Does Schema Get You Into AI Search Results? An Honest Look

We've had the same conversation three times in the last month. A potential client gets in touch, and somewhere in the first ten minutes, they ask whether we can "do their schema", so they'll start appearing in ChatGPT and Google's AI answers. They've read a LinkedIn post, or sat through a webinar, or been quoted a tidy sum by someone promising that structured data is the secret tunnel into AI search results.

And honestly, our hearts sink a little. Not because schema is a bad idea (it isn't, and we'll get to that), but because we've watched this exact film before. For twenty years, we've been untangling clients from SEO myths that got sold to them as silver bullets, and the schema-for-AI pitch has the same smell. A real thing, with real but modest value, being repackaged as a magic lever and priced accordingly.

So when an SEO called Mark Williams-Cook published an experiment that pokes a large hole in the hype, we wanted to walk you through what he found, because it's genuinely useful and it'll save you some money. To be clear upfront: the research and the experiment we're about to describe are his, not ours. We're just translating it into plain English and adding our two penn'orth from the agency trenches. You can (and should) read his full article on Search Engine Journal.

First, what schema actually is

If the word "schema" means nothing to you, you're in good company, and you don't need to feel bad about it. Schema (its full name is Schema.org structured data) is a bit of invisible code you add to a web page to spell things out for machines. It sits in the background, where visitors never see it, and its whole job is to remove guesswork.

Here's the idea. Natural language is messy. The word "Apple" could mean a fruit, the tech company, or a record label. When a person reads your page, they work it out from context, but a computer has to guess. Schema is how you stop it from guessing. You use the code to say, in effect, "this string of numbers is a phone number, this is our address, this is the name of our company, and here's our Wikipedia page so you know exactly which company we mean." It's a polite little contract between your website and a machine, agreeing on what each thing on the page means.

That's been useful for years, mostly for Google. It's part of how you end up with those rich results in search (star ratings, FAQs, event dates, that sort of thing), and it feeds Google's Knowledge Graph, the big database of facts that powers the info panels you see on the right of a search. Williams-Cook makes the point that this is what schema was built for, and it does that job perfectly well. The argument is about something else entirely: whether schema is your way into AI chatbots like ChatGPT and Perplexity. If you want the bigger picture of how AI search differs from the search you're used to, our plain-English guide to GEO is a good companion read to this one.

The duck experiment, and why it matters

This is the bit that made us sit up. Williams-Cook wanted to test a claim that's everywhere right now, which goes: "an AI returned a fact that was only in my schema, therefore the AI is reading and using my schema." Sounds reasonable. It's also, as he demonstrates rather beautifully, not what's happening.

So he built a deliberately daft test page for a made-up T-shirt company called DUCK YEA. On the visible page, he mentioned no address at all. But hidden in the code, he buried a chunk of schema that was complete nonsense, on purpose. Instead of using the proper, recognised labels, he invented his own. The company "type" was listed as a "MallardEnterprise". The address fields had names like "reedNumber", "puddle" and "featherCode". As far as the actual rules of schema are concerned, this was gibberish. Any system genuinely reading it as schema should have taken one look and binned it.

Then he asked ChatGPT and Perplexity for the company's address. Both cheerfully handed it over: 77, The Muddy Bank, South Pondshire, and so on. Perplexity even claimed it had found the answer "in the page's embedded structured data", with the satisfied air of a student who'd done the reading.

Here's the punchline. The AI tools didn't reject the made-up nonsense because they were never really reading it as schema in the first place. To them, that hidden code was just more text on the page, a bit oddly punctuated. They scanned the page, spotted what looked like an address, and read it aloud. As Williams-Cook puts it, if he'd wrapped the address in flashing text and surrounded it with duck emojis, it would have made no difference. The whole point of schema, telling the machine precisely what each thing is, was simply ignored.

That's why the "the AI returned my schema fact, so it's using my schema" argument falls apart. The fact was also sitting there in plain text for the AI to grab. The schema wrapper did nothing. Williams-Cook is careful (and we'll be too) not to overclaim here: his test doesn't prove AI tools ignore schema in every situation forever. It proves that this particular, very common bit of "proof" people wave around isn't proof of anything.

The two theories about how AI "uses" schema

When people argue that schema feeds the AI, Williams-Cook points out they're usually leaning on one of two ideas. Both are worth understanding because once you do, the sales pitch becomes much easier to see through.

Theory one: schema gets "baked in" when the AI is trained

The thinking goes that when companies like OpenAI build their models, they hoover up the whole web, your lovely schema included, and it ends up stored inside the AI somewhere. The problem, as Williams-Cook explains, is that the process of building these models starts by cleaning the raw web pages, and that this cleaning step deliberately strips out the code, the menus, the cookie banners, and, yes, the schema. The goal is to keep the readable prose and throw away the clutter. Schema lives in exactly the kind of code tag that gets thrown away.

And even if some of it slipped through, there's a second problem. These models don't store facts like a filing cabinet with a folder marked "your company" and your address tucked inside. They store patterns, a blurry statistical sense of which words tend to follow which other words, learned from reading enormous amounts of text. Williams-Cook's comparison is someone trying to recall the lyrics of a song they last heard in 2011. For your specific address to "stick", the model would need to have seen it many, many times across the web. For nearly every business, that simply doesn't happen. As he memorably puts it, you'd be paying a consultant to whisper your postcode into a hurricane.

