Structured Data for AI Search: What Still Matters

What Structured Data Actually Does for AI Assistants

Every AI assistant answering a question like “best dentist near me” or “where can I buy this part” first sends out a crawler—a program that visits web pages and copies their content—to read your site. Without labels, that crawler sees a wall of text and has to infer meaning: is “555-0134” a phone number or a product code? Structured data removes the guesswork by wrapping facts in a small code block called JSON-LD, which sits quietly on the page and says, in effect, this is the phone number, this is the price, this is the opening time.

OpenAI’s own documentation confirms its crawler, GPTBot, fetches publicly available web pages much the way a traditional search engine does OpenAI, and Google’s structured data guidance is explicit that these labels help a program “understand the content of the page” Google Search Central. Neither company has said it ignores this labeling, so in practice, the same JSON-LD block you would add for classic search still gets read by the assistants your customers are switching to. If you’ve never added it, the assistants answering questions about your business today are working from guesses, not facts.

The Four Schema Types Worth Your Time

Schema.org lists hundreds of item types, from Recipe to Movie to Airport, but a typical small or mid-sized business only needs four of them to cover almost every fact an assistant might quote. Add these, and don’t worry about the rest.

• Organization — establishes your business as a single named entity, with a logo and links to your official social profiles, so an assistant can tell that the Instagram account and the website both belong to the same “Dolores’ Bakery.”
• LocalBusiness — an extension of Organization built for anyone with a physical address or service area; it carries your hours, phone number and neighborhood, the exact details an assistant needs to answer “are they open right now.”
• Product and Offer — attach a real price, currency and availability to what you sell, so an assistant quotes your actual number instead of averaging a few competitors’ guesses.
• FAQPage — a set of question-and-answer pairs in the exact shape assistants already like to lift and quote back to a user, word for word.

What Machine-Readable Confirmation Does for Entity Resolution

Picture Dolores, who owns a bakery in Austin. She has a Google Business Profile, a Yelp page, an Instagram account and a website—four separate mentions of “Dolores’ Bakery” scattered across the internet. Before an assistant can safely recommend her, it has to solve entity resolution: deciding whether all four mentions describe the same real business, or whether one of them is a different Dolores’ Bakery in a different city.

Structured data gives the assistant a shortcut. A schema.org property called sameAs lets Dolores list her Instagram, Facebook and Yelp URLs directly inside her own website’s Organization markup—a machine-readable confirmation, in her own words, that these accounts belong to her schema.org. That’s a stronger signal than an assistant inferring a match from similar names and addresses across sources it doesn’t fully trust; it’s Dolores herself vouching for the connection, in a format nothing can misread.

This matters more as competition for a name grows. A single “Dolores’ Bakery” in a town of eight thousand people resolves itself easily. A “Riverside Dental” competing against four clinics with similar names needs every disambiguating detail—the exact address, the exact phone number, the confirmed social links—to keep an assistant from blending it with a rival down the street.

A Copyable JSON-LD Example You Can Adapt This Week

You don’t need a developer to write this from scratch; you need to fill in a template. A JSON-LD block for a local business sits inside a single script tag, marked as type application/ld+json, and starts with a line declaring its vocabulary as schema.org’s context. Inside it, the type is set to LocalBusiness, followed by a name field with your exact business name and an image field pointing at your logo.

Next comes an address, written as its own nested object of type PostalAddress, broken into streetAddress, addressLocality (your city), addressRegion (your state or province) and postalCode—four separate labeled fields rather than one line of text, because that’s what lets an assistant answer “is this near me” without misreading a ZIP code as a suite number. A telephone field and a priceRange field (something as simple as “$” for moderate pricing) follow. An openingHoursSpecification field lists your days and hours in a structured format rather than a sentence, so “closed Sundays” can’t be misread as “closed Mondays.” Finally, a sameAs field holds an array of URLs: your Instagram, your Facebook page, your Yelp listing, and any other profile that is unmistakably you.

Once written, that block gets pasted once, near the top of your homepage’s code, and left alone. If you use a website builder like Squarespace, Wix or Shopify, most now offer a settings field for custom code or structured data where this can be added without touching the rest of your site.

What You Can Skip

Not every schema.org type is worth your time, and chasing all of them wastes effort better spent on the four above.

• BreadcrumbList markup, which describes your site’s folder structure (Home > Menu > Cakes), helps classic search results show a navigation trail but tells an assistant nothing about your business’s facts—skip it unless you already have it from a past project.
• AggregateRating without real underlying reviews. Google’s own structured data guidelines explicitly prohibit markup that misrepresents ratings you don’t actually have Google Search Central, and an assistant repeating an inflated score you can’t back up is a reputation risk, not a win.
• Deeply nested or duplicated markup on every single page. Your Organization identity only needs to be declared clearly once, usually on the homepage; repeating it inconsistently across dozens of pages creates conflicting versions for an assistant to reconcile, which is worse than declaring it once, correctly.
• Brand-new, rarely supported schema.org types. The vocabulary itself lists hundreds of types built for narrow uses like recipes or airports schema.org; if a type isn’t among the handful search engines and assistants are documented to actually use, the label does nothing for you today.

How to Check Your Structured Data Is Actually Working

Adding the code is only half the job; confirming it’s readable is the other half. Google’s free Rich Results Test lets you paste a page URL or the code itself and see exactly which types it detects and which fields are missing Google. Running it after every change catches a typo—a missing comma can silently break an entire block—before it costs you anything.

That test tells you the code is valid. It doesn’t tell you whether ChatGPT actually used your confirmed hours in an answer, or whether Perplexity still lists a competitor first. That’s a separate question, and it’s the one Growli was built to answer: we track what assistants actually say about a business across platforms and flag where the gap between correct code and a correct answer still exists. See how it works if you want to check your own results rather than assume the markup is doing its job.

FAQ