Generative engine optimization · Schema markup
Does schema markup get you cited by AI? We tested 26 med spas.
We wanted to find out whether schema markup actually gets you cited by AI. It doesn't. Here's what turned up instead.
Every agency pitching AI optimization sells the same three-step: throw on some schema markup, upload an llms.txt file, and watch ChatGPT start recommending you. We got tired of debating whether that's true, so we went and looked.
We took 26 med spas in Scottsdale, checked what structured data each one carries, and then put twelve real customer questions to ChatGPT to see who gets named. Fourteen of the 26 came up at least once.
What the numbers say
The result did not go the way the industry promises. Businesses with LocalBusiness markup were named 69% of the time against 38% without it, which looks like a win until you test it: at this sample size that gap sits well inside chance. Medical markup, Service markup, FAQ markup, any JSON-LD at all, every one of them came back with no detectable link to being named.
The clearest evidence sits in the contradictions. One clinic carries MedicalOrganization, Physician and Service markup together, holds 349 reviews, and appeared in none of the twelve answers. The only business in the pool with FAQ markup also scored zero. Meanwhile two spas carrying no LocalBusiness markup whatsoever were each named in four answers out of twelve.
One thing did come back significant: review count. Businesses above the median were named 77% of the time against 31% below it, and that difference is unlikely to be chance. It also explains the schema numbers, because the spas with markup happen to be the ones with more reviews, and once you account for that the schema gap nearly disappears.
What did matter
So the honest reading is that schema markup was not what got these businesses into the answer. Their reputation was.
What this does and doesn't prove
Let's be straight about what this is. Eleven clinics isn't a controlled experiment. We didn't hold every variable constant, and AI answers drift between runs, so this is a snapshot, not a law. But the pattern's clear enough that "add more schema" stops being a credible explanation for who gets cited. If markup were the lever, that clinic with twenty-plus schema types and a working llms.txt should have been unavoidable. It wasn't there.
So what put those five clinics in the answer, if not their code? We didn't find a single site among the sources ChatGPT actually cited. Every source was a directory or roundup, the review aggregators and "best of" lists you see when you search "med spa Miami" yourself. Two of those directories showed up again when we ran the identical query for San Diego, same week, opposite coast. That's not coincidence. That's the model returning to the same trusted shortlist, regardless of what any individual clinic's homepage claims.
Where we'd actually start
Here's what we'd actually prioritize, in order, based on what we saw:
- Find which directories are already cited for your city and category. We did this by running the query ourselves and reading the actual sources the model lists, rather than guessing from "what everyone recommends." The two directories that repeated across our two-city test weren't obvious beforehand.
- Get listed in those specific directories before touching your website's code. That's what actually correlated with citation in our test. Markup didn't.
- Treat schema markup as basic hygiene, not a growth lever. Do it once, keep it accurate, stop there. It can hurt you if it's wrong or missing, but piling on more isn't what moves you from invisible to cited.
- Check where you stand before spending a dollar. We built a free check that runs your business name and city against ChatGPT and Perplexity and shows you the actual sources behind the answer, the same thing we did by hand for this test, in about thirty seconds.
We're not selling schema packages, so we've got no incentive to claim it works when it didn't. What we found: the clinics getting named earned it somewhere off their own site, and the ones with the cleanest on-page code sat this one out completely.