Schema and Structured Data for AI Search: The Practical Guide

Schema is the most oversold idea in AI search. Spend ten minutes on LinkedIn and you will be told that a few tags will get you quoted by ChatGPT, pinned to the top of an AI Overview, or waved past the queue into Google’s knowledge graph. It is a tidy story. It is also wrong, and Google has said as much in writing.

This is the technical piece in our answer engine optimisation series, so we will be straight with you. Structured data matters, but not for the reasons the hype merchants give. Below: what schema actually is, what it honestly does for AI retrieval, the types worth your time, and how to implement and test it without losing a week to markup that changes nothing.

What is structured data, and what does schema actually do?

Structured data is a small block of code that spells out the meaning of a page in a format machines read without guessing. A person reads “The Second Floor, Leicester” and knows it is a company in a city. A crawler sees a string of characters. Schema removes the guesswork by labelling it: this is an Organization, this is its name, this is its address, this is the addressLocality.

“Schema” and “structured data” get used interchangeably, and that is close enough. Schema.org is the shared vocabulary, a joint project backed by Google, Microsoft, Yahoo and Yandex, that defines the types and properties everyone agrees on. Structured data is the practice of putting that vocabulary on your pages. The format you should use is JSON-LD: a self-contained script in the page’s code that describes the content without touching the visible layout.

Think of it as labelling the boxes in a warehouse. The stock is the same either way. Labelled boxes just get found, sorted and trusted faster, by search engines and by the language models now reading those same pages.

Does Google need special schema to show you in AI search?

No. Google is unusually direct about this, so we will quote it. In its documentation on AI features, Google states there is “no special schema.org structured data that you need to add” to appear in AI Overviews or AI Mode (Google Search Central). The bar for eligibility is the ordinary one: a page must be indexed and eligible to show in Google Search with a snippet. That is it.

Google’s dedicated guide to generative AI features, published in 2026, makes the same point a different way. There is no AI-only markup, no secret file, no tag that flips a switch. The things that make you eligible for AI answers are the things that make you eligible for normal search: useful content, crawlable pages, clear structure.

If you want the cautionary tale, look at FAQ markup. For years, agencies sold FAQPage schema as a growth lever. Then Google restricted FAQ rich results to a handful of health and government sites in 2023, and in May 2026 switched them off for everyone (Search Engine Journal). Anyone who had sold FAQ schema as the product, rather than as clean labelling, was left holding nothing. That is the risk of treating markup as a trick. Tricks expire.

If schema is not required, why bother at all?

Because clarity compounds. Structured data cannot force a machine to cite you, but it removes the ambiguity that makes machines hesitate. Three things happen when your markup is clean and honest.

• You remove guesswork. When a page explicitly says it is a Product with a price and a review, or an Article by a named author on a stated datePublished, a crawler does not have to infer it from layout and hope. Less inference means fewer wrong conclusions about what your page is.
• You feed entity understanding. Search engines and AI systems model the world as entities (people, places, companies, products) and the relationships between them. Consistent markup helps them connect your pages to the right entity, which is the groundwork for being recognised as a credible source. This is the same machinery behind ranking in AI Overviews: the system has to understand you before it will quote you.
• You stay eligible for rich results. Plenty of markup still earns visible search features: review stars, product pricing, breadcrumbs, event details. Those features win clicks in ordinary search, which is not going anywhere just because AI answers sit alongside it.

None of that is a guarantee, and we will not pretend otherwise. Schema is hygiene and clarity. It is the difference between handing someone a labelled file and handing them a shoebox of loose paper. Both contain the same facts. One of them gets read correctly.

Which schema types actually matter, and when?

For most UK businesses, fewer than ten types cover almost everything, and half of those you will never touch. Ignore the catalogue of hundreds of obscure types. Here is the shortlist that earns its place, and when to use each.

