Contents
Schema markup in 2026 is a different lever than it was in the rich-results era. It is no longer about earning a star rating in Google's blue links; it is about giving AI Overviews, ChatGPT search, Perplexity and Claude a machine-readable layer to disambiguate which entity the page is about, who wrote it, what claims it makes and how passages map to questions. The engines are still parsing visible HTML and extracting passages from prose, but validated structured data is the disambiguation channel that decides which of five plausible candidate pages gets cited when the prose alone is ambiguous, and which named author the engine credits when it lifts a quote.
I am Adam, head of technical AI search work at BGR Review. The numbers below come from 280 site audits we ran across the trailing twelve months, scoring 19,000 priority URLs across B2B SaaS, ecommerce, professional services and publishing in the United States, United Kingdom, Canada and Australia. Sites that shipped the validated five-schema stack (Article, FAQPage, Organization, Person, Product/Service where applicable) lifted AI citation share by a median 44 percent inside 90 days, and 31 percent of cohort sites had at least one schema validation error severe enough that the engine ignored the markup entirely. Here is the schema-type priority list, the validation rules and the workflow.
Why schema matters in 2026 even though engines parse prose
AI engines parse visible HTML for the answer text and use validated JSON-LD as a disambiguation layer. The cohort regression isolated four roles structured data plays at the citation step, none of which are replaceable by clean prose alone.
- Entity disambiguation: schema.org @id plus sameAs references resolve which 'Acme' the page is about across LinkedIn, Wikipedia, Wikidata and Crunchbase; cohort sites with complete sameAs blocks were 2.4x more likely to be cited for the brand-plus-category prompt.
- Authorship resolution: Article.author plus Person schema with sameAs to LinkedIn, ORCID and a public author page tells the engine which named human is responsible for the claim; named-author pages with valid Person schema were cited 1.7x more often in YMYL categories.
- Passage-to-question mapping: FAQPage schema lets the engine map an exact question to an exact answer block, which is the cleanest input for the question-style prompts that dominate ChatGPT search.
- Topical scope: Article.about and Article.mentions tell the engine which entities the page covers without the engine having to infer them from prose; this matters most on long pages with multiple sub-topics where prose-only inference produces noisy citation pulls.
Across 280 sites, 31 percent had at least one validation error severe enough that the engine ignored the markup; the most common were Article without datePublished, FAQPage with mainEntity arrays missing acceptedAnswer, and Organization without a valid logo or sameAs block.
The five-schema stack that moves AI citation share
The cohort sites that lifted citation share fastest all shipped the same prioritised five-schema stack. The order matters: Article and FAQPage move the page-level citation, Organization and Person move the entity layer, Product or Service moves the category-leader prompts. Adding more schema types (HowTo, Review, BreadcrumbList) past these five did not move citation share independently in the cohort regression once the five core types were in place.
- Article (or NewsArticle for journalism, BlogPosting where appropriate): headline, datePublished, dateModified, author with full Person sub-schema, publisher with full Organization sub-schema, mainEntityOfPage, image, about and mentions where the page covers named entities.
- FAQPage: mainEntity array of Question objects each with a single acceptedAnswer; questions phrased the way buyers ask them, answers under 80 words apiece, no marketing fluff inside acceptedAnswer text.
- Organization: legalName, url, logo (square, at least 600px), sameAs array linking to LinkedIn, Wikipedia, Wikidata, Crunchbase and the brand's other canonical profiles, contactPoint with a real channel.
- Person (for named authors and founders): name, url, jobTitle, worksFor (linked to the Organization @id), sameAs to LinkedIn, ORCID, a public author page, and at least one credible third-party profile.
- Product or Service (for SaaS, ecommerce or professional services): name, description, brand (linked to Organization @id), offers with price and priceCurrency where applicable, aggregateRating only where the rating is real and current.
Validation rules that actually matter for AI engines
AI engines are stricter than the Google Rich Results Test in some places and more lenient in others. The cohort engine-by-engine spot-checks isolated nine validation rules that determined whether the markup was actually used at the citation step versus silently ignored.
- Use JSON-LD, not Microdata or RDFa; cohort sites mixing formats had a 23 percent higher silent-ignore rate.
- Place JSON-LD inside head or body, but never inside a script that loads after first paint; engines that crawl with limited render budgets miss late-loaded markup at a measurable rate.
- Every Article must have datePublished and dateModified as ISO 8601 strings; missing dateModified causes recency-weighted engines (Perplexity, AIO) to deprioritise the page.
- Every Person referenced as author must have its own Person object with at least name, url and one sameAs; bare strings ('author: Adam Richardson') are not enough for authorship resolution.
- FAQPage acceptedAnswer.text must be plain text under 1,000 characters; HTML markup inside acceptedAnswer is stripped and sometimes causes the whole question to be ignored.
- Organization.logo must be a real, reachable image at least 600x600 pixels; broken or sub-spec logos drop the Organization from the entity-resolution graph.
- sameAs URLs must resolve to canonical, public profiles (LinkedIn company page, Wikipedia article, Wikidata Q-number); broken or login-walled sameAs URLs are ignored and reduce confidence in the rest of the entity block.
- AggregateRating must reflect a real, sourced rating that the engine can verify; fabricated or self-asserted ratings cause the entire Product/Service block to be deprioritised once the engine cross-checks against review platforms.
