Schema and Entity SEO for AI Search

Five studies disagree on whether schema earns AI citations. This guide reconciles them and shows the exact JSON-LD to ship, with a UK entity layer.

TopicHow AI Reads the Web

Schema markup will not, on its own, get you cited by ChatGPT. The best 2025 and 2026 evidence shows structured data is read at crawl, index and training time by Google and by Microsoft's systems, then ignored during live retrieval, at which point assistants read only your visible HTML. So treat schema as entity infrastructure. Build one connected graph with an entity home, @id references, a Wikidata record and UK company identifiers, then measure citation changes against a control group. This guide shows the exact JSON-LD.

Nobody agrees on this topic. In May 2026, Ahrefs published a controlled study of 1,885 pages that added JSON-LD and found citation movements "statistically indistinguishable from zero". The same month, Analyzify aggregated five studies, several reporting schema-marked pages cited up to 2.3 times more often. Google's documentation says no special markup is needed at all. Microsoft says schema helps its LLMs understand your content. Agency blogs pick whichever position suits their pitch and present it as settled.

It is not settled, but the contradictions can be reconciled.

I am Ian Duncan, founder of AxiomAI. I have been in marketing for around thirty years, since HotBot was a serious search engine, and I have spoken on SEO at a Guardian industry event. I have watched markup hype cycles inflate and pop before. What follows explains why the studies disagree, says plainly where schema does nothing, and merges schema and entity SEO into a single implementation. It adds a UK layer the other guides skip, and a measurement protocol that would survive a sceptical finance director.

Does schema markup actually get you cited by AI assistants?

As a direct cause, mostly no. As a correlation, strongly yes. The gap between those two answers is behind almost every argument you will read on the subject.

Take the strongest data first. Between August 2025 and March 2026, Ahrefs followed 1,885 pages that added JSON-LD and compared them against roughly 4,000 matched control pages that never had it. Citations in AI Overviews fell 4.6%, AI Mode rose 2.4% and ChatGPT rose 2.2%, movements the authors describe as "statistically indistinguishable from zero". Adding markup moved nothing.

How strong is your own AI marketing? Take the free two minute scorecard and get an honest read plus the next steps that matter most.

Take the scorecard

Yet the very same dataset shows AI-cited pages were almost three times more likely to carry JSON-LD than pages that were never cited.

Both findings are true. Sites that implement structured data properly tend to be the sites with strong content and existing authority, so schema travels with citations without causing them. That confusion, correlation dressed up as causation, powers the "2.5x more likely" claims that circulate between agency guides with no traceable origin. When you cannot find the methodology, assume there is none.

The positive studies still deserve a fair hearing. An AirOps and Kevin Indig analysis of 16,851 ChatGPT queries across 353,799 pages found pages with JSON-LD cited at 38.5% against 32.0% without, with BreadcrumbList at 46.2%, FAQPage at 45.6% and Organization at 44.3% as the strongest performing types. The same aggregation reports Digital Applied finding schema-marked pages cited 2.3 times more often across 1,000 AI Overviews, rising to 2.8 with HowTo, and a UC Berkeley GEO-16 framework measuring a 39% lift from structured data. Those figures reach me second-hand through Analyzify rather than from the raw papers, so I hold the precise decimals loosely. You should too.

The Ahrefs null comes with a caveat the headline leaves out. The treated pages each already had 100 or more AI Overview citations before any markup went on. "If a page is already getting picked up, our data suggests that adding schema isn't going to push it higher", the study notes. The null result speaks to pages that were already winning. It says far less about an organisation that AI systems cannot yet identify at all. That is a different problem, and it is the one the rest of this guide is about.

Why do the five studies disagree?

Because schema is consumed in one pipeline and ignored in the other, and each study measured a different pipeline. Once you see the split, most of the disagreement makes sense.

Pipeline one runs at crawl, index and training time. Google's own generative AI optimisation guide confirms that its AI features are built on its core ranking systems, so anything that helps conventional indexing and rich results feeds through to AI Overviews indirectly. Microsoft has been more explicit. Speaking at SMX Munich, principal product manager Fabrice Canel said schema markup helps Microsoft's LLMs understand your content, the first plain confirmation from a major AI platform.

