The Real Job of Schema Markup (And Why You Still Need It)
If you’ve talked to an SEO or GEO consultant in the last year, you’ve probably heard some version of “add schema markup” as step one for AI visibility. It comes up constantly, and almost always with the same quiet promise attached: mark your pages up correctly, and ChatGPT, Perplexity, and Google’s AI Overviews will start citing you.
I like that idea a lot. I just don’t think it’s true, and I went and checked.
The evidence is thinner than the advice suggests, and honestly, once you actually look at it, the real story is more interesting than the hype was. So let’s start from the beginning.
Schema Markup, Quickly
Structured data is a bit of code you add to a page, usually as a block called JSON-LD, that spells out what the page is about in a way machines can parse cleanly: this is an Organization, its name is X, its founder is Y. It sits quietly in the background, separate from anything a visitor actually sees.
Google recommends JSON-LD as the preferred format, and they’re refreshingly honest about what it does and doesn’t do: it makes a page eligible for certain search features and helps their systems understand context. That’s it. It’s an invitation, not a guarantee.
The Five Types That Actually Matter for a Brand
Most companies don’t need to touch more than five schema types. Here’s what each one is actually doing for you:
Organization carries your entity’s core facts ,name, logo, founding date, address. The property that matters most here is sameAs, which links your site’s version of “who we are” to your canonical profiles elsewhere: Wikipedia, LinkedIn, Crunchbase. It’s the thread that ties your own claims to independently verifiable ones, and it’s the exact mechanism behind something I wrote about a while back ,if you missed it, Your Brand Has a Wikipedia Problem You Don’t Know About covers why that matters for how AI models talk about you.
Person does the same thing for your people. jobTitle, worksFor, and a sameAs pointing to their own verified profiles is what lets a knowledge graph confidently say “Gadi Samet, COO of Buzz Dealer” instead of guessing which Gadi Samet you mean.
WebPage declares what a given page is ,its headline, its author, its publisher, when it was last reviewed. It’s not glamorous. It’s the difference between a system trusting your date or byline and a system guessing at it.
FAQPage pairs your actual questions with your actual answers in a structured block. It works well on pages that are genuinely FAQ-shaped. Where I see it misused most is when it gets bolted onto pages that aren’t FAQs at all, just to tick a box.
Get Gadi Samet’s stories in your inbox
Join Medium for free to get updates from this writer.
Service spells out what you sell ,the service type, the provider, where it’s available. It’s the type I see left half-finished more than any other, which is a shame, because it’s the most direct, machine-readable answer to “what does this business actually do.”
The Uncomfortable Data on Schema and AI Citations
Here’s where it gets genuinely interesting, and where I’d love to tell you I was right to be skeptical from day one ,I wasn’t, not really, until I saw the numbers.
Ahrefs ran the test most of us in this industry have only ever assumed the answer to. They tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, compared against 4,000 control pages that didn’t, and measured citation activity 30 days before and after, across three platforms. The result: nothing moved in any meaningful way. ChatGPT citations shifted +2.2%. Google AI Mode moved +2.4%. Both are statistically indistinguishable from doing nothing at all. Google AI Overviews actually dropped -4.6% ,small, but per their statistical testing, a real effect rather than noise.
Now, here’s the twist that explains why the “add schema” advice took hold in the first place. In a separate analysis of 6 million URLs, Ahrefs found that pages cited by AI were almost three times more likely to be running JSON-LD than pages that weren’t. That’s a real pattern. It’s just not the pattern people think it is. Ahrefs is upfront that this is correlation, not causation: the kind of site that bothers implementing clean schema is usually also the kind of site already winning on the things that genuinely earn citations ,strong content, real structure, third-party coverage. Schema is a symptom of a well-run site, not the cause of its visibility.
An experiment from Otterly.ai fills in the mechanical “why.” Testing five schema types over three months, they found that six out of seven AI search platforms couldn’t even retrieve schema data when asked to fetch it directly. And when a piece of information existed only inside FAQ schema, with nothing matching in the visible page text, not a single AI platform used it to answer a related question. Most of these systems are reading your page roughly the way a person would, by scanning the visible content, not parsing the JSON-LD tucked away in your code.
Building It Correctly (Because It’s Still Worth Doing)
None of this means skip schema. It means stop expecting it to do a job it was never actually doing. If you’re implementing it, a few things genuinely matter:
Use JSON-LD. It’s Google’s own recommended format, and it’s the easiest to keep clean without touching your visible page code.
Only mark up what’s true and actually visible on the page. Google’s guidelines are explicit that hidden or inaccurate markup breaks policy, and it can cost you rich-result eligibility altogether, not just fail to help.
Fill in the properties that carry real information, not just the bare minimum a validator will accept ,sameAs, founder, and foundingDate for Organization; jobTitle and sameAs for Person.
Run it through Google’s Rich Results Test before you ship it, and go back and keep sameAs links current as profiles change. A stale one is worse than none.
And mentally file it under ORM and entity clarity, not GEO citations. That’s the job it’s actually good at.
The Real Takeaway
Schema markup’s real job is telling machines, clearly and consistently, who you are. That’s genuinely valuable ,it’s core infrastructure for reputation management, and it feeds the knowledge graphs that sit upstream of a lot of what AI systems end up saying about you. What it isn’t, based on everything I could actually find, is a way to get ChatGPT to start quoting you.
If your entire GEO plan begins and ends with adding schema, you’re pulling the wrong lever. The things that actually move AI citations is content these systems can read directly, and real third-party coverage ,are a separate, harder project. Worth doing both, honestly. Just not for the same reason, and not with the same expectations.