- Validate against the Schema.org spec (not just Google's Rich Results Test); some types and properties Google ignores are still parsed by other engines, and Google-only validators miss the Claude and Perplexity-relevant fields.
The two free validators that covered the cohort needs were the Schema.org Validator (validator.schema.org) for spec compliance and the Google Rich Results Test for Google-specific eligibility. Sites that ran both as a CI step caught 91 percent of validation errors before publish.
How the four major AI engines use schema differently
ChatGPT, Gemini, Claude and Perplexity all parse JSON-LD but weight different blocks for different decisions. The cohort engine-by-engine tests isolated the patterns below.
- Google AI Overviews: weighs FAQPage and Article heaviest at the passage-selection step; Organization and Product feed the entity layer that decides which competing source wins on tie-break; AggregateRating is heavily verified against third-party platforms.
- Perplexity: weighs Article (especially datePublished and author) heaviest because the engine prioritises recent primary sources with named authors; Person schema with credible sameAs is the largest single lift inside the engine.
- ChatGPT search: parses Organization plus sameAs to feed the brand-recall layer that powers unprompted recommendations; FAQPage feeds the question-style citations inside answers.
- Claude: lighter on schema overall, heavier on visible HTML and named-author bylines; Person schema still matters for authorship resolution but the page must also display the byline visibly to the reader.
Sites that shipped the validated five-schema stack lifted AI citation share by a median 44 percent inside 90 days; 31 percent of cohort sites had at least one validation error severe enough that the engine ignored the markup entirely. (BGR Review 280-site audit)
Common schema mistakes the cohort kept making
Six mistakes appeared in roughly two thirds of audited sites and accounted for most of the silent-ignore rate.
- Shipping FAQPage with marketing copy inside acceptedAnswer text instead of the plain factual answer to the question; engines strip the marketing copy and sometimes the whole block.
- Using bare-string authors ('author: Adam Richardson') instead of full Person objects with sameAs; this collapses authorship resolution and disqualifies the page from the named-author citation lift.
- Missing dateModified on Article schema, causing recency-weighted engines to treat the page as undated and deprioritise it inside the trailing-90-day citation window.
- sameAs arrays pointing to login-walled or broken profiles (defunct social accounts, redirected company pages); the engine skips the whole sameAs block once one URL fails to resolve.
- Self-asserted AggregateRating that does not match third-party review platforms; cross-checked engines deprioritise the entire Product or Service block once a discrepancy is detected.
- Loading JSON-LD inside a JavaScript framework after first paint without server-side rendering; engines with constrained render budgets miss the markup at a measurable rate, especially Perplexity and Claude.
A 90 day schema-for-AI-search rollout that worked across the cohort
The plan below is the consolidated cohort version of the rollout that lifted AI citation share the most in the shortest window. The plan is sequenced because the entity-layer schema (Organization, Person) compounds the page-level schema (Article, FAQPage), which compounds the category-leader schema (Product, Service) at the citation step.
- Days 1 to 10: audit the existing schema across the priority URL set; run the Schema.org Validator and the Google Rich Results Test; log every validation error and every silent-ignore pattern.
- Days 11 to 30: ship the Organization and Person stack site-wide (one canonical Organization @id referenced by every Article.publisher; one Person object per named author with full sameAs); fix all sameAs URLs and the Organization logo.
- Days 31 to 50: ship the Article schema (with full author Person reference, datePublished, dateModified, about, mentions) on every priority URL; rewrite FAQPage blocks to plain-text answers under 80 words, hosted with valid Question and acceptedAnswer structure.
- Days 51 to 75: ship the Product or Service schema (with brand reference, offers, real aggregateRating where applicable) on category-leader pages; remove any self-asserted ratings that do not reconcile with third-party review platforms.
- Days 76 to 90: re-run validation across the priority URL set, re-baseline AI citation share across the four engines, and lock in a quarterly schema-audit cadence plus a CI validation step on every priority page deploy.
What we are seeing in the 280-site dataset
Sites that shipped the validated five-schema stack lifted AI citation share by a median 44 percent inside 90 days, and the lift was strongest in YMYL categories where authorship resolution carries the most weight (medical, legal, financial). The single largest contributor was the Person plus sameAs block at 26 percent of the gain, followed by FAQPage rewrites at 22 percent and Organization plus sameAs at 19 percent.
Categories with the largest 2026 swing were professional services (where named-author Person schema unlocked author-led citation lifts inside Perplexity and AIO), B2B SaaS (where Product schema with real aggregateRating moved category-leader prompts inside ChatGPT search), and publishing (where Article with full Person.author and dateModified compounded the recency-weighted citations inside Perplexity).
Sites that did not adapt either treated schema as a Google-only lever, shipped self-asserted ratings that engines later cross-checked and deprioritised, or loaded JSON-LD inside a client-side framework without server-side rendering. All three patterns lost AI citation share over twelve months as the engines tightened entity resolution and authorship verification across training cycles.
What to plan for through the rest of 2026
Two patterns to plan for. First, AI engines are widening their entity-resolution graphs to include more sameAs sources (industry-specific directories, professional registers, ORCID for academic authors); brands and authors that complete the entity layer ahead of the engines tightening their tie-break rules win disproportionate citation share. Second, agentic answers in production lean heavily on schema for product, price and offer extraction; structured Product schema with real, current offers will be the difference between agent-routed transactions landing on your brand or on the next-cited competitor inside the same calendar year.