• Organization: your top-level identity, set on the home page or across the site. Name, logo, URL, and the sameAs links that tie you to your real profiles. Start here.
• LocalBusiness: a more specific Organization for anyone with a physical location or a defined service area. Adds address, opening hours, geo and phone. Essential for local and map visibility; Google’s Organization guidance lists the fields that matter.
• Article: for blog posts and guides, with a named author, a datePublished and a dateModified. Author and dates are the parts worth getting right, because they support the credibility signals both search and AI systems weigh.
• Service: for each distinct thing you sell. Useful on service and money pages where you want the offering understood precisely.
• Product: for e-commerce, or anything with a price, availability and reviews. This is where rich results still pay off directly.
• BreadcrumbList: describes where a page sits in your site. Small effort, tidy result, and it reinforces structure for crawlers.
• Person: for the humans behind the work: authors, founders, consultants. Ties a name to a real, describable entity rather than a floating byline.
• FAQPage: still valid, but read the room. The rich result is gone as of May 2026, so add it only where you have a genuine, visible FAQ. Google will still parse it to understand the page; you just will not get a search feature for it.

The pattern: describe who you are (Organization or LocalBusiness), what you publish (Article, Person), and what you sell (Service, Product), then stop. Depth of accurate markup on a few types beats a scattergun of half-filled ones.

A worked example makes it concrete. A Leicester plumber needs LocalBusiness with address, hours and service area, Service for each job type, and Organization-level sameAs links to their Google Business Profile and a review platform. A software company with no premises needs Organization rather than LocalBusiness, Article on every guide, Person for the authors, and Product or SoftwareApplication for the tool itself. Same vocabulary, different subset. Pick the subset that matches what you actually are, and leave the rest.

How do you tell AI systems who you are?

With the sameAs property, which is the most underused line in most companies’ schema. Entity confusion is a genuine problem: there are other firms with your name, other people with your founder’s name, and a machine that cannot tell you apart will hedge or pick someone else. sameAs is how you disambiguate. It links your Organization to the authoritative profiles that describe the same entity.

Picture an assistant asked to recommend a Leicester agency. If three firms share a similar name and none of them state, in machine-readable terms, which social profiles and directory listings are theirs, the assistant either guesses or leaves all of them out. A complete sameAs block is you raising your hand and saying, unambiguously, this one is us.

Point sameAs at things that genuinely confirm your identity:

• Your verified social and professional profiles (LinkedIn, and the platforms you actually maintain).
• Your entry on reputable directories and review platforms. The UK directories that influence AI recommendations are worth being listed in for exactly this reason.
• Knowledge bases such as Wikidata, and Wikipedia if you genuinely warrant an article. Wikidata in particular is machine-readable by design and feeds knowledge graphs directly.

The goal is a consistent identity across the open web. When your name, address and profiles line up everywhere a machine looks, it stops hedging and starts treating you as a known quantity. That consistency, not a clever tag, is what underpins being recommended by ChatGPT and the assistants like it. Schema states the identity; the wider web has to corroborate it.

How do you implement JSON-LD, step by step?

Put one JSON-LD script in the page’s HTML, describe only what is on that page, and keep it accurate. JSON-LD is Google’s recommended format because it sits apart from your content and is simple to maintain. A minimal Organization block looks like this:

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Your Company Ltd",
  "url": "https://www.yourcompany.co.uk",
  "logo": "https://www.yourcompany.co.uk/logo.png",
  "sameAs": [
    "https://www.linkedin.com/company/your-company",
    "https://uk.trustpilot.com/review/yourcompany.co.uk",
    "https://www.wikidata.org/wiki/Q00000000"
  ]
}

To roll it out sensibly:

1. Map types to templates, not to individual pages. Decide that all blog posts get Article, all service pages get Service, the home page gets Organization or LocalBusiness, then implement once per template.
2. Pull values from real page data. The name, datePublished and price in your markup should come from the same source as the visible content, so they cannot drift apart.
3. Add the script to the rendered HTML. The <head> or the <body> is fine; Google reads both. If your site renders client-side, confirm the markup appears in the served HTML, not only after JavaScript runs.
4. Use one block per entity. You can include several types in a page’s markup, but keep each accurate and non-duplicative. More is not better.

