.
What about the Search Engine Land experiment where the no-schema page wasn't indexed?
It is the most-cited evidence for schema, and worth reading in full rather than in summary. Molly Nogami and Ben Tannenbaum, September 2025, built three single-page sites with good, poor, and no schema; only the good-schema page reached an AI Overview, and the no-schema page was crawled but never indexed. But all three were brand-new single-page Vercel deployments with no custom domain, no sitemap, no robots.txt, and no canonical tags, and each targeted a different keyword. A brand-new zero-signal page failing to get indexed has plenty of candidate explanations besides missing schema. The authors are appropriately careful, calling a coin-flip fluke possible and planning a follow-up that swaps the markup between pages.
Should I remove FAQPage schema now that rich results are gone? No. Keep it where the questions and answers are real and visible on the page. Remove it only where the Q&A existed purely to win the dropdown.
Does schema still matter for Copilot? This is the strongest case for it. Microsoft's Fabrice Canel confirmed in March 2025 that schema helps Bing's LLMs understand content, and Bing indexation is a documented precondition for appearing in Copilot answers.
Where this leaves the audit
The useful reframe is that structured data is technical hygiene with a real but bounded payoff, and it competes for engineering time against work with a much better evidence base. If a page is invisible to non-rendering crawlers, or not eligible for generative AI features in Search Console, no amount of correct JSON-LD will rescue it. Fix the gate before you decorate the room.
RankSage is in pre-launch. We are building a platform that joins GA4, Google Search Console, first-party behavioral data, and AI citation data at the page level across five engines (ChatGPT, Claude, Gemini, Perplexity, and Copilot), then ranks the actions worth taking.
An honest note on what that can and cannot tell you, since this post has been about resisting overclaims: page-level citation tracking shows you which of your pages get cited and how that changes over time. It cannot prove why. Nobody's tooling can currently isolate the causal contribution of a single technical change to an AI citation, which is exactly why the Ahrefs study needed a control group of 4,000 pages to say anything confident. What good measurement gives you is a reliable baseline and the ability to run your own controlled test instead of trusting an industry average.
Join the waitlist for early access.