How LLMs actually use schema (and how it differs from Googlebot)
Googlebot parses JSON-LD into a structured knowledge graph — it knows that the Person on this page has a name, a job title, a sameAs link to their LinkedIn. The major LLMs in 2026 (GPT-5, Claude 3.7, Gemini 2.0, Perplexity Sonar) extract entities and facts from rendered HTML during their crawl, then use schema as a confirmation or disambiguation signal.
What this means in practice: schema that confirms what is visibly on the page lifts trust and improves citation rate. Schema that contradicts the page (or claims facts not present on the page) is increasingly detected and downweighted. The era of 'add FAQ schema with answers nobody can find on the page' as a free lunch is over.
Treat JSON-LD as a structured caption for what is on the page, not as a secret backchannel to convey extra information that isn't visible. If a fact is in your schema, it should be in your visible body. If it isn't, it shouldn't be in your schema.
The schema types that move LLM citation in 2026
| Schema type | Leverage | Notes |
|---|---|---|
| Person (with sameAs) | High | Bylines + author bios; sameAs to LinkedIn/Twitter/ORCID |
| Organization (with knowsAbout, areaServed, sameAs) | High | Site-wide; knowsAbout disambiguates topical authority |
| Article (with author, dateModified, datePublished) | High | Recency + authorship; LLMs prefer recent, attributed content |
| Most-specific Service/LocalBusiness sub-type | High | LegalService, Dentist, HVACBusiness vs generic LocalBusiness |
| BreadcrumbList | Medium | Site structure / hierarchy; helps LLM understand URL relationships |
| FAQPage | Medium | If answers visibly on page; harmful if not |
| HowTo | Medium | Mobile rich results + AI Overview citation; steps must be on-page |
| Review (with reviewBody, datePublished) | Medium | Real reviews lift; templated/inflated reviews demote |
| AggregateRating | Low / risky | Frequently misused; LLMs detect inflation |
| Speakable | Very low | Voice search use case is small; rarely drives LLM citation |
Person schema: the highest-leverage 2026 add
Bylined articles with Person schema and a linked author bio page outperform anonymous articles on LLM citation by a meaningful margin in our 2026 audit set. The Person schema should appear on both the article (as the article's author) and on the author's bio page (as the page's primary entity).
Required fields: name, jobTitle, image, url. Strongly recommended: sameAs (LinkedIn, X, ORCID, Wikipedia if applicable), worksFor (Organization reference), description, knowsAbout (array of topics).
sameAs links are how Google and the LLMs disambiguate your author from someone with the same name. A 'Sarah Chen' with sameAs to a LinkedIn profile is identified as a specific person; a 'Sarah Chen' without sameAs is one of a thousand candidate entities. The disambiguation lifts citation rate measurably.
Organization schema: knowsAbout and areaServed are under-used
Most Organization schema we audit has only the basics: name, url, logo, contactPoint. The fields that move topical authority are knowsAbout (an array of topics the organisation has expertise in) and areaServed (the geographies it operates in). Both are direct signals to the LLM about what queries you should be cited for.
knowsAbout is best populated with 8–15 specific entities — not 'marketing' but 'search engine optimization,' 'local search optimization,' 'answer engine optimization,' 'technical SEO audit.' Each entity should ideally be a Wikidata or schema.org defined concept. areaServed for a Canadian business should list provinces and major cities served, not just 'Canada.'
Article schema: dateModified is a recency signal LLMs read directly
Article schema with a recent dateModified consistently outperforms older or undated articles on LLM citation in our 2026 audit set. The recency signal is read directly by the LLM at retrieval time. Stamping a real, accurate dateModified when content is genuinely updated is a free lift.
Bumping dateModified without actually updating content is detectable and increasingly correlated with citation suppression in our probe set. Major LLMs hash a fingerprint of content over time; a dateModified change with no content change is a trust erosion signal.
Schema anti-patterns to remove this quarter
- 1FAQPage with answers not present in the visible body. Either add the answers to the page or strip the schema.
- 2AggregateRating with inflated values, no review count, or no source. Remove or replace with a real Review block.
- 3Service schema for services you don't actually offer (sometimes inherited from a template). Audit and trim.
- 4Multiple competing primary entities on a single page. Pick one (Article OR Service OR Product), not all three.
- 5Schema for a different entity than the visible page topic (e.g., Recipe schema on a dental page). Sounds absurd, found in 4% of our audits.
- 6Speakable schema with text that doesn't read aloud naturally. Strip it; the use case is too small to chase.
A practical schema validation workflow
- 1Validate on the Schema.org Validator first — catches structural errors regardless of Google preference.
- 2Validate on the Rich Results Test — tells you what Google will surface as rich results.
- 3Spot-check 5 random pages of each template (article, service, location, author bio).
- 4Run a quarterly site-wide schema audit with Screaming Frog (Custom Extraction) or a JSON-LD parser.
- 5When schema changes, monitor the GSC 'Enhancements' report for 7 days for warning spikes.