Theory two: schema gets read live, the moment you ask a question

The other idea is that schema is read in real time, when you ask the AI something and it nips off to fetch your page. This is closer to how some AI search actually works, but Williams-Cook walks through why the strong version of the claim still doesn't hold up. One version says your schema feeds Google's Knowledge Graph, which is true, but that's a slow, carefully built database, not something assembled on the fly in the two seconds between you hitting enter and getting an answer.

Another version leans on a Microsoft quote that supposedly "confirmed" schema feeds Copilot. When you read the actual quote, Williams-Cook shows it was about something completely different (telling search engines when your content has changed so they come and look again, which is about freshness, not schema). And the third version is the duck experiment we just covered. So when you stack them up, the live-reading argument is thinner than it first appears.

We're not flagging all this to be gloomy about AI search, quite the opposite, we think it matters enormously, and we've written plenty about getting ready for it, including our practical guide to getting your business AI-ready. The point is narrower: schema specifically isn't the magic key it's being sold as.

Even Google can't get this right yet

This was the example that really landed for us, because it's so concrete. If anyone on earth were going to wire structured business data neatly into AI answers, it'd be Google. They've got the Knowledge Graph, they've got Google Business Profiles (a tidy, structured database of business details), they own the AI (Gemini), they own the AI Overviews that appear at the top of search, and they own the search index underneath all of it. Every advantage you could ask for, under one roof.

And yet Williams-Cook shares a screenshot of a single Google results page where the AI Overview at the top confidently states that a particular car dealership is open, lists its address and even its opening hours, while the Google Business Profile panel on the very same page, for the very same business, carries a big red "Permanently closed" banner. Two parts of the same company, on the same screen, in the same moment, flatly contradict each other. The AI answer wasn't even checking Google's own authoritative, structured record.

His conclusion, and ours, is hard to argue with. If the company with every possible advantage can't reliably plug its own structured business data into its own AI answers, the idea that a consultant can do it for your website by adding some schema is, to put it gently, optimistic.

Why this feels like SEO all over again

Here's where we'll put our own oar in, because this pattern is painfully familiar. Back when Google could be gamed, the SEO world was awash with people selling "one weird trick" to get you to the top. Keyword stuffing, dodgy link schemes, meta tag voodoo. Most of it was either useless or actively harmful, and we spent years cleaning up after it and gently explaining to clients why the magic beans hadn't grown a beanstalk. We've written before about why SEO is so misunderstood, and the short version is that complicated, slow-moving things attract simple, confident, wrong explanations.

The schema-for-AI pitch is cut from the same cloth. There's a genuine shift happening (more people are getting answers from AI instead of a list of links), everyone's anxious about being left behind, and into that anxiety steps a confident voice with a neat, sellable solution. The trouble is that "add schema, appear in ChatGPT" is a claim resting, as Williams-Cook says, on a remarkably thin pile of evidence. A fact appearing in an AI answer that also happens to be in your schema is not proof that the schema did the work, because that fact is almost always sitting in your visible text too.

So our advice is the same as it's always been when a new shortcut appears: ask whoever's selling it to show their working. If the proof is "the AI repeated something that was in the schema", that's the duck experiment, and it doesn't survive contact with a nonsense test. If they've got something better, brilliant, we'd love to see it. We treat the new "GEO best practice" with the same healthy scepticism we learned to apply to SEO orthodoxy, and you should too.

So should you bother with schema at all?

Yes. And we want to be really clear about this, because it would be easy to read all of the above and conclude that schema is pointless. It isn't. Williams-Cook recommends still using it, and so do we. The trick is to use it for the right reasons and not pay silly money for the wrong ones.

Schema is cheap to add; the downside is virtually nil, and the benefits compound over time. It still helps with ordinary Google results. And if it does turn out to matter more for AI down the line (it might), then the work's already done and you can feel smug about it. What it won't do, on the current evidence, is single-handedly drag your brand into ChatGPT's answers because you sprinkled some code on a page.

There's one situation where Williams-Cook reckons schema genuinely earns its keep for AI, and we agree it's the interesting case. It's not for big, well-known brands that the AI can already identify with confidence (adding schema there is solving a problem you don't have, a bit like introducing yourself by name to your own mother). It's for newer or smaller businesses, or companies whose name clashes with others, where the machines aren't yet sure who you are. For them, schema is groundwork. It helps you become a recognisable, resolvable thing in the first place, which earns you the right to be a candidate for those AI answers later, even if it doesn't buy you a mention today.

Where we'd steer your money instead, in the meantime, is the unglamorous stuff that actually moves the needle: genuinely useful content that answers the questions your buyers ask, clearly structured pages, and the slow accumulation of credibility across the web. That's the same advice we give for traditional search, which is no coincidence, because good SEO-friendly content remains the foundation on which everything else, AI included, sits. If you're weighing up where to spend, our take on whether a marketing agency is worth it might help you sense-check any pitch you're being given.

If someone's quoted you a big number to "do your schema for AI", we're more than happy to have a quick, no-strings chat and tell you straight whether it's worth it for your particular situation. Sometimes the honest answer is "yes, but it should cost a tenth of that", and sometimes it's "save your money and spend it on content". Either way, we'd rather talk you out of something you don't need than sell you a shortcut that doesn't exist. Drop us a line if you fancy a chat, and bring the quote with you.