Schema Markup for AI Search: What Helps and What Doesn't (5.5M Responses)

Which Types of Schema Actually Help, Which Don’t, and What 5.5 Million Responses Tell Us

Here’s a contradiction that confuses a lot of GEO practitioners. Semrush’s study of 304,805 URLs found a +22% citation lift associated with structured data, making it the fifth-strongest predictor of AI citation. But Search Atlas’s study of 5.5 million responses across Perplexity, Gemini, and OpenAI found schema markup has no effect on LLM citation frequency.

Both studies are credible. Both used large sample sizes. And both are right.

I’m Gabriel Bertolo, founder of Radiant Elephant, a boutique SEO and GEO agency headquartered in Northampton, Massachusetts. I’ve been implementing entity-rich schema markup for clients across New England and nationally for over 13 years. And the schema question is one I get asked more than almost anything else right now: “Does structured data help with AI search or not?”

The answer is both. And I can show you the line where it matters and where it doesn’t, because we’ve tested it across our own client base.

The resolution is simpler than it seems: Google’s AI systems (AI Overviews, AI Mode) clearly use structured data as an input signal. Standalone LLM platforms (ChatGPT, Perplexity, Claude, Gemini) primarily rely on raw text extraction and semantic similarity. Schema helps you on Google’s AI surfaces. It doesn’t directly influence citation decisions on non-Google platforms.

That distinction should shape exactly how much time and budget you allocate to schema for GEO. I covered schema as one of 15 evidence-backed tactics in the research review synthesizing 12 studies and 17 million citations. This article goes deeper on which types produce results, which don’t, and the priority order I use with every client.

The evidence is genuinely split, and that’s actually useful

On the positive side: Google’s Search team confirmed in April 2025 that structured data gives an advantage in search results. Microsoft’s Fabrice Canel stated at SMX Munich in March 2025 that schema markup helps Copilot’s LLMs understand content. Semrush’s analysis ranked structured data as the fifth-strongest predictor of AI citation behind clarity (+33%), E-E-A-T signals (+31%), Q&A format (+25%), and section structure (+23%).

On the negative side: Search Atlas analyzed 5.5 million AI responses and found zero measurable correlation between schema coverage and citation rates on OpenAI, Gemini, or Perplexity. ALM Corp’s research concluded that “no markup type guarantees inclusion” in AI results.

The practical takeaway: implement schema for the Google AI surfaces where it demonstrably helps, and for the entity disambiguation benefits it provides across all platforms. But don’t treat schema as a standalone GEO strategy. It’s a foundation, not the whole building.

What we’ve seen in our own client work

I’ll give you a specific example from our work here in the Pioneer Valley. We had a local service business, a multi-location company operating across Western Massachusetts, running basic LocalBusiness schema from a generic WordPress plugin. No sameAs links. No Person schema for the owner. No FAQPage markup on their service pages. Their Google Business Profile was solid, reviews were strong, but AI Overviews weren’t citing them for any of the service queries they should have been appearing in.

We rebuilt their schema from scratch. Organization schema with @id, foundingDate, areaServed covering their specific service towns, and sameAs links connecting them to Wikidata, LinkedIn, Yelp, BBB, and their Google Maps listing. Person schema for the owner with jobTitle, knowsAbout, and a LinkedIn sameAs. FAQPage schema on each service page with genuine questions and location-specific pricing: “What does [service] cost in Northampton?” with a real answer including a real price range.

Within six weeks, two things happened. Their organic click-through rate on service pages jumped from 3.7% to 10.3%. And they started appearing in Google AI Overviews for three local service queries they’d never shown up in before.

Did schema alone cause both? No. We also restructured the content with answer-first formatting and added specific local pricing data. But the schema was the piece that connected their entity to verifiable external platforms and gave Google’s AI confidence that this was a real, local business with a real owner and real service areas.

