◆ Semantic Web Catalogue

Where we are

For each design question the semantic web has had to answer: what is settled, what is open, what pushes it one way or another, and the next moment when a decision can change it.

As of 23 Oct 2026. Each paragraph sums up the problem threads listed under it, where every claim is sourced. Coming back? The board lists what is on now.

Across all issues

Works that address the Semantic Web as a whole.

The 2001 vision's software agents arrived in another form: AI systems that call services through protocols such as MCP, which counted more than 10,000 public servers when it moved to a Linux Foundation fund in December 2025. Whether agents will rely on shared vocabularies and published data, as the W3C argued in October 2026, or on each service's own tools, is open.

Threads
Forces
Language models need knowledge to check against

TechnicalIdentification

What is a thing called?

Mostly working, with the hard parts open. The 2005 httpRange-14 rule, that a 303 redirect lets a URI name anything, is still how Linked Data publishers such as Wikidata work. For people, platform accounts became the names in practice: Wikidata records X, Facebook and Instagram usernames on hundreds of thousands of items, and those names change or close at one company's choice. Open: how a name can outlive its host (redirect services depend on whoever runs them, and DID 1.1 and DID Resolution reached W3C Candidate Recommendation in 2026); who controls the registries most identifiers come from; and how to say two datasets name the same thing without overclaiming, since owl:sameAs proved too strong.

Threads
Forces
Laws that require verified identity and labelled content · Running infrastructure costs money and effort · Distrust of platform owners
Next decision
24 Dec 2026: EU digital identity wallets due

TechnicalLinking

How do things connect?

Reopened. The Web's one-way links won because anyone can make them, and the cost is link rot: in 2024 Pew found that 38% of pages from 2013 no longer load. Webmention and ActivityPub bring back links that the target hears about, archives keep copies of what was cited, and in AT Protocol every link is a public record that indexers can count. None has displaced the plain one-way link. Readers still have no standard place on someone else's page: annotation tools such as Hypothesis use the W3C's 2017 Web Annotation format, but most notes stay inside the tool that made them.

Threads

TechnicalData model

What shape is the data?

Two halves. How to say who made a statement is largely settled: named graphs and PROV, and now RDF 1.2's triple terms, whose core specification became a W3C Candidate Recommendation in April 2026, with canonicalization (2024) and Data Integrity proofs (2025) for signing. Whether triples or property graphs are the right shape is open: ISO published GQL for property graphs in 2024, and proposals such as Amazon's OneGraph try to serve both from one model.

Threads
Next decision
30 Apr 2027: RDF & SPARQL Working Group charter ends

TechnicalSerialization

How is it written down?

Largely settled in favour of what developers already know. After RDF/XML, the formats that spread were the familiar-looking ones: microformats and RDFa inside HTML, then JSON-LD in 2014. What is still argued is whether the rigour survives when developers never see the RDF underneath. Newer work goes into compact forms, such as the CBOR-LD draft of April 2026.

Threads
Forces
Language models need knowledge to check against
Next decision
1 Nov 2026: New Wikidata query endpoints open

TechnicalReasoning

What follows from it?

Reopened by AI. Description logics mapped how expressiveness costs tractability, and OWL 2's profiles settled a working compromise in 2009. Language models reopened both threads: they predict facts rather than derive them, losing the guarantees of sound reasoning, yet they need knowledge to be checked against. Proposals range from graphs that ground models, such as GraphRAG and Wikidata's 2025 vector database, to the argument that general statistical methods keep winning.

Threads
Forces
General methods that scale keep winning · Language models need knowledge to check against
Next decision
31 Dec 2026: Data Shapes Working Group charter ends

TechnicalQuery & validation

How do we ask and check?

Largely settled. SPARQL (2008) is the standard query language, and SHACL (2017) gave a standard way to check that data is complete despite OWL's open-world assumption; ShEx stays in use for Wikidata's EntitySchemas. SHACL 1.2 is in drafts to cover RDF 1.2, and its working group is chartered to the end of 2026.

Threads
Forces
Running infrastructure costs money and effort
Next decision
1 Nov 2026: New Wikidata query endpoints open

TechnicalScale & performance

Does it work at the size needed?

Open, and expensive. The hope of one query across the whole Web of data gave way to central endpoints and cheaper interfaces after monitoring found public SPARQL endpoints often down or slow. Wikidata's query service, the largest public one, split its graph in 2025 and is moving from the unmaintained Blazegraph to QLever, with new endpoints opening to the community on 1 November 2026.

Threads
Forces
AI companies harvest what is published openly · Running infrastructure costs money and effort · Public money for open internet software
Next decision
1 Nov 2026: New Wikidata query endpoints open

BridgeVocabulary

What terms do we share?

Open on every front. Schema.org became the Web's common vocabulary through the search engines; its 2026 releases added mappings to GS1, Dublin Core and Open Graph and terms for EU Digital Product Passports. Hundreds of other vocabularies coexist, catalogued by LOV and the OBO Foundry, and matching between them is measured each year by the Ontology Alignment Evaluation Initiative. Libraries are moving from MARC to BIBFRAME with no date set to stop MARC, and vocabularies grown from use, as in Wikidata, rely on community proposals and shape checks.

Threads
Forces
General methods that scale keep winning · Language models need knowledge to check against

HumanTopology

Where does data live, and who hosts it?

Open. The Web was built without a centre, and data recentralized on large platforms. Solid, ActivityPub, local-first software and AT Protocol each move data or identity back toward users, in different ways. In 2026 ActivityPub got a new W3C working group, AT Protocol an IETF one and Solid's protocol a first W3C draft, while the EU declined to require social networks to interoperate and Bluesky's active use fell by about a quarter in a year.

Threads
Forces
Data protection law · Laws that require verified identity and labelled content · Running infrastructure costs money and effort · Laws that require portability and interoperability · Distrust of platform owners · Protocols change slowly, platforms quickly · Public money for open internet software
Next decision
3 Nov 2026: NLnet funding deadline

HumanTrust

Who vouches for it?

Open. Signing is solved in parts: Verifiable Credentials 2.0 and Data Integrity became W3C Recommendations in 2025, and Content Credentials sign photos, on the Pixel 10 for every picture. Whose signatures to trust is not settled. PGP left it to individuals and PICS to rating services; now governments and device makers step in, through EU digital identity wallets due by the end of 2026 and the AI Act's duty, from August 2026, to mark AI-generated content.

Threads
Forces
Data protection law · Laws that require verified identity and labelled content · Laws that require portability and interoperability
Next decision
2 Dec 2026: AI Act marking deadline for existing systems

HumanGovernance

Who decides, who may use it, and on what terms?

Open, and the busiest. Who may reuse openly published data has reopened with AI: Creative Commons proposed CC signals, the IETF is drafting a vocabulary for AI-use preferences, and the W3C plans ODRL 3.0. Commons projects face crawler load and fewer human visits, and Wikimedia now sells high-volume access to AI companies. Standards come from many venues: W3C and IETF groups, the search engines' steering group for schema.org, and a Linux Foundation fund for AI agents. The Web's habit of keeping everything now meets a legal right to be forgotten, which federated copies, archives and trained models make hard to honour. Privacy rules keep moving; the EU dropped its proposal for browser consent signals in June 2026.

Threads
Forces
AI companies harvest what is published openly · Search answers instead of sending visitors · Data protection law · Laