Ask ChatGPT or Claude to write you some JSON-LD schema and it’ll hand you a clean-looking code block in seconds. Paste it in, and nine times out of ten it’ll be fine. It’s the tenth time that gets people — the block that validates, sits on the page for six months, and quietly does nothing, or worse, gets your site flagged in Search Console.
We use AI for schema generation constantly now. It’s genuinely one of the better uses of the technology in SEO work. But there’s a right way and a wrong way to do it, and the difference isn’t obvious until something breaks.
What schema is actually for
Schema markup (usually written as JSON-LD, sitting quietly in a page’s <head>) tells search engines what’s on a page in a format they can parse without guessing. This business is a dentist. This page sells a product. It costs £45. It has 212 reviews averaging 4.8 stars.
Some of that gets rewarded with visible perks in search — star ratings under a listing, breadcrumb trails, product pricing shown before someone even clicks. Some of it just helps machines understand your site without any visible reward at all. Worth knowing that distinction before you build a strategy around chasing a specific rich result, because Google has been quietly retiring several of these in the last couple of years. FAQ rich results, for instance, are gone entirely as of May 2026 — Google pulled the visible dropdown feature after years of sites stuffing FAQ schema onto pages that weren’t really FAQs. The markup still validates fine. It just doesn’t do anything visible anymore. If your agency (or your AI prompt) is still pitching FAQ schema as a way to win extra SERP real estate, that pitch is out of date.
Product, Review, LocalBusiness, Article and BreadcrumbList schema are still very much worth doing properly. That’s where the effort should go.
Where AI genuinely helps
It’s fast, and it knows the shape of most common schema types reasonably well — Organization, LocalBusiness, Product, Article, BreadcrumbList. If you’re staring down fifty product pages that all need the same schema type with different values swapped in, that’s a task AI will chew through far quicker than doing it by hand.
It’s also decent at explaining what a property actually does, which helps if you’re not a developer and want to understand what you’re pasting into your site rather than just copying it blind.
Where it quietly goes wrong
Three patterns show up again and again when we review AI-generated schema for clients.
The first is invented or outdated properties. Schema.org’s vocabulary changes, and AI models will sometimes include something that sounds plausible but isn’t actually part of the current spec, or was deprecated. It looks fine to the eye. It fails silently when Google actually parses it.
The second is bad nesting. A Product needs an AggregateRating, which needs a ratingValue and reviewCount nested correctly inside it — get the structure slightly wrong and you end up with JSON that’s syntactically valid but means something different to a parser than you intended.
The third, and the one we see most often in practice, is schema that stops matching the page. Someone generates a clean block, ships it, and six months later the price changes, or a product goes out of stock, or the review count updates — and the schema doesn’t, because nobody wired it to update automatically. Mismatched structured data (a price or rating in the markup that isn’t visible on the actual page) is specifically the kind of thing Google’s guidelines call out as manipulative, even when nobody meant it that way.
How we’d actually use it
Treat AI output as a first draft you check, not a finished file you paste and forget.
Give it real specifics — the actual product name, price, and review numbers straight off the page, not a vague “write me Product schema.” Ask for one schema type at a time rather than bundling several together, since that’s when nesting mistakes creep in. Then validate it: run it through the Schema Markup Validator at validator.schema.org (Google’s Rich Results Test works too, for the types it still supports) before it goes anywhere near a live page. If it throws an error, feed that exact error back to the AI rather than guessing at a fix yourself — usually faster and more accurate than manual troubleshooting.
Then check it against the actual page. Anything the schema claims that a visitor can’t verify by reading the page should come out.
The actual point
AI hasn’t removed the need to understand schema — it’s made the drafting part faster, which is a real time saver across dozens of pages. The validation step is still on you. Every site we’ve audited with schema issues had markup that looked right and wasn’t quite — usually because nobody went back and checked it once the page content changed. That’s the gap worth closing, AI-assisted or not.
If you’re not sure what state your site’s structured data is actually in, that’s a quick thing for us to check — book a free audit and we’ll show you.