Pipeline two is live retrieval. Mid-conversation, an assistant fetches your page and reads the rendered, visible HTML. Not the JSON-LD. OtterlyAI demonstrated this with an elegant trap. The team planted unique facts only inside schema, invisible on the page, and over three months zero platforms ever surfaced the schema-only information. Asked to read raw markup directly, only Gemini managed it. ChatGPT, AI Overviews, AI Mode, Perplexity, Copilot and Claude all failed, and Google AI Mode went one worse by confidently describing a "Service" schema that was not on the page.

Now run each study back through that split. Otterly's own sitewide rollout saw Google AI Overview appearances rise 1,500% while ChatGPT visibility fell 71%, Copilot fell 64% and Perplexity did not move. That is exactly the pattern you would expect if consumption happens inside Google's indexing systems and nowhere near a live fetch. The correlational studies capture index-time and site-quality effects. The Ahrefs null shows the marginal causal effect is tiny once a page is already indexed, understood and heavily cited. Otterly's own conclusion is that schema's value "appears to flow primarily through traditional SEO performance, not through direct consumption by most AI Search platforms".

These studies all have weaknesses. Otterly tested one SaaS brand with no control group, and its headline percentages were reported without ruling out confounds. Ahrefs only studied pages that were already heavily cited. The aggregated positive studies never published their confound handling. The clean experiment, on companies AI systems have never heard of, has not been run. Anyone telling you this science is settled is selling something.

Where schema works and where it does nothing

At index time it works. During a live assistant fetch it does nothing, unless you are on Gemini and you ask it to read the markup. Nowhere is it required, and Google says so in writing. "There's also no special schema.org structured data that you need to add", states Google Search Central's AI features documentation, which also names the real controls for AI surfaces, the unglamorous nosnippet, data-nosnippet, max-snippet and noindex settings. The same page explains that AI Mode and AI Overviews may issue multiple related sub-searches, a technique Google calls query fan-out, to assemble one answer. A page can now be retrieved for a sub-question no human ever typed.

Here is the platform picture the evidence actually supports.

Platform Consumes your schema Evidence
Google AI Overviews and AI Mode Indirectly, at index time, never as a requirement Official guidance says no special markup is needed (Google Search Central). AI Overview appearances rose 1,500% after a sitewide rollout (OtterlyAI)
Gemini Yes, the only assistant that fetched raw JSON-LD on request Seven-platform retrieval test (OtterlyAI)
Bing and Copilot Yes, confirmed on the record for Microsoft's LLMs Fabrice Canel at SMX Munich (Search Engine Land)
ChatGPT No direct reading observed, correlation only 38.5% citation rate with JSON-LD versus 32.0% without (Analyzify). Visibility fell 71% in the rollout test (OtterlyAI)
Perplexity No evidence of consumption Failed the raw-schema fetch and showed no visibility change (OtterlyAI)
Claude No evidence of consumption Failed to retrieve schema when prompted (OtterlyAI)

So schema earns its keep at index time. It helps machines resolve which real-world organisation your pages describe. It keeps you eligible for the rich results and knowledge surfaces that Google's AI features draw on, and it is what Microsoft has confirmed its own systems read. Where it does nothing, on current evidence, is inside a live assistant fetch, as a hiding place for facts missing from your visible HTML, or as a rescue for content nobody wants to cite. Your visible words carry the weight, and quotable sentences with cited statistics beat markup tricks. There is more on that pattern in our breakdown of the Princeton GEO study.

What is the difference between schema markup and entity SEO?

Schema is syntax. Entity SEO is the strategy that syntax serves.

Schema markup labels your content in a vocabulary machines can parse. Entity SEO makes sure those machines resolve your organisation, people and services to one unambiguous node in their knowledge graphs rather than three half-matched guesses. Online, the two are usually written about as separate topics. Schema guides skip Wikidata, entity homes and @id stitching entirely, while entity guides describe JSON-LD in words and never show a line of code. That split is daft, because in practice you implement them as one system. That is the approach behind our guide to entity maps and the wider complete guide to generative engine optimisation.

One number does the rounds here. Brand mentions correlate with AI visibility at 0.664, against 0.218 for backlinks. Even Digital Applied, which repeats it, flags it as vendor research. The number has been relayed through several hands with no published methodology, so treat it as directional rather than proven. It does fit the mechanism, though. Systems trained on a corpus lean on how often and how consistently an entity is described across that corpus, and far less on your link profile alone.

How do you implement schema and entities as one system?

Anchor everything to stable @id identifiers, publish from a single entity home, then corroborate with external records.

Pick one entity home

Choose the single page that defines your organisation, normally the homepage or the about page. Every schema block, external profile and directory listing should agree that this URL, and no other, is where the entity lives.

