Structured data can tell a search engine what a business is; it cannot rank it or make an AI recommend it

A vocabulary adopted by hundreds of thousands of sites for a feature that no longer exists

When the Web Data Commons project first counted it in 2018, the FAQPage type from the schema.org vocabulary appeared on 19 domains in the whole of the Common Crawl; by the 2024 crawl it had been adopted by 308,000 ^[14]. One incentive drove that growth. Google would show the questions and answers as an expandable rich result beneath the page's listing, and the plugins that add structured data to small business websites made the markup a single click.

The incentive has now gone. In August 2023 Google limited the FAQ rich result to what it called well-known, authoritative government and health websites ^[6]. On 8 May 2026 it added a deprecation notice saying the feature would no longer appear in Search from 7 May 2026, and on 15 June 2026 it removed the documentation altogether, on the grounds that the feature is no longer shown ^[5]. The markup on all those domains still validates. It simply no longer does the thing it was installed to do.

That is the condition of structured data for a local business in 2026. It is oversold, because an industry promises that a few lines of JSON-LD will improve rankings or persuade an AI assistant to recommend you. It is underused, because the parts Google does document, a correct statement of who the business is, where it is and what it offers, are often missing, out of date or contradicted by a second plugin on the same page.

I wrote the structured data for my own clinic's website myself in 2026, with no agency, so what follows is partly the documentation I had to read and partly the claims I had to check against it. Nothing here relies on a rich result my own site has or has not received, because I have not measured that.

What Google says structured data does

Google's definition is narrow. Structured data is a standardised format for providing information about a page and classifying its content, which Google uses to understand the page and to gather information about the web and the world in general ^[1]. Google recommends JSON-LD, a block of script in the page's HTML, as the format easiest to implement and maintain at scale ^[1]. Three concrete things follow, and the documentation describes each.

• Eligibility for specific search features. Adding structured data can enable search results that Google calls rich results, which display more than the standard blue link ^[1]. Each feature has its own page in the documentation with required and recommended properties.
• Disambiguation of the business as an entity. The Organization type exists so that Google can better understand a company's administrative details and tell it apart from others in search results, and it can influence what appears in a knowledge panel, including the logo ^[4]. Google recommends placing it on the home page or a single page that describes the company ^[4].
• Machine-readable facts. LocalBusiness markup can help pages appear in a distinct kind of result, including a prominent knowledge panel with details about a business, and the documented properties are mundane: name, address, telephone, opening hours, coordinates, price range, URL ^[3].

Note the verbs. Enable, understand, influence, help. None is a promise, and Google's general guidelines say so.

Using structured data enables a feature to be present, it does not guarantee.

Type and property names are written here as the vocabulary spells them, in US English, because a JSON-LD block with a UK spelling is not a valid type.

What Google says it is not

The claim that structured data improves rankings is the oldest in the field, and Google's staff have denied it for as long as it has been made. The most direct recent statement came from John Mueller of Google in April 2025, who wrote that structured data will not make a site rank better and is used for displaying the search features listed in the documentation ^[11].

The general guidelines are consistent with that. They describe a manual action for spammy structured markup whose effect is to remove eligibility for rich results, and note that this does not affect the page's ordinary ranking ^[2]. Google treats the two as separate systems.

There is a reasonable counter-argument. A rich result can take more space and attract more clicks than a plain listing, and a business understood as a distinct entity may be matched to more of the queries it deserves. Neither is a ranking change, but both can change traffic. The honest position is that structured data can alter what a listing looks like and how confidently a search engine identifies the business behind it, and that any effect on the order of results is undocumented and, on Google's account, absent.

What has been withdrawn, and what the pattern tells you

• FAQ. Restricted in August 2023 to well-known, authoritative government and health websites ^[6], then withdrawn entirely from 7 May 2026, with the documentation removed in June ^[5].
• HowTo. Restricted to desktop in August 2023 ^[6], then dropped from desktop on 13 September 2023, at which point Google described the result type as deprecated ^[7].
• Sitelinks search box. Google announced in October 2024 that it would stop showing the feature from 21 November 2024 because usage had dropped ^[18].
• Five more types in one day. Documentation for course info, estimated salary, learning video, special announcement and vehicle listing was removed on 9 September 2025, and practice problems followed on 6 January 2026 ^[5].

Two things are consistent. Google has said each time that unused markup does no harm; in 2023 it stated that structured data not being used does not cause problems for Search ^[6]. And the withdrawn features were almost all presentational. The types that describe what a business is, Organization, LocalBusiness, BreadcrumbList and WebSite, remain, and when the search box went Google said that site names use a variation of WebSite structured data that continues to be supported ^[18].

The inference, and it is an inference rather than a Google statement, is that markup describing stable facts about a business has outlasted markup designed to win a particular display.