We saw similar patterns in our national GEO case study. Entity-rich schema connecting the client to regulatory terms, industry certifications, and Wikipedia/Wikidata references was part of the foundation that produced 283 AI citations in seven months. The schema didn’t produce citations on its own. But without it, the entity layer was incomplete, and the content optimizations we built on top would have underperformed.

Which schema types produce measurable results

Not all schema is equal for AI visibility. After 13 years of implementation and now specifically tracking AI citation impact across our client portfolio, here’s the hierarchy I’ve found to be most accurate.

Organization schema with sameAs has the strongest individual evidence. Schema App’s controlled study measured a 46% increase in impressions and 42% increase in clicks from sameAs entity linking alone. This works because sameAs addresses entity resolution, not content citation. It’s the bridge between schema implementation and entity optimization.

In our own work, I’ve found that sameAs is the single schema property that produces the most consistent, measurable impact. Every client engagement at Radiant Elephant starts with an entity audit: does your Organization schema exist? Does it have a unique @id? Does it include sameAs links to at least three external authority platforms? If the answer to any of those is no, that’s the first thing we fix. Before content. Before link building. Before anything else.

FAQPage schema nearly doubles ChatGPT citation chances according to SE Ranking’s data. FAQ markup wraps explicit question-answer pairs in structured format, giving both Google’s AI and text-extraction systems clean, structured answers to extract. Each answer should be a self-contained 40-60 word response, the same answer-first format that works for unstructured content.

We’ve tested this extensively on client service pages. The combination of a visible FAQ section on the page (so humans can read it) with FAQPage schema (so machines can parse it) and each answer containing location-specific data or pricing consistently outperforms pages with identical content but no FAQ section. For one Amherst-based professional services client, adding FAQPage schema with five locally-specific questions produced a measurable increase in featured snippet captures within three weeks. Two of those snippets later appeared verbatim in AI Overview responses.

Person/Author schema produces 3-4x higher AI citation rates for domains with strong social proof profiles (SE Ranking). The key properties are jobTitle, worksFor, knowsAbout, and sameAs. This connects your content’s author to a verified entity, giving AI systems confidence in the expertise behind the claims. It directly supports E-E-A-T signal building.

I implement Person schema for myself on every page I author on radiantelephant.com, linking to my LinkedIn, connecting my worksFor to the Radiant Elephant Organization @id, and listing specific knowsAbout topics (SEO, generative engine optimization, web design, schema markup). This is the same approach I recommend for every client: named authors with verifiable credentials, connected to the organization entity through schema.

Article schema with datePublished and dateModified provides the machine-readable freshness signal AI systems use to evaluate content recency. A dateModified value reflecting a genuine recent update compounds with the freshness advantage.

HowTo schema gets preferentially cited by AI systems generating procedural answers. High-value for tutorial and instructional content.

Speakable schema is being floated as an AI extraction priority signal by some practitioners. One unverified claim of +127% increase in voice search referrals. Worth testing as a micro-optimization. Not worth significant investment.

Structured data helps machines read your page. So does proper image alt text, and most sites skip it.

Schema as infrastructure: what it does and what it doesn’t do

I need to be direct about this because I see agencies, including a few here in the Northeast that I won’t name, selling “schema optimization for AI” as though adding structured data alone is going to transform your AI visibility overnight.

It won’t. And I say that as someone who has been doing schema implementation longer than most agencies have been in business.

Schema helps machines understand your content. It reduces ambiguity. It strengthens entity signals. All good things that contribute to citation eligibility. But they’re foundational, not transformational.

The sameAs property is the highest-leverage element because it does something text content alone cannot: it provides machine-verifiable proof that your brand exists as a real entity across multiple authoritative platforms. That’s an entity signal, not a content signal.

Everything else in schema for GEO falls into the “makes your content easier for machines to parse correctly” category. Missing schema can cost you citations because the AI couldn’t confidently identify your entity, your author, or the structure of your Q&A content. But adding schema to thin content won’t make that content citation-worthy.