Anchor the organisation with @id

{ "@context": "https://schema.org", "@type": "Organization", "@id": "https://axiomai.uk/#organization", "name": "AxiomAI", "url": "https://axiomai.uk/", "logo": "https://axiomai.uk/logo.png", "founder": { "@id": "https://axiomai.uk/#ian-duncan" }, "sameAs": [ "https://www.wikidata.org/wiki/Q_YOUR_ITEM_ID", "https://www.linkedin.com/company/your-company/", "https://find-and-update.company-information.service.gov.uk/company/01234567" ]
}

The @id is the load-bearing property. It is a stable URI that names the entity itself, distinct from any page URL, and every other block on your site can point at it instead of repeating the organisation's details. For sameAs, create a Wikidata item, which has a far lower bar than Wikipedia, and reference verifiable records such as your Companies House listing. Swap the placeholders for your own identifiers, obviously.

Stitch the graph across the site

{ "@context": "https://schema.org", "@graph": [ { "@type": "Person", "@id": "https://axiomai.uk/#ian-duncan", "name": "Ian Duncan", "jobTitle": "Founder", "worksFor": { "@id": "https://axiomai.uk/#organization" } }, { "@type": "Article", "@id": "https://axiomai.uk/resources/example-guide/#article", "headline": "Example guide", "inLanguage": "en-GB", "author": { "@id": "https://axiomai.uk/#ian-duncan" }, "publisher": { "@id": "https://axiomai.uk/#organization" } } ]
}

Because the Article, the Person and the Organization reference identical @id strings, a parser treats them as one connected graph rather than duplicate fragments. Organization to Person to Article to Service, each entity defined once and referenced everywhere else. This is the stitching that entity guides name but never show. It is also what stops a machine deciding your founder and your author are two different people.

Keep the vocabulary up to date as well. Schema.org version 30.0 shipped on 19 March 2026 with equivalence annotations and EU Digital Product Passport examples, following v29.4's ConferenceEvent type in December 2025. Centralise your JSON-LD in one template include so vocabulary updates propagate sitewide from a single edit.

This is the pattern I run on my own sites. I publish the entity graph, an llms.txt file and an EntityMap, and I adopted EntityMap v1.0 the day after it launched. Only the graph carries real evidence, though. The rest costs minutes and might matter later, so I ship it and say exactly that. If you would rather have the gaps found for you, my free entity audit checks these properties on any URL you give it.

The standard itself has serious backing. R.V. Guha, one of the founders of schema.org, reviewed EntityMap and endorsed it ahead of launch, per Search Engine Journal. For a standard this young, that counts for a lot.

What extra markup should a UK business ship?

Your registered identity, which none of the pages ranking for this topic bother to mention.

Every registered UK company already holds a unique public identifier at Companies House. Your VAT registration adds another, and some firms have a Legal Entity Identifier on top. Placing them in your Organization block hands machines exact keys they can match against public registries instead of guessing from a name that fifty other businesses share.

{ "@context": "https://schema.org", "@type": "Organization", "@id": "https://axiomai.uk/#organization", "identifier": { "@type": "PropertyValue", "propertyID": "Companies House", "value": "01234567" }, "vatID": "GB123456789", "leiCode": "213800EXAMPLELEI0000", "address": { "@type": "PostalAddress", "addressCountry": "GB" }
}

Because this block repeats the same @id, it merges with the Organization node you already published rather than creating a rival copy. Add inLanguage of en-GB to your WebSite and Article nodes so nothing silently defaults to US English, and if you have physical premises, give them a LocalBusiness type with a full PostalAddress and postcode. No study isolates the effect of registry identifiers on AI citations, so this layer is reasoning rather than proof. It costs an hour. It cannot contradict your visible content, because it comes straight from statutory records. And misidentification is one of the few AI failure modes you can directly prevent.

How do you measure whether schema changed your AI visibility?

With a control group. Without one you cannot tell whether the schema caused anything.

Agency guides quote four-to-twelve-week timelines with no method for separating schema's effect from everything else that moved in the same quarter. Content edits, model updates, seasonality, plain luck. That separation problem is precisely what the Ahrefs treated-versus-control design was built to solve, and it is worth copying at small scale.