Self-serving reviews and the invented star

The withdrawal that matters most to local businesses is not a feature but an eligibility rule. Google's review snippet documentation says that ratings must be sourced directly from users, that sites must not rely on human editors to create, curate or compile ratings for local businesses, and that where the entity being reviewed controls the reviews about itself, the page is ineligible for the star feature, whether the reviews arrive through direct markup or an embedded third-party widget ^[8].

The LocalBusiness documentation says the same from the other direction: aggregateRating and review are recommended only for sites that capture reviews about other local businesses ^[3]. A clinic, a solicitor or a plumber marking up its own Google rating on its own home page is doing something Google's documentation tells it not to do.

Yet this is one of the most common things plugins and agencies add, because a row of stars is the outcome clients most easily recognise. Where the rating is invented, or assembled by the business from testimonials it selected, the general guidelines say reviews or ratings not by actual users may result in manual action ^[2]. Where it is genuine but self-hosted, the markup is ignored for the star feature. Either way it is not doing what it was sold to do.

What the two testing tools each check

Two free tools are routinely confused, and the confusion produces false confidence.

The Schema Markup Validator at validator.schema.org is the descendant of Google's old Structured Data Testing Tool, provided by Google as a service for the schema.org community ^[13]. It extracts JSON-LD, RDFa and Microdata, shows the graph it found and reports syntax errors; it is focused on schema.org and will not interpret other vocabularies ^[13]. It answers whether the markup is valid schema.org. It does not answer whether Google will do anything with it.

The Rich Results Test answers that second question, and only for the specific list of rich result types Google Search uses: it shows which of those types were found on the page, with errors and suggestions ^[12]. A valid block of Organization or Service markup can produce a result saying no items detected, because there is no rich result for those types. That is not an error. It is the tool saying that what you marked up is not a display feature.

The sequence follows. Validate the graph with the schema.org validator. Check the display-eligible types with the Rich Results Test. Then watch Search Console's rich result status reports, which show valid and invalid items for supported types, an invalid item being one with at least one critical issue preventing it from appearing as a rich result ^[20]. They are samples rather than an inventory ^[20].

The evidence that schema helps AI systems

The newest sales claim is that structured data makes a business visible to AI assistants and answer engines. The sellers deserve a fair hearing, because there is a real argument on their side, and then they should be held to the evidence, most of which is inference.

The case for

Microsoft has said something specific. Fabrice Canel of Bing told an SMX Munich audience in March 2025 that schema markup helps Microsoft's language models understand content, and later confirmed the remark in writing ^[15]. The logic is also sound: a system handed a machine-readable statement of a business's name, address, services and people has less to guess at than one reading marketing copy.

The case against

Google's documentation is unusually plain. Its page on AI features says there is no special schema.org structured data you need to add, and no new machine-readable files, AI text files or markup are needed to appear in AI Overviews or AI Mode ^[9]. Its guide to generative AI features, updated in July 2026, repeats that structured data is not required and there is no special schema.org markup to add, while calling it still useful for rich results ^[10]. From Google's perspective, the guide says, this work is still SEO ^[10].

Then there is the question of whether the models see the markup at all. In May 2026 the SEO practitioner Mark Williams-Cook published a test in which he created a fictitious company with deliberately nonsensical schema, using made-up properties, and asked ChatGPT and Perplexity about it ^[16]. Both extracted the address anyway. His reading was that the systems were not parsing it as schema but reading the visible text, and he noted that common pipelines for preparing training data discard script tags, where JSON-LD lives ^[16]. He still recommended implementing schema because the cost is low; his point was that a fact appearing in an answer and also in the markup is not evidence the markup was used, since the fact appears on the page as well ^[16].

My own testing points the same way. I have asked ChatGPT, Gemini and Perplexity about clinics and treatments in Bradford, including my own, and when the answers were wrong I traced them back to out-of-date or inconsistent third-party pages, not to anything missing from my own markup. The free Live AI Visibility Test on HighRegard runs real prompts and records which sources were cited; those sources are pages, directories and review profiles, which is where the evidence about a business lives.

So the state of the evidence is this. Microsoft says markup helps its models ^[15]. Google says none is required and none is special ^{[9] [10]}. One independent test suggests general-purpose assistants read the visible page rather than the markup ^[16]. The rest is inference from the design of the vocabulary. Inference is not worthless, but a business paying for schema as an AI visibility service should know that is what it is buying.

What a local business should actually mark up

The types worth a local business's attention describe stable facts, are documented by Google, and have survived every withdrawal above. That is five things, in order.

