Structured data and AI search: a working glossary

Schema.org is a collaborative project launched in 2011 by Google, Bing, Yahoo, and Yandex. It defines hundreds of entity types — Organization, Product, Article, LocalBusiness, Review, FAQPage, BreadcrumbList, and many more — each with a structured set of properties.

Pages add Schema markup by embedding a JSON-LD script block in the page HTML (a script tag with type "application/ld+json"). Search engines parse the block when crawling the page and use it to populate rich results — review stars, FAQ accordions in search, product price ranges — and to enrich knowledge graph entries.

For local businesses, the highest-leverage Schema types are: LocalBusiness (or the more specific subtype like Restaurant, Dentist, RealEstateAgent) for the business itself, Review and AggregateRating for review surfaces, FAQPage for FAQ sections, BreadcrumbList for navigation, Service for service descriptions, and Organization with proper sameAs linking to Wikipedia, LinkedIn, and other authoritative profiles.

AI search engines also use Schema markup heavily — Perplexity, ChatGPT, and Google AI Overviews extract structured data to answer questions confidently. A page with clean LocalBusiness markup is far more likely to be cited when someone asks AI about the business.

Schema.org defines two relevant types: Review for an individual review (with author, reviewBody, reviewRating) and AggregateRating for the summary (ratingValue, ratingCount, reviewCount). They're typically nested inside a parent entity like Product, LocalBusiness, or Service.

Google's structured data guidelines have tightened over the past five years. To qualify for star rating rich snippets in search results, the rated entity must be of an eligible type (Product, Service, LocalBusiness, etc.), the ratings must be aggregated from real customer reviews visible on the page, and the ratings must not be self-reported without external verification.

Self-claiming a 5-star average from "847 customers" without a visible review widget on the page is increasingly likely to be ignored or trigger manual review. Google's quality raters check the page for an actual review surface — embedded testimonials, links to Trustpilot/G2/Yelp, or pulled-in third-party widgets.

For tools-and-software businesses, the cleanest implementation is to embed verifiable third-party reviews (G2 widget, Trustpilot widget) and let Schema markup describe what's actually on the page rather than asserting ratings the page doesn't prove.

DefinedTerm extends Schema.org's Intangible type. Each entry typically has: name (the term), description (the definition), inDefinedTermSet (a reference to the parent glossary), and optionally identifier (a unique slug or ID).

Google and AI search engines use DefinedTerm markup as a signal that the page is a structured definition source — typically pulled from these sources preferentially when generating AI Overviews for "what is X" queries.

Investopedia's glossary is the canonical example of DefinedTerm at scale — over 30,000 financial term entries, all marked up with DefinedTerm schema, all linked through a single DefinedTermSet. This is why Investopedia dominates AI citations for financial definitions even though many newer sources have better-written content.

This glossary uses DefinedTerm markup on every individual entry page, linked through a single DefinedTermSet at /glossary. The implementation pattern follows Investopedia's — clean definition first, supporting context after, related terms linked inline.

Launched globally as Search Generative Experience in 2024 and renamed AI Overviews in 2025, the feature uses Google's Gemini model to generate a multi-paragraph answer to many search queries before the traditional ten blue links are shown. Each Overview cites three to seven source URLs that contributed to the answer.

For publishers, ranking in the source pool that feeds Overviews is increasingly important and structurally different from ranking in traditional results. AI Overviews tend to favor: definition-pattern content (glossaries, "what is X" pages), step-by-step procedural content (how-to guides), and content with named entities clearly tagged through schema markup.

Definition pages are particularly effective because the Overview model needs to authoritatively define terms used in answers. Pages that lead with a clean single-sentence definition — like every entry in this glossary — get extracted verbatim more often than pages that bury the definition under setup text.

Google has not published the ranking factors for AI Overviews, but observable correlations include: long-form content (1500+ words), structured data markup, named author attribution, freshness signals (recent dateModified), and high topical authority (a site about reviews is more likely to be cited on review topics than a generalist site).

Proposed by llmstxt.org in late 2024, llms.txt is to AI crawlers what sitemap.xml is to search engines — a curated index that surfaces the pages worth reading. The file lives at /llms.txt and contains:

- A top-level H1 with the site name - A blockquote summary describing what the site is and who runs it - H2 sections grouping the key URLs by topic (Product, Guides, Pricing, etc.) - Each URL paired with a one-line description of what the page contains

The format is intentionally minimal: human-editable, no required schema, plain Markdown. Unlike sitemap.xml which exists to maximise crawl coverage, llms.txt exists to maximise citation accuracy — telling the model "if someone asks about X, this is the page to cite."

Adoption is voluntary and asymmetric — Anthropic's Claude appears to consume llms.txt actively, OpenAI's ChatGPT consumes it sporadically, Google's AI Overviews currently does not. The cost of adding it is near-zero (a single file) and the upside in AI citation accuracy is real, so most sites with a clear product or content strategy now ship one.

A separate proposed standard, llms-full.txt, includes the full content of each linked page concatenated into one file. Useful for smaller sites where total content is under the typical context window of an LLM (around 100k tokens). Larger sites usually omit llms-full.txt and let the model crawl individual URLs on demand.

E-E-A-T evolved from E-A-T with the addition of "Experience" in December 2022. The four factors are evaluated together, not separately:

Experience — first-hand experience with the topic. A restaurant review by someone who ate there, a product review by someone who used it, a guide written by someone who has actually worked in the field.

Expertise — domain knowledge, demonstrated through credentials, methodology, or sustained practice. Medical content from a board-certified physician scores higher than the same content from an anonymous writer.

Authoritativeness — recognition by other authorities in the field. Inbound links from respected publications, citations in industry research, professional credentials displayed on the byline.

Trustworthiness — accuracy, citations, transparent author identity, and honest disclosure of conflicts. Pages with named human authors, verifiable credentials, and cited sources score materially higher than the same content with anonymous bylines.

E-E-A-T is not a direct ranking factor — Google's quality raters evaluate content for E-E-A-T as part of training the ranking algorithms, but Google doesn't maintain a numerical E-E-A-T score per page. The practical effect is that pages with strong E-E-A-T signals tend to rank better over time, especially in competitive verticals.

For local businesses, the leverage points are: named author bylines on guides, real testimonials with verifiable sources (G2, Trustpilot), explicit editorial standards, and proper schema markup naming the business as the publisher.