  • Baseline first. Track citations for four weeks across ChatGPT, Perplexity, Copilot and AI Overviews with a fixed prompt panel before touching any markup.
  • Split comparable pages into a treated set and a control set, matched on traffic and existing citations.
  • Deploy schema to the treated set only, and change nothing else on either set while the test runs.
  • Push the updated URLs through IndexNow and resubmit your sitemaps so index-time systems see the change quickly.
  • Confirm recrawl in Search Console and Bing Webmaster Tools before starting the clock.
  • Compare the change in citation rate between the two sets, never the raw before-and-after on treated pages alone.
  • If both sets moved together, the cause was not your schema. Write that down and resist the urge to claim it anyway.

The IndexNow step is not decoration. Microsoft advises that generative AI values fresh content partly as a reference check on its training data and recommends pushing updates through IndexNow as they happen. And when citations do move, turning that into a number a budget holder respects is its own discipline. Our guide to measuring GEO ROI when the traffic is invisible covers it properly.

What should you do next?

Ship the Organization block with a stable @id on your entity home this week. Add the Companies House identifier while you are in there. Start the citation baseline before changing anything else, so future you can prove what worked and what was coincidence.

Then step back. Markup is one factor among several, and on the evidence above it matters less than visible, quotable, well-cited content. If you want to see where your biggest gaps sit before spending another hour on JSON-LD, the free AI marketing scorecard takes about ten minutes.

Schema will not rescue weak content. It never has. A clean entity graph, though, is one of the few things in AI search you control completely, and it costs little enough that the unsettled science should not stop you building it.

Scoring a page before you publish it

The signals in this section are checkable, which means they can be scored rather than argued about. Whether the schema validates, whether claims carry sources, how recently the page was genuinely revised, and whether the technical basics let a crawler read it at all.

Axiom GEO, our own platform, scores landing pages with 59 checks across schema, content, freshness, E-E-A-T and technical SEO, so the gap between two pages is a number instead of an opinion. Run it on a page you are proud of before you run it on one you are not, because the surprise is usually in the first.

Frequently asked questions

For the AI Overviews slice specifically, see Schema Markup for AI Overviews, What You Actually Need.

Does FAQ schema still do anything now Google restricts FAQ rich results?

The visible questions and answers matter more than the markup. In the AirOps ChatGPT dataset, FAQPage markup carried a 45.6% citation rate, among the strongest types, though that is correlation rather than proven cause. Well-structured question and answer content gives assistants quotable, self-contained passages, which helps whether or not Google ever shows a rich result. Keep the format, keep the markup honest, and expect no miracles from either.

How do I get into Google's Knowledge Graph without a Wikipedia page?

You do not need Wikipedia. Create a Wikidata item, which has a far lower bar, reference it from your Organization sameAs, and make your entity home unambiguous, one page stating who you are, what you do and where you are registered. Corroborate with consistent external records such as Companies House and LinkedIn, then wait for a recrawl. Graph builders reward consistency across independent sources more than any single listing.

JSON-LD, microdata or RDFa, which format should I use for AI search?

JSON-LD. It keeps markup separate from your templates, it is the format Google's documentation examples use, and @id graph stitching is far easier to maintain in it. The format war barely matters to assistants anyway, since in Otterly's seven-platform test only Gemini could read raw markup of any kind. Choose JSON-LD for maintainability, not for AI magic.

Does llms.txt help my site show up in AI search?

Not in Google. Its optimisation guidance states plainly that Google Search does not use llms.txt files. I publish one on my own sites anyway, because it costs five minutes and other crawlers may choose to read it, but I treat it as an unproven courtesy rather than a tactic. The agencies selling llms.txt packages have not published evidence that it moves anything.

Can wrong or spammy schema markup hurt my AI visibility?

Yes. The clearest warning in the research is that markup contradicting your visible content is worse than no markup at all. Stale prices, inflated ratings and copy-pasted blocks describing someone else's business teach systems to distrust your data. Audit quarterly, generate schema from the same source of truth as the visible page, and delete anything you cannot keep accurate.

How long until schema changes show up in AI answers?

Nobody has published a reliable timeline, and the four-to-twelve-week figures you will see are asserted without measurement. Index-time effects cannot begin until a recrawl, so confirm reindexing first, then judge against a control set over at least eight weeks. For pages that are already cited, Ahrefs found no lift at all across its August 2025 to March 2026 window, so set expectations accordingly.

Put this to work on your business

We build exactly this, SEO, AI visibility, PPC, email and automation, as one measured system. Talk it through with a real person, no obligation.

Book a free strategy call or take the two minute scorecard