Do AI Assistants Read JSON-LD at Answer Time?

"Machines read structured data, so feed the AI machines structured data" is the syllogism holding up a lot of schema retainers. It fails at the first measurement: when researchers watch what AI systems actually pull from a page at answer time, the markup goes unread. This page separates the two pipelines — what reads JSON-LD, what reads your visible words — and says where the effort belongs, from a team that ships both on every page.

Do AI assistants read JSON-LD when building an answer?

The best available evidence says no. Ahrefs' May 2026 schema study reports a searchVIU experiment that tested whether AI systems consume structured data during direct retrieval, and the result was unambiguous: "every system extracted only visible HTML content. JSON-LD, hidden Microdata, and hidden RDFa were all ignored."

Evidence tier, stated plainly: that is one third-party test, reported secondhand by a source on our closed list — supported, not platform-documented. What tips the scale is the silence on the other side: no AI platform's documentation on our source list claims its assistant parses JSON-LD at answer time, and Google states outright that its AI features require no special structured data. A tactic with zero platform documentation and a failed direct test is not where budget goes first. How answers actually get assembled — retrieval, chunking, synthesis — is walked through in how AI search works.

What did the controlled schema test find?

That adding JSON-LD moved AI citations by roughly nothing. Ahrefs tracked 1,885 pages that added schema between August 2025 and March 2026 against about 4,000 matched control pages, using difference-in-differences analysis to strip out platform-wide trends:

Ahrefs, 1,885 treated pages vs ~4,000 controls, August 2025–March 2026, published May 2026.

The study's honest limit belongs next to its headline: it measured pages already visible to AI systems, and its author notes schema "might still play a role" for pages those systems have not discovered at all — an open question, labeled as one. The full accounting of that study, including the 53% correlation that launched a thousand vendor decks, is in what the evidence shows about schema and AI citations.

What is JSON-LD actually for, then?

Google's pipeline — which is a real and documented consumer, just not the one being sold. Google's structured-data introduction describes markup as helping its systems understand page content, and as the eligibility mechanism for rich results: the enhanced listings Google grants for specific, documented types. Those guidelines also carry the rule that settles the visible-vs-hidden question by fiat: markup must reflect the visible content of the page, not claims the reader cannot see.

Entity disambiguation is the second documented job. Organization markup and its relatives state in machine-readable form who published a page and how the publisher relates to its other properties — plumbing for any system resolving who asserts this. Both jobs survive the AI null result untouched, and both are why the markup stays.

Where should the effort go — markup or visible words?

Visible words, by every evidence tier we track. Platform-documented: Google's AI features need no markup, only an indexed, snippet-eligible page. Study-supported: the Princeton GEO benchmark measured visibility lifts of up to 40% from content-level changes — adding citations, quotations, and statistics to the text — and its wins were in the words, not the markup layer. Field-observed, with its limit stated: the 1 AI Overview citation our fleet can document went to a page whose visible opening line is a 1-sentence definition written to be lifted whole [our data] — consistent with the retrieval evidence, though no outside observer can see which layer the engine actually read.

The operational rule that falls out: any fact you want an engine to quote must exist as a self-contained visible sentence in the served HTML. On our builds every page ships server-rendered, so the direct answer is present in the raw response a crawler fetches, not assembled later by JavaScript [our data] — extraction cannot quote what the fetch never contained.

That rule comes with a 30-second test we run on every template change: fetch the page the way a bot does — curl with an AI crawler's user agent — and search the raw response for the answer sentence. If the words are there, any extractor can quote them; if they only materialize after scripts run, you are betting your visibility on rendering behavior no assistant documents. The same check, run on the JSON-LD block, shows you exactly what the markup contains that the visible page does not — and per Google's guidelines, the correct answer is nothing.

Should you remove your JSON-LD?

No — run both layers, priced honestly. We ship JSON-LD on all 400+ pages across our 3 production builds and will keep doing so for the 2 documented jobs, rich results and entity disambiguation [our data]. It costs almost nothing once templated, which is exactly the level of investment the evidence supports.

Against our own interest as people who sell builds with schema included: if your stack makes structured data expensive — plugin sprawl, manual per-page markup, an agency line item — skip the markup before you skip the writing. A page with clean visible answers and no JSON-LD can still be quoted everywhere that matters at answer time; a page with immaculate JSON-LD and mushy visible text has optimized the layer nothing quotes. The FAQ-specific version of this trade gets its own honest accounting in is FAQ schema still worth adding, and the passage-writing discipline that actually moves extraction is core to our generative engine optimization guide.