You still need to pack it with statistics and expert quotes. You still need to structure it for extraction. You still need to make sure AI crawlers can technically access your site. Schema is one layer in the stack. Not the whole stack.

Implementation priority order

This is the order I follow with every Radiant Elephant client, whether they’re a local business in Springfield or a national B2B manufacturer.

First: Organization schema with complete sameAs. Every site needs this on the homepage or about page. Include @id, foundingDate, description, and sameAs links to Wikipedia (if applicable), Wikidata, LinkedIn, Crunchbase, and major social profiles. This is your entity foundation.

Second: Person schema for every named author. jobTitle, worksFor (linked to the Organization @id), knowsAbout, and sameAs linking to LinkedIn, institutional pages, and any Wikipedia entries. If your content doesn’t have named authors, fix that first. Anonymous content has a measurably harder time earning AI citations.

Third: FAQPage schema on pages with genuine Q&A content. Not forced FAQ sections tacked onto product pages for SEO. Genuine questions your audience asks with substantive, standalone answers. Each answer is a potential “citation block” for AI extraction. For local businesses, include location-specific data in the answers. “How much does [service] cost in [city]?” with a real price range.

Fourth: Article schema on all editorial content. Include author (linked to Person @id), publisher (linked to Organization @id), datePublished, and dateModified. This tells AI systems when your content was last updated and who stands behind it.

Fifth: Use @graph to connect entities. If a page has multiple related entities (company + author + article), the @graph technique links them into a coherent entity graph AI systems can traverse. More advanced but produces the cleanest semantic signal. This is how we build schema for every client at Radiant Elephant: connected entity graphs, not isolated schema blocks.

Use JSON-LD exclusively. Not microdata, not RDFa. Google explicitly recommends JSON-LD. It’s separated cleanly from your HTML, easier to maintain, and what AI systems parse most reliably.

Five implementation examples

Example 1: Local plumbing company in Western Massachusetts

Current state: Basic LocalBusiness schema from a WordPress plugin. No sameAs. No Person schema. No FAQPage.

Implementation: Enhanced Plumber-type Organization schema with @id, foundingDate, areaServed listing specific towns (Northampton, Amherst, Hadley, South Hadley, Easthampton), priceRange, aggregateRating from actual Google reviews, and sameAs links to Wikidata, LinkedIn, Yelp, BBB, and Google Maps. Person schema for the owner with jobTitle (“Owner and Master Plumber”), knowsAbout, and LinkedIn sameAs. FAQPage schema on each service page: “How much does drain cleaning cost in Northampton?” “Standard drain cleaning in the Northampton area costs $150-$350 depending on drain type and severity. Kitchen drain cleaning averages $175.”

Why this combination works: The entity signals establish who and where you are. The FAQPage wraps locally-specific pricing in extractable format. When someone asks an AI “how much does drain cleaning cost in Northampton,” your schema-marked answer with a local price range, connected to a verified local entity, has strong citation probability. We’ve seen this exact pattern produce results for clients across the Pioneer Valley.

Example 2: SaaS product pages

Current state: No schema beyond basic website metadata.

Implementation: SoftwareApplication schema with applicationCategory, operatingSystem, pricing via AggregateOffer (lowPrice, highPrice), aggregateRating pulling from G2/Gartner reviews, and featureList. Organization schema with sameAs to Wikipedia, Wikidata, Crunchbase, G2, LinkedIn. FAQPage on the pricing page: “[Product] offers three pricing tiers: Starter at $29/user/month, Professional at $79/user/month, and Enterprise at $199/user/month. All plans include a 14-day free trial.”

Why it works: SaaS pricing queries are among the most common AI search patterns. Structured pricing data gives AI clean, extractable information. The G2 and Gartner sameAs links strengthen entity signals on Copilot (which heavily favors Gartner sources) and ChatGPT (which values review aggregation platforms).

Example 3: Recipe blog

Current state: Basic Recipe schema from a WordPress plugin. Missing author details, nutrition data, and FAQ.

