The marketing director at a 9-location dermatology group spent six months publishing expert content, earning backlinks, and cleaning up her site speed. ChatGPT still cited her competitors. The difference wasn't the writing. It was 14 lines of JSON-LD she hadn't touched.
Structured data has had a quiet year. While everyone argued about content format and brand mentions and E-E-A-T prose, schema markup became the machine-readable proof layer that generative engines use to decide who is citation-worthy. It's not the whole game. But it's the part most teams are getting wrong.
This isn't a technical SEO refresher. This is about the specific schema patterns that ChatGPT, Claude, and Perplexity are actually pulling from right now, and why the way most agencies implement structured data was designed for Google's 2019 crawlers, not 2026 LLM retrieval.
Why LLMs care about schema
Generative engines don't read your page the way a human does. They're not absorbing the narrative flow of your 2,400-word service page. They're scanning for structured signals they can verify, extract, and attach to an entity. Schema markup is how you hand them those signals in a format they don't have to interpret.
Think of it this way. Your content is the argument. Your schema is the footnotes. A good argument without footnotes might be convincing to a human reader. To an LLM deciding whether to cite you, the footnotes are what gets checked.
The mechanism matters here. When ChatGPT or Perplexity surfaces a citation, it's drawing on training data plus real-time retrieval. The retrieval layer needs to confirm signals fast. Structured data is what makes your entity machine-confirmable in under a millisecond. Pages without it make the LLM do more inferential work, and when a cleaner source exists, yours gets skipped.
The five schema types that are working
Not all schema is equal for AI citation. The patterns that matter for generative engines are different from the ones that drive rich snippets in Google. Here's what's actually moving the needle.
Organization schema with sameAs
Organization schema is the identity layer for your brand. Name, URL, logo, address, phone. Most sites have a version of this. What most sites are missing is the `sameAs` array. That's the list of external URLs where your entity is confirmed: your Google Business Profile, LinkedIn company page, Crunchbase, Wikidata if you have it, industry directories.
LLMs use `sameAs` to triangulate. It's the difference between "I found a page that claims to be Level Up Digital Marketing" and "I can confirm this entity exists across seven authoritative sources." The second version gets cited. The first gets skipped in favor of someone else.
FAQPage schema
FAQ schema does two jobs simultaneously. It feeds Google's AI Overviews directly, which is the most visible AI citation surface most service businesses encounter. And it signals to LLMs that your page answers discrete questions in a structured, extractable format, which is exactly how generative engines prefer to retrieve information.
The mistake most teams make: they write three generic FAQ questions at the bottom of a page and mark them up. The pages getting cited have 6 to 10 questions that map directly to how a buyer actually searches. "What does an HVAC tune-up include?" outperforms "What are your services?" by a wide margin, because the LLM is trying to answer the first kind of question.
Article schema with author markup
This one is consistently underimplemented at service businesses. Author markup with explicit credentials is one of the strongest E-E-A-T signals a generative engine can parse without reading your prose. The `author` property should include `jobTitle`, `affiliation`, and a `sameAs` pointing to the author's LinkedIn profile or a Google Scholar page if applicable.
A blog post attributed to "Staff Writer" is invisible to LLMs looking for authoritative sourcing. The same post attributed to a named expert with verifiable credentials is a citable source. Same words. Completely different retrieval outcome.
HowTo schema
HowTo schema is purpose-built for the kind of queries that generate AI responses. "How to choose a commercial roofing contractor." "How to prepare for laser skin resurfacing." These are exactly the questions ChatGPT and Perplexity answer in detail. If your page answers a how-to question and doesn't have HowTo markup, you're making the LLM guess at your structure when a competitor with proper markup is right there.
Service schema
Underused and high-value. Service schema lets you declare specific offerings, service areas, descriptions, and pricing in a format LLMs can parse directly. For a multi-location service business, Service schema attached to location-specific pages is one of the clearest authority signals available for local AI citations. Most agencies skip it because it doesn't drive a Google rich result. That logic misses the point entirely now that AI Overviews exist.
“Your content is the argument. Your schema is the footnotes. An LLM deciding whether to cite you checks the footnotes.”
What most sites actually have
The average service business website has generic WebPage markup auto-generated by their CMS plugin, a basic LocalBusiness schema that's missing half its properties, and maybe a BreadcrumbList. That schema was designed to pass Google's structured data validator in 2020. It passes. It also does almost nothing for AI citation.
