A Study Tracked 1,885 Pages That Added Schema. It Didn’t Buy Them AI Citations.

If you have sat through a GEO deck in the last year, you have seen the slide. Add JSON-LD, mark up your organization entity, describe your products and your FAQ, and the AI engines will file you properly and start citing you. It sounds like engineering, so it feels more reliable than link building, and it is dramatically easier to sell internally than “write to 40 directories.”

Ahrefs tested it. The result is worth reading in full, but here is the headline: adding schema produced no major uplift in citations on any platform.

What the study actually measured

Ahrefs tracked 1,885 web pages that added JSON-LD schema between August 2025 and March 2026, matched them against 4,000 control pages, and measured what happened to their citations across Google AI Overviews, AI Mode, and ChatGPT.

No platform showed a meaningful gain. AI Overviews was the one place where the difference ran the other way: a 4.6% decline, small in absolute terms but statistically significant against the matched controls. On AI Mode and ChatGPT, treated pages performed slightly better than control pages on average, but by margins that could easily be noise across thousands of URLs.

That is a negative result, and negative results are the most useful kind of research in a field this full of confident claims.

The second finding is the interesting one

Within the same week, Mark Williams-Cook published a test that explains why. He built a fake company — DUCKYEA t-shirts — and put its address nowhere on the page except inside invented, invalid JSON-LD. Then he asked ChatGPT and Perplexity where the company was located.

Both assistants found the address.

The schema was not valid. It did not correspond to any real entity. It did not use an approved vocabulary correctly. And both models read it and used it, because the words were present in the HTML and looked relevant to the question.

His conclusion is the cleanest sentence written about this all year: the model is picking up whatever you list in the HTML. It does not matter whether it is valid schema. And the practical version of that: schema is good, repackaging the basics as a magical new GEO formula is bad.

What this implies about how assistants choose sources

Put the two findings side by side and a coherent picture emerges.

Structured data is a machine-readable assertion made by the page about itself. Assistants appear to treat it as text: they read it, they may use it, they do not privilege it. A page asserting “we are an organization with these attributes” is one page making a claim. It is not corroboration, because there is nothing to corroborate against — the claim originates from the party being described.

What the models appear to need, for anything beyond reading comprehension, is the same thing search engines have needed for twenty-five years: independent descriptions of the same entity, from sources that do not share an owner. When an assistant decides whether a product is a real player in a category, “the product’s own site says so” is weak evidence and “thirty other sites describe it the same way” is strong evidence.

This is also why the schema study’s null result is not a scandal. Schema does something useful — it disambiguates your entity for crawlers and it can earn rich results in traditional search. It just does not manufacture the corroboration that recommendation decisions are made from.

What to do instead

The replacement for schema-led GEO is unglamorous and measurable.

Describe yourself in one consistent sentence, everywhere. Pick the description, keep it stable across your homepage, your about page, your LinkedIn profile, your industry directories, and your review profiles. Variance in your own description is the most common self-inflicted wound we see — five different taglines across five listings means five weak claims instead of one reinforced one.

Count your corroborating properties, not your backlinks. For AI visibility specifically, the useful number is how many independent pages describe what you do and who you serve. Fifty links from blog comments are worth less here than fifteen real listings, because the listings are what a model reads when it is deciding whether you are a plausible answer.

Get the basics right once, then move on. Organization, Product, FAQPage where it genuinely applies, correct canonical, clean headings. That is a two-hour job. Everything past that is not what is holding you back.

Stop optimizing for the answer box and start optimizing for the sources. You cannot influence the synthesis step. You can influence whether, when the synthesis step looks for evidence about your category, your name appears in more than one place with a description that matches.

A concrete weekly test

Ask ChatGPT, Perplexity, and Google AI Mode a question a real buyer would ask about your category — not your brand name, the category. Then look at which sources got cited. Not the answers, the sources.

Run it weekly for a month and tabulate two things: how often a domain you appear on is cited, and how often your brand appears in an answer without your domain being cited. The first number tells you whether your citation work is reaching the models. The second tells you whether you have become the assumed answer — which is what a genuinely corroborated entity looks like.

Both numbers are actionable. Neither one moves because you added JSON-LD.

Where this leaves directory work

The honest framing is that the schema study did not discover a new strategy. It removed a distraction.

For the last year, a lot of teams spent their AI-visibility budget on markup they could implement themselves, because it felt controllable. In the same period, the work that actually shapes AI answers — being present and consistently described across the places that define your category — went under-funded, because it is tedious, it involves filling in forms, and nobody gets promoted for it.

The study does not say “citations don’t matter.” It says the cheap technical substitute does not work. Which leaves the original work, done manually and reasonably well, as the thing that does.

Which makes the placement work the substitute that actually behaves like an input: https://backlinksubmit.org matches your product to relevant directories and communities from a handpicked list, submits manually with written copy, and shows in a console which platforms have been submitted and how each is progressing.

FAQ

Should I remove my schema markup?

No. It still helps traditional search understand your entity, and it costs nothing once implemented. Just stop expecting it to move AI citations.

Does this apply to product feeds?

Partly different. OpenAI has said it sources shopping data from structured feeds, and Microsoft has documented how schema helps Copilot. So schema matters more in commerce contexts than in general recommendation contexts. Test it in your own category rather than generalizing from either result.

What’s the single best proxy for AI visibility?

How often your brand is mentioned in third-party sources that describe your category. That is the input, and unlike referral data, it is visible to you without platform cooperation.

Is this an argument for buying directory listings?

It’s an argument for being present in enough independent places that your description repeats without you being the only author of it. Whether you pay a service or do it yourself is a separate question about your time.