What Is Structured Data? – Definition and How It Works
Learn what structured data means, how it differs from semi-structured and unstructured, and why extraction is a conversion problem.
Structured data has a defined shape: known fields, known types, and a schema that says what may appear. It is what databases hold and what analysis requires.
1. The Three Categories
| Category | Shape | Examples |
|---|---|---|
| Structured | Fixed schema, typed fields | Database tables, Parquet, validated CSV |
| Semi-structured | Self-describing, flexible | JSON, XML, JSON-LD |
| Unstructured | No inherent field structure | Prose, images, audio, rendered HTML |
The boundary between the first two is enforcement rather than appearance: JSON can carry a rigid structure without anything guaranteeing it.
2. Why the Distinction Matters
Analysis needs structure. Grouping, aggregating, joining, and filtering all assume fields with types and consistent meaning.
Most of the world's information is not structured, which is why so much data work is conversion — turning documents and pages into rows and columns.
3. Web Pages Sit Awkwardly
An HTML page has structure in the markup sense and none in the data sense. The nesting describes presentation, not meaning, and there is nothing to tell a parser which element holds the price.
That gap is the entire discipline of extraction: imposing a schema the source never declared.
4. The Structured Islands Inside Pages
Many pages carry genuinely structured data internally:
- JSON-LD blocks describing products, articles, and organisations.
- Microdata and RDFa attributes on markup.
- State blobs in script tags, holding what the page rendered from.
- JSON endpoints the page calls.
Finding one of these is worth more than a well-crafted selector, because it is a data contract rather than incidental layout — and contracts change far less often.
5. Structuring Unstructured Text
Where no island exists, options are pattern rules, a trained extraction model, or a language model with a strict schema.
The last has changed the economics for long-tail sources: a schema plus validation extracts fields from prose without per-site rules. It costs more per page and needs validation, because a model will confidently produce a well-formed wrong value.
6. Structure Is a Contract, Not a Format
Data can sit in JSON and still be unusable if fields are inconsistently named, types vary between records, and nulls mean different things in different places.
Declaring a schema and enforcing it at the boundary is what actually makes data structured. The file format is secondary.
7. Practical Guidance
- Look for structured islands before writing selectors.
- Declare a schema for what you extract, and validate against it.
- Normalise values so the same thing has one representation.
- Keep the raw source, since extraction improves.
- Record null semantics per field.
8. Summary
Structured data has a defined, enforced schema; semi-structured data is self-describing but unconstrained; unstructured data has no field structure at all.
Web extraction is conversion between these categories, and the highest-leverage move is finding the structured island a page already contains rather than inferring structure from its layout.