Yes, but almost certainly not for the reason it was sold to you. In May 2026 Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026, matched them against 4,000 control pages, and measured the change in citations: plus 2.4 percent in Google AI Mode and plus 2.2 percent in ChatGPT, both close enough to zero that the researchers classed them as noise. Google AI Overviews moved the other way, down 4.6 percent, and Ahrefs said plainly it could not pin that on the markup either.
That is the best-controlled public study anyone has run on this question, and it lands like cold water on a year of conference slides. It is also not the whole answer, because two of the four companies that matter have said opposite-sounding things on the record, and both of them were telling the truth. The useful version of this question is not whether schema works. It is which parts of it are still buying you something, and which parts are a line item your agency has not had the nerve to delete.
What the largest test of schema and AI citations actually found
Ahrefs started where everyone else stops. Across 6 million URLs, pages cited by AI were almost three times more likely to carry JSON-LD than pages that were not, and 53 percent of AI-cited pages had schema on them. That gap is real, and it is the number that has been circulating for eighteen months as proof that structured data is an AI visibility lever.
The problem is that schema lives on well-maintained sites. Sites that bother with structured data also publish better content, fix their technical debt, earn links, and rank in ordinary search. The markup could be doing the work, or it could be a passenger on a vehicle driven by everything else. To their credit, Ahrefs went and checked.
The design is the part a CEO should care about. They found pages that flipped from no schema to schema on a known date, gave each one three control pages from other domains with similar citation levels beforehand, and compared how the two groups moved over the following 30 days. That control set is what separates a real effect from a platform trend. Raw before-and-after growth in AI Mode looked like plus 43 percent, which would make a spectacular slide. The control pages, which added nothing at all, gained almost as much. Strip out the trend and the effect shrinks to 2.4 percent.
One caveat matters more than the headline. Every page in that dataset was already being cited heavily, with more than 100 AI Overview citations before any schema was added. So the study answers a narrow question well: if a page is already inside the consideration set, markup will not push it higher. It does not answer whether schema helps a page nobody has noticed get crawled, parsed, and included in the first place. Ahrefs says so themselves. Anyone quoting this study as proof that schema is dead is quoting the half they liked.
Why Google and Microsoft appear to disagree, and why both are right
Google’s position is documented and blunt. The Search Central page “AI features and your website”, last updated in December 2025, states that there are no additional requirements to appear in AI Overviews or AI Mode, that a page only needs to be indexed and eligible for a snippet, and, in as many words, that there is no special schema.org structured data you need to add. The same page also says you do not need to create new machine readable files or AI text files, which is worth remembering the next time someone quotes you for one. We covered that particular fashion in what is llms.txt.
Microsoft says something that sounds contradictory. Fabrice Canel, a principal product manager on Bing, told SMX Munich in March 2025 that schema markup helps Microsoft’s large language models understand your content, referring specifically to Copilot. That is a named engineering lead confirming it in public, which is rarer than the industry’s confidence on this subject would suggest.
The two statements reconcile once you notice they are describing different moments. Schema is extracted when a page is crawled and indexed. It is not extracted when a chatbot reaches out and grabs your page live. In October 2025 the German firm searchVIU built a test page with prices hidden in eight different places and asked five AI systems what things cost. A price that existed only inside JSON-LD was found by none of them. Gemini scored best at four of eight and was the only system to render JavaScript. ChatGPT found three, Google AI Mode two, Perplexity one, Claude none at all. Visible microdata and visible RDFa were read. Hidden markup, in every format, was ignored.
So the honest picture is a layered one. Where a search index sits between your page and the answer, as it does with AI Overviews, AI Mode and Copilot, your markup is in the record and can be used. Where a model fetches your page live, as ChatGPT and Claude often do, your markup is invisible and only the prose survives. The practical rule falls out of that on its own: if a fact matters, write it in a sentence a human can read. The JSON-LD is a second copy for the filing cabinet.
Schema is a legibility tax. You pay it so a machine does not have to guess who you are, and you should stop expecting a receipt.
Which schema types are worth the effort, and which are theater
Start with the theater, because it is expensive and easy to stop. FAQPage markup is finished. Google deprecated the FAQ rich result outright, the results stopped appearing on May 7 2026, the Search Console report goes in June and API support in August, according to the notice on Google’s own FAQ structured data documentation reported by Search Engine Journal in May 2026. HowTo went the same way in August 2023. Both are still sitting in AI-readiness checklists being sold this quarter. That sentence could describe a schema audit or a timeshare presentation.
What survives is the unglamorous end of the vocabulary: the types that tell a machine who you are rather than what shape your page is. Organization with sameAs links. Person for real authors. Article with an honest byline and date. BreadcrumbList. LocalBusiness if you have locations. Product and Offer if you sell things, where Google’s own AI features guidance points you at keeping Merchant Center and Business Profile current rather than at any clever markup. None of that will win you a citation on its own. All of it reduces the number of things a retrieval system has to guess.
Mark the entity, not the article
One Organization block, site-wide, with sameAs pointing at your LinkedIn, your Crunchbase entry, and anywhere else your company is already described. This is a half-day of developer time, once, and it is the highest-value markup you own.
Stop paying for FAQ schema
Google stopped rendering FAQ rich results on May 7 2026. Leaving existing markup in place is harmless. Commissioning new FAQ blocks in 2026 is not a strategy, it is an invoice.
Make the markup match the page
Google’s AI features documentation lists this as a fundamental. Any number that lives in JSON-LD but not in visible copy is invisible to a live-fetching model and a compliance risk everywhere else. Ask for a one-page audit of every mismatch.
Write the facts in prose
Pricing model, locations, founding year, who you serve, what you refuse to do. Four of five AI systems tested by searchVIU read only what a human could see. If a fact exists solely in markup, treat it as unpublished.
Hold a control group
Before you change anything, record baseline citations for ten pages you will touch and ten you will not. Ahrefs published the method; it takes an afternoon. See how to measure brand in AI search for the wider scoreboard.
The part nobody wants to hear
Schema keeps its grip on marketing budgets because it is a task. It has a ticket number, a named owner, a completion date, and a validator that turns green when you are done. You can report it to a board. The thing that actually determines whether a machine cites you is far less tractable: whether your company has spent years saying things specific enough to be worth quoting, in places credible enough to be indexed, consistently enough that three different systems recognise the same entity. Nobody can put a completion date on that.
So the work bifurcates. The tractable half gets done, gets reported, and moves nothing. The consequential half belongs to whoever has ten free hours a week to think about positioning, and in a company doing thirty or eighty million dollars, that person does not exist. What you end up with is a growing collection of capabilities purchased first and outcomes defined later, if at all. A markup audit, an llms.txt file, a visibility dashboard, an agency retainer, and no operating system connecting any of them to a decision anyone will make.
The companies that get cited are not the ones with the most complete structured data. They are the ones where somebody decided what the company is for, wrote it down in plain language, and then repeated it without flinching for long enough that the machines stopped guessing. What gets measured gets cited, but only after somebody has decided what is worth measuring.
Fix your Organization markup this month, delete the FAQ line item, and then go spend the saved budget on having something specific to say.
Markup only helps if the crawler can see it; see can AI crawlers read our JavaScript website?