1. A single Organization or LocalBusiness node. Choose the most specific type that is true. Schema.org defines LocalBusiness as a particular physical business or branch, with subtypes including Dentist, MedicalBusiness, LegalService and HomeAndConstructionBusiness ^[21]. Give it the properties Google lists: name, address, telephone, URL, opening hours, coordinates and, where accurate, a price range ^[3]. Give it a stable @id so every other page can point at the same node, and sameAs links only to profiles that genuinely belong to the business ^[4].
2. Person, for named professionals. A clinic, practice or firm whose credibility rests on named people should say who they are, with job title, qualifications where they are visible on the page, and a link to the professional profile. It is the same principle of disambiguation that Google describes for the company ^[4], applied to its people.
3. Service, for each thing you actually offer. There is no Google rich result for it, so the Rich Results Test will report nothing, and that is expected. Its value is that it links a named service to its provider and area served in a form a machine can read, on a page whose visible text says the same.
4. WebSite, on the home page only. This is how Google recommends you state the site name shown in results, with the name and canonical home page URL ^[19], and it is the part of WebSite markup Google said it would continue to support when it dropped the search box ^[18].
5. BreadcrumbList, on every page below the home page. Google uses breadcrumb markup to categorise the page's information in search results and requires at least two items, each with a name and position, and a URL for all but the last ^[17]. It is the one display feature on this list, and a modest one.

FAQPage, HowTo, aggregateRating on your own reviews and types added because a plugin offered them are not on the list.

The common ways it goes wrong

The failures fall into four groups.

Marking up what is not on the page. Google's guidelines say not to mark up content that is not visible to readers of the page, and to provide up-to-date information ^[2]. Markup listing services the business no longer offers, an address it has left, or a practitioner who has moved on is worse than none, because it asserts a falsehood in the one place machines are told to trust, and the manual action for structured data issues covers exactly this ^[2].

Duplicate emitters. An SEO plugin, a theme, a page builder and a hand-added block can each output their own graph, leaving two or three statements of the business's name and address on one page that disagree on the telephone number or the type. The schema.org validator shows every one of them ^[13]. There should be one.

Invented or self-served ratings. Covered above: ineligible if self-controlled ^[8], manual action if fabricated ^[2].

Default output left as installed. A generic LocalBusiness type where a specific one exists, no people, no services, no identifiers connecting pages. It is valid, and it says almost nothing.

What this means for a local business

Structured data is a one-off piece of technical work with light maintenance, not a monthly line on a retainer. Google's documented LocalBusiness properties are few ^[3] and the Organization properties are all optional ^[4]. Once the graph is right, the job is to change it when the facts change and to check the Search Console reports when the site is rebuilt or a plugin updated ^[20].

Its place in the order of work is after crawlability and a well-maintained Google Business Profile, alongside making the business's name, address and telephone number agree everywhere they appear. A JSON-LD block stating one phone number while three directories state another has added a fourth opinion, not settled the argument.

The adoption figures show why the work is still worth doing without a display feature to win. In the October 2024 Common Crawl, 44.1 per cent of the 37.4 million domains examined carried some structured data, JSON-LD was found on 11.6 million domains against 7.6 million for Microdata, and LocalBusiness appeared on 1.26 million domains ^[22]. A local business stating its facts in the vocabulary the rest of the web uses is not chasing an advantage. It is meeting a baseline, and the Web Data Commons project reports a continuous increase in deployment ^[14].

Anyone selling schema as a subscription, a ranking lever or a route to being recommended by an AI assistant should be asked which Google document supports the claim. For rankings, Google's staff say the opposite ^[11]. For Google's AI features, no special markup is needed ^[9]. For other assistants, there is a Microsoft statement about its own models ^[15] and otherwise inference.

What the evidence does not settle

Several questions remain open.

Whether structured data influences which sources Google's AI Overviews and AI Mode draw on is not addressed by Google's documentation. Google says none is required ^[9]; it does not say it is ignored. Structured data feeds the same understanding of pages that the rest of Search uses ^[1], so an indirect effect is plausible. Nobody outside Google has measured it in a way that separates the markup from the visible content it duplicates.

Whether ChatGPT, Gemini, Perplexity and their successors will start to parse markup as markup is unknown. Williams-Cook's test describes their behaviour in mid-2026 ^[16]; it does not bind their designers, and Microsoft's statement shows one search-backed system already claims to use it ^[15].

What is settled is narrower than the sales claims. Structured data lets a business state its facts in a form machines can read. Google will use those facts to understand and disambiguate the business and, for a short and shrinking list of types, to change how a listing looks. It will not use them to rank the business higher, and it does not require them for its AI features. Everything beyond that is a platform's word about its own models or an inference from how the vocabulary was designed. A business should do the documented part carefully, keep it true, and treat the rest as a hypothesis until someone publishes a measurement.