Does Schema Markup Help in AI Search? What the Evidence Says

By Muntazir · Published 27 September 2026

One tag told the more interesting story. A homepage belonging to a small UK heating and plumbing firm carries the script element a schema plugin uses to hand structured facts to a search engine — present, correctly typed, completely empty. Somewhere between installing the plugin and publishing the site, nobody filled it in. To a visitor, the page reads like any other local trade site. To anything parsing source code instead of rendered text, there’s no business there at all.

That gap between what a site’s markup claims to offer and what a machine can actually read runs through most attempts to answer whether schema helps in AI search. A check of nine similar small business homepages — drawn from two independently published “best local trade website” roundups, read on 27 September 2026 — found that empty tag on one of them. The other eight already carried some form of Organization, LocalBusiness or industry-specific markup. Absence wasn’t the pattern this sample turned up. Completeness was.

What does “schema markup” actually give an AI system?

Schema markup, in practice, means a block of JSON-LD sitting in a page’s source: a set of key-value facts — a business’s name, its type, its address, links to its social profiles — written in a format a machine can parse directly instead of inferring from paragraphs. An Organization or LocalBusiness node is the version aimed at answering “who runs this site.” It’s the same information a human reader already gets from the header and footer, offered a second time in a form nothing has to guess at.

The theory for why this should matter to an AI answer is straightforward: a system generating a response from retrieved text has to work out who said something and how much to trust it, and a clean structured-data block removes one layer of that inference. Whether the theory holds up against what these systems actually do is a separate question.

How complete is the schema that’s already out there?

Of the eight small business sites carrying an entity node, none had every property the format supports — name, url, logo, description, sameAs, address, telephone. Links tying the node to a social or directory profile, the sameAs property, were present on three of the eight. The other five had a name and a URL and stopped there.

The choice of type mattered more than expected. Three of the eight sites declared plain Organization rather than LocalBusiness or a more specific subtype like Plumber. None of those three included an address or phone number in the schema — not because the business hid that information, but because Organization has no field built for it, and nobody moved to a type that does. The same phone number sat two inches away in the page footer, readable to a visitor and invisible to whatever reads the markup instead.

Does adding schema actually increase how often an AI engine cites a page?

This is the part the theory doesn’t survive intact. In May 2026, Ahrefs published a study that tracked 1,885 pages which added JSON-LD schema between August 2025 and March 2026, against 4,000 matched pages that made no such change, and measured citation rates on Google AI Overviews, Google AI Mode and ChatGPT. The result: AI Overview citations fell 4.6%, AI Mode citations rose 2.4%, and ChatGPT citations rose 2.2%. The researchers described the two increases as statistically indistinguishable from zero, and the decline as unexplained rather than caused by the markup itself. Adding schema, on its own, produced no measurable citation gain on any of the three systems.

That’s a specific, falsifiable finding, and it complicates the case for schema more than most advice on the topic admits. A correlation still shows up elsewhere: other 2026 tracking has found a large majority of pages already cited by ChatGPT and Google AI Mode happen to carry structured data. Read next to Ahrefs’ controlled result, the two facts point the same way — schema tends to travel with sites that were already going to be cited for other reasons, rather than causing the citation itself.

So why does Google keep telling site owners structured data matters?

Because it still does, just not as an AI-specific requirement. Google’s own developer documentation is direct about this: “there’s also no special schema.org structured data that you need to add” to appear in AI Overviews or AI Mode. The guidance that follows is the same guidance that predates generative search — keep content crawlable, keep it well-organized, and, notably, make sure any structured data matches the visible text on the page. That last condition is exactly what an empty JSON-LD tag fails, whatever intention put the tag there in the first place.

What does that leave a business with?

A machine-readable statement of facts a page already makes to a human — useful for disambiguation, for a knowledge graph trying to work out who’s who, and for the quieter parts of technical SEO Google has recommended for a decade. Not a citation lever. Not something that moves the needle on its own, according to the one controlled study that tried to isolate its effect.

The identity side of that — one consistent name and set of facts across a site’s title tag, its markup and its social profiles — is a distinct problem from AI citations, and worth solving on its own terms; entity SEO covers the disambiguation case directly. A schema checker confirms a block parses and is present in the first place, which is the more basic failure this small sample turned up — an empty tag doesn’t fail more elegantly than a missing one, and the wider mechanics of how a model reads a page go well past what any one markup block can fix. Getting the block right is worth doing. It isn’t the lever most people writing about it are reaching for.