A maintenance team adds FAQ, HowTo and Organization markup to every entry in a directory. The JSON-LD validates, the rich-result tester raises no objection, and a quarter later the assistants are still naming a competitor when the question touches a category the directory knows well. Nothing is broken in the code. The invoice was written in the right format and handed to the right office, and nobody agreed to pay it. That gap, between a page a machine can parse and a page a machine chooses to quote, is where most of the confusion about schema markup lives.
Structured data does a specific job. It turns a block of prose into typed entities: this is an Organization, that is a HowTo with ordered steps, this FAQ block is a set of question-and-answer pairs rather than a rhetorical heading. JSON-LD became the usual way to express those types because it sits alongside the page without rewiring the HTML. For a directory, the payoff is concrete. An entry with a stable identifier, an explicit type, and a declared relationship to its category can be lifted out of the page and reused in an answer without a crawler having to guess what the label means. Markup is the invoice machines scan before deciding whether the rest of the page is worth parsing at all.
Adoption is already broad enough that skipping it is a choice with consequences. The Web Data Commons corpus, rebuilt from the October 2024 Common Crawl and reported by the University of Mannheim and Web Almanac 2025, found structured data in 51.25% of 2.4 billion HTML pages. JSON-LD is the format of choice for roughly 70% of sites that do mark up, used by about 11.5 million sites. What was a specialist optimization a few years ago is close to table stakes now. A directory that ships no structured description is not merely less tidy; it is asking a parser to work harder than the majority of competing pages require.
| Signal | Reading | Origin and year | Consequence for a directory |
|---|---|---|---|
| Pages with structured data | 51.25% of 2.4B HTML pages | Web Data Commons / Web Almanac, from Oct 2024 crawl | Markup is a baseline, not an edge |
| JSON-LD adoption | ~70% of marked-up sites, ~11.5M sites | University of Mannheim / Web Almanac 2025 | One serialization, applied consistently |
| GEO visibility ceiling | up to +40%; expert quotes +41%, statistics +33%, fluency +29%, cited sources +28%, keyword stuffing −8% | Princeton GEO study, KDD 2024 | Content strategies, not schema fields |
| Google public line | no extra technical requirement; no special schema.org data needed | Google Search Central, updated 10 Dec 2025 | Schema is not an entry gate |
| Madrid counterweight | structured data remains essential in an AI world | Google Search Central Live Madrid, 2025 | Legibility still decides what gets used |
| Controlled JSON-LD test | cited pages ~3x likelier to carry JSON-LD; test group AIO −4.6%, AI Mode +2.4%, ChatGPT +2.2% | Ahrefs, Aug 2025 – Mar 2026 | Correlation is not a lift curve |
The Princeton GEO study presented at KDD 2024 is often cited as proof that structure wins. It found that generative engine optimization could raise visibility in generated answers by up to 40%, with expert quotes associated with a 41% gain, statistics 33%, fluency 29%, and cited sources 28%. The same study associated keyword stuffing with an 8% decline. Those numbers describe content strategies observed in one specific experimental setup, not a bonus attached to a JSON-LD block, and they should be read that way. A directory can mark up every field correctly and still publish prose that gives a model nothing worth repeating. Schema makes a claim legible; it does not make the claim persuasive.
Google's own messaging adds a genuine tension worth stating carefully. In guidance updated on 10 December 2025, the company said AI features impose no additional technical requirement on sites and that no special schema.org markup needs to be added for them. Months earlier, at Search Central Live Madrid in 2025, Google speakers argued that structured data remains essential in an AI world. Both statements hold at once. The first says there is no hidden schema that unlocks AI visibility; the second says a page that cannot be typed and parsed is a poor candidate for being assembled into an answer. One is about entry requirements, the other about legibility. Reading them as a contradiction misses the more useful lesson: markup is necessary plumbing that almost nobody gets credited for, and plumbing is not the same thing as a reason to quote.
The number that should temper any schema celebration comes from a controlled test by Ahrefs. Between August 2025 and March 2026, the team added JSON-LD to 1,885 pages and held 4,000 pages as a control. Pages that already drew AI citations were almost three times as likely to contain JSON-LD, a strong association and one worth noticing. But the group that actually received the markup moved in uncoordinated directions: Google AI Overviews −4.6%, AI Mode +2.4%, ChatGPT +2.2%. A cited page tends to carry an invoice; handing an invoice to a page does not make it cited. Correlation and causation are not the same measurement, and the distance between them is the honest boundary of what schema can promise. The study does not prove markup is useless. It proves that a strong association is a bad forecast when the intervention itself is the variable being tested.
What survives that boundary is a narrower, more boring discipline. Keep identifiers stable when a display name is translated, declare the entity type instead of leaning on category context, mirror the human description in the machine-facing one, and treat FAQ and HowTo blocks as a real answer rather than a keyword container. None of that forces a mention. All of it lowers the cost for any crawler or assistant that decides the directory is worth reading carefully. For an open catalog, the useful move is architectural rather than editorial: emit the structured record as a first-class output of the same data model that renders the browsable page, and hold the identifiers constant across every surface that displays it. A shared runtime, the kind behind Cross Mini App, turns that into a single decision instead of a per-page chore. Machines decide which invoices to honor. The least a directory can do is write one they can read.