Most content management systems and SEO plugins generate solid JSON-LD for the common types, and that is fine to start with. Just check what they output rather than trusting it blindly, which brings us to testing.

How do you test structured data?

With two free tools, used together, before and after anything goes live.

1. Google’s Rich Results Test tells you which Google rich results a page is eligible for, and flags Google-specific errors and warnings. Use it to answer “will Google do anything visible with this?”
2. The Schema Markup Validator checks your JSON-LD against the Schema.org vocabulary itself, with no Google-specific filtering. Use it to answer “is this valid, well-formed markup at all?” Google retired its old Structured Data Testing Tool and handed this one to Schema.org, so it is now the neutral reference.

Run the validator first to catch syntax and vocabulary errors, then the Rich Results Test for eligibility. After launch, the structured data reports in Google Search Console show issues at scale across your indexed pages, which is where you catch problems the one-page tests miss. Fix every error, and treat the warnings as a to-do list rather than noise.

Testing proves the markup is valid; it does not prove it is doing anything. For that, watch outcomes. Search Console’s rich results reports tell you what Google has parsed and where it choked, and its Generative AI performance report shows how your pages surface in AI features on Search. Neither report credits a citation to a specific tag, so resist drawing a straight line from one FAQPage block to one AI mention. Use them to spot regressions and trends, not to award schema medals it never earned.

What are the most common schema mistakes?

Nearly all of them come down to schema saying something the page does not. Search engines treat that as a trust problem, and the consequences run from ignored markup to manual action. The usual offenders:

• Marking up content that is not on the page. If your FAQPage schema lists questions a visitor cannot see, or your review markup invents ratings, that breaks Google’s guidelines and can get your structured data ignored across the whole site. Only mark up what is genuinely visible.
• Schema that contradicts the page. A price of £99 in the markup and £129 on the page. An author in the JSON-LD who is credited nowhere on screen. Machines notice the mismatch, and mismatch reads as manipulation.
• Over-marking. Tagging every element with every type you can find does not add authority. It adds clutter and error surface. Mark the things that matter and leave the rest alone.
• Letting it go stale. Markup rots. Prices change, authors leave, opening hours shift, and a dateModified sits frozen on a page you rewrote last month. Stale schema is worse than none, because it actively misinforms. Review it when you review the content.
• Wrong or invalid types. Using a type that does not fit, or inventing properties Schema.org does not define. This is exactly what the validator catches, so there is no excuse for shipping it.

The honest test for any block of markup is simple: does this describe the page accurately, right now, to someone who cannot see it? If yes, keep it. If not, fix it or bin it.

Where schema fits in the bigger technical picture

Schema is one layer of a technical foundation, and it cannot carry a site on its own. Markup helps a machine understand a page it can already reach, render and trust. If your pages are not crawlable, your rendering hides content from bots, or your site is slow and duplicative, no amount of JSON-LD rescues you. That is why we treat structured data as one line item in a full technical SEO audit rather than a standalone fix.

The sequence that actually works is unglamorous:

1. Get the fundamentals right: crawlability, indexing, rendering, speed, clean information architecture.
2. Publish genuinely useful, clearly structured content, the sort that answers a real question. That is the substance AI systems quote, and the foundations of GEO start there, not with markup.
3. Then add accurate schema to label what you have built, so machines interpret it correctly.

Do it in that order and structured data amplifies a strong site. Do it out of order and you are labelling boxes in a warehouse nobody can get into.

If you would rather have someone check whether your schema, and the technical foundation under it, is actually working for AI search, that is what our free growth review is for. We will tell you what is helping, what is stale, and what is snake oil, in plain terms. Our engagements are scoped monthly with weekly updates and no lock-in, so you are never paying for markup that does nothing.

Schema will not get you cited by an AI on its own. Nothing will. But clean, honest, accurate structured data makes you easier to understand, and easier to understand is the closest thing to a durable advantage that AI search currently offers.