Implementation: Enhanced Recipe schema with author linked to Person schema (including credentials like “Q-Grader certified” or “Le Cordon Bleu trained”), dateModified reflecting actual updates, complete NutritionInformation (calories, protein, fat, carbs per serving), and precise recipeIngredient measurements. FAQPage schema addressing common substitution and technique questions: “Can I use pancetta instead of guanciale for carbonara?” with a substantive answer explaining the flavor trade-off, not just “yes.”

Why it works: Food queries are among the highest-volume AI search categories. Nutrition data serves diet-related fan-out queries. The FAQ addresses substitution questions matching common AI prompts. The author credentials satisfy E-E-A-T standards for food content. The dateModified signal tells AI the recipe has been verified or updated recently.

Example 4: Dermatology practice in the greater Boston area

Current state: Basic MedicalOrganization schema. No physician Person schema. No FAQPage. No credential markup.

Implementation: MedicalBusiness schema with medicalSpecialty (“Dermatology”) and sameAs to Wikidata, Healthgrades, ZocDoc. Physician schema as @type: Physician with medicalSpecialty, hasCredential (board certification from the American Board of Dermatology, with recognizedBy linking to the ABD organization entity), alumniOf (medical school), and sameAs to Healthgrades, Doximity, LinkedIn, and PubMed author profile. FAQPage on condition pages with medically specific answers: “Most acne treatments require 8-12 weeks to show measurable improvement. Topical retinoids typically show initial results at 6-8 weeks. In our practice, the average patient achieves 70-80% clearance by week 12 on a customized regimen.”

Why it works: Healthcare is YMYL. The Physician schema with board certification, medical school alumniOf, and verified sameAs links to Healthgrades and Doximity gives AI machine-readable proof that the content comes from a credentialed medical professional. The practice-specific outcome data in the FAQ (“70-80% clearance by week 12”) adds first-party evidence that generic health sites can’t match. For a dermatology practice competing in the Boston metro area, this entity layer is the difference between being cited and being invisible.

Example 5: B2B manufacturing company

Current state: No schema at all. Standard WordPress site with product pages and certifications.

Implementation: Organization schema with NAICS industry code (332710 for machine shops), numberOfEmployees, hasCredential for ISO certifications (ISO 9001:2015, AS9100D, each with the certifying body listed), and sameAs to Wikidata, LinkedIn, ThomasNet, and D&B. FAQPage on capabilities pages: “What tolerances can you hold on CNC machined aluminum parts?” “For aluminum 6061-T6, we hold standard tolerances of ±0.001 inches on critical dimensions, with capability down to ±0.0005 inches for precision applications. Surface finish options range from 125 Ra to 16 Ra. All critical dimensions verified on a Zeiss CMM with NIST-traceable calibration.” Article schema on all blog and resource content with author Person schema and dateModified.

Why it works: An engineer asking an AI about CNC tolerances needs a specific, verifiable answer. The FAQPage wraps technically precise data in structured format AI can extract directly. The Organization schema with ISO certifications, NAICS code, and ThomasNet sameAs establishes the entity in the manufacturing ecosystem where AI systems look for credible suppliers. This is the kind of schema implementation I build for every B2B manufacturing client at Radiant Elephant, whether they’re a shop in the Connecticut River valley or a national operation.

The bottom line on schema for GEO

Schema is foundational. It won’t win you citations on its own. But missing it creates a gap that competitors with proper implementation will exploit, especially on Google’s AI surfaces where the +22% citation lift is real and documented.

The implementation priority is clear: Organization + sameAs first, then Person for authors, then FAQPage on Q&A content, then Article schema everywhere else. Connect them with @graph. Use JSON-LD only. And don’t let anyone sell you schema as a standalone GEO strategy without the content quality, data density, and technical accessibility to back it up.

I detailed the full schema evidence alongside 14 other tactics with implementation guidance in the complete GEO research review.