What is being sold
Add structured data and AI engines will start citing you. It appears in GEO packages, in agency proposals, and in a great deal of content marketing. It is intuitive, it sounds technical, and it is easy to bill for, because the deliverable is visible and countable.
The evidence does not support it. Three separate lines point the same way, and the most useful one comes from a company that sells the tool.
The controlled test
Ahrefs ran 1,885 treatment pages against 4,000 controls between August 2025 and March 2026. Adding schema moved AI citations by +1.2% on ChatGPT and −4.6% on Google AI Overviews.
Those numbers point in opposite directions and are small in both. That is what noise looks like. It is not a small positive effect, it is an absence of effect.
Ahrefs sells SEO tooling, including structured data features. A null result here runs against their own upsell, which makes it more credible rather than less. Note the direction of the commercial interest whenever you read a study: most published GEO statistics come from vendors whose conclusion happens to match what they sell.
Google, in plain language
Google's AI optimization guidance is unusually direct on this. Two quotes, verbatim:
You don't need to create new machine readable files, AI text files, markup, or Markdown
And:
you don't need to write in a specific way just for generative AI search
Source: Google Search Central, AI optimization guide.
Google has an obvious interest in discouraging attempts to game its systems, so this is not disinterested either. But it is consistent with the controlled test, and it removes the excuse that nobody was told.
llms.txt at 300,000 domains
SE Ranking analysed roughly 300,000 domains in November 2025. An XGBoost feature-importance test found that removing llms.txt as a variable improved model accuracy. No relationship to citation frequency. Reported via Search Engine Journal.
SE Ranking sells AI-tracking tools, so the same caveat applies, in the same direction: the finding works against the product.
Same pattern, different file. A machine-readable artifact was proposed, adopted enthusiastically, sold as a lever, and does not appear to move the thing it was sold to move.
The newest spec does not reference it either
Google launched the Universal Commerce Protocol for agent-driven commerce. If on-page markup were the substrate AI systems use to understand products, UCP is where you would expect to see it.
It is not there. We checked the specification directly on 30 August 2026.
UCP's documentation has a page titled "Schema Authoring", which is the sort of title that gets cited as proof of a connection. It is about something else entirely. It covers UCP's internal JSON Schema for describing capabilities, a validation vocabulary. The only schema-related URL on the page is json-schema.org/draft/2020-12/schema.
A vocabulary for validating the shape of JSON data. What UCP uses to define capability contracts between agents and merchants.
A vocabulary for describing things on web pages so search engines can produce rich results. What people mean when they say "schema markup".
Two different specifications with confusingly similar names. Searching the page for schema.org structured data, JSON-LD, rich results, or microdata returns nothing at all. The word "product" appears once, as an ordinary noun in a sentence about search responses, not as a markup type.
Third-party content describing a "structured data plus WebMCP plus UCP stack" is marketing, not specification. The stack it describes is not in the documentation. Anyone can verify that in about a minute, which is worth doing before paying for work premised on it.
Two findings converging from different directions, one measured and one documentary, both saying machine-readable markup is not the lever it is sold as for AI systems.
Where schema genuinely matters
Schema is not useless. It is misfiled. This distinction is the entire point of this piece, and skipping it would make the argument false.
Structured data is a rich results and Merchant Center deliverable, and in those roles it is not optional. For e-commerce it is close to mandatory. Merchant listing eligibility requires name, image, offers.price and offers.priceCurrency, per Google's merchant listing documentation.
Schema does real work:
- Rich results in Google Search, including review stars, prices, availability and FAQs.
- Merchant Center and Shopping eligibility, where missing required properties means exclusion.
- Entity clarity, by making business facts explicit and consistent rather than implied by page copy.
- Breadcrumbs, site links and other search appearance features.
None of that is small. Losing Shopping eligibility over a missing price field is a commercial problem with a direct revenue line attached. The correction here is one of category, not of value.
Schema belongs in the SEO column. Selling it from the GEO column is where the misrepresentation happens.
What we had filed wrong
We delivered 16 validated JSON-LD blocks as part of a client engagement, filed under the AI-visibility heading. The work was correct. The heading was wrong.
The blocks were accurate, they validated, and they will do useful work in search appearance. They were not an AI visibility deliverable, and presenting them under that heading implied a mechanism the evidence does not support. That work should have been filed under SEO, and it is now.
If you are buying GEO
Jeremy Moser, CEO of uSERP, on the record:
80 percent of GEO is good, fundamental SEO. If a GEO service does not openly tell you that, they are selling you snake oil.
That matches what we see. It is also worth knowing that publishers who top AI citation counts reportedly receive under 1% of their traffic from AI platforms. Being cited and being sent customers are not the same outcome, and most GEO pricing quietly assumes they are.
Three questions worth asking any GEO proposal:
- What is the mechanism? Not the deliverable, the mechanism. Why would this specific change alter whether an engine retrieves and names you?
- What is the evidence, and who produced it? If the only support is a vendor study whose conclusion matches what the vendor sells, treat it as marketing until something independent agrees.
- What is being measured, and against what baseline? If nobody recorded a dated before state, no after state can be interpreted.
GEO is not a separate premium service. It is SEO done by someone who understands how retrieval has changed. The technical work that helps AI systems is mostly the same technical work that helps search engines: pages that can be reached, facts that are consistent, claims backed by evidence, and independent sources that corroborate you.
One controlled test is not settled science, and AI retrieval changes. The Ahrefs study ran from August 2025 to March 2026, and the llms.txt analysis was November 2025. If someone produces a controlled test showing structured data moves citations, that would be worth reading and we would update this page. Nothing above rests on a claim we cannot point at a source for.