Most agencies audit schema for Google compliance and stop there. Passing validation is table stakes. Optimizing for LLM retrieval requires a second pass with different criteria. You're not asking "does this validate?" You're asking "can an LLM use this to confirm my entity, extract my expertise, and cite me with confidence?"
The gap looks like this in practice. We've audited sites where 147 location pages were flagged as duplicate-content risk by a standard SEO crawl. But the real problem wasn't duplication. It was that every location page had identical schema with no location-specific Service or LocalBusiness properties. Fourteen identical JSON-LD blocks are worse than none, because they actively confuse entity resolution.
How ChatGPT and Perplexity actually retrieve
Understanding the retrieval mechanism changes what you prioritize. ChatGPT's browsing and Perplexity's retrieval both use a two-stage process: a fast scan for structured signals, then a deeper read if the initial signals pass a confidence threshold. Schema markup lives entirely in stage one.
Perplexity in particular is aggressive about source verification. It's pulling `sameAs` data, checking domain authority signals, and looking for consistent entity representation across your pages. A brand that appears as three slightly different entity names across its schema blocks ("Level Up Digital," "Level Up Digital Marketing," "Level Up Digital Marketing Group, LLC") creates resolution ambiguity. The engine picks the cleaner source.
Claude's retrieval behavior skews toward authoritative citations when generating responses in professional or research contexts. Named authorship with verifiable credentials is the highest-leverage schema investment for Claude citations specifically, because Anthropic's training heavily weights source authority. A law firm's attorney bios with proper Person schema and bar admission data will outperform a generic practice-area page every time.
This connects to something we covered in why your AI content strategy is invisible to AI indexing systems: the indexing criteria for LLM retrieval and the indexing criteria for traditional Google rankings have diverged. Optimizing for one no longer automatically optimizes for the other.
The implementation gap
Schema markup for AI citation is not a one-afternoon fix. It's closer to a data architecture project. You're declaring a consistent entity graph across your entire site, making sure every page type has the right markup, and maintaining it as you publish new content.
Most teams do a one-time schema implementation and walk away. Schema for AI citation is a living layer, not a checkbox. Every new service page needs Service schema. Every new blog post needs Article schema with proper author markup. Every location expansion needs a new LocalBusiness block with complete `sameAs` and Service children.
The operational ask is real. A Webflow or WordPress site needs either a custom JSON-LD injection workflow or a CMS schema integration that generates markup from structured content fields. We've built this with Sanity as the content layer, generating schema programmatically from the same fields editors fill in anyway. The schema is never an afterthought because it's derived from the content model, not bolted on after.
Most paid "schema tools" are the self-checkout at CVS. They produce something that looks like the right output, but the thinking behind it is still the operator's job. The tool generates a block. You still have to know what properties matter, what values to populate, and how the whole entity graph hangs together. If you don't know, the output is technically valid and strategically useless.
This is also where our SEO and GEO work diverges from what most agencies deliver. A standard SEO engagement audits existing schema for errors. A GEO-aware engagement audits for AI citation readiness. which is a different rubric, a different set of schema types, and a different ongoing maintenance requirement.
Where we're putting the bets
The brands that will dominate AI citations in 2027 are building their schema layer now, while most competitors are still debating whether AI search traffic is worth tracking. The window for early-mover advantage in structured data is open, and it closes the same way every SEO window closes: slowly, then all at once, when a competitor locks in the entity graph and you're playing catch-up.
The AI search traffic numbers support the urgency. Visitors referred by AI tools convert at dramatically higher rates than traditional search visitors, per the Microsoft Clarity data. The channel is small but the intent is surgical. Getting cited by ChatGPT or Perplexity for a high-consideration query in your category is worth more than ten times the equivalent organic traffic. You want to be cited there. Schema is the infrastructure that makes citation possible. See what's actually moving in AEO right now if you want the measurement side of this picture.
The bet we're making: schema markup is about to get the same mainstream attention that page speed got in 2018. Everyone knows it matters. Few have done it right. Then Google makes one announcement, or a major study drops showing citation correlation with schema completeness, and every agency scrambles. The operators who built the foundation before that moment will already be getting cited. The rest will be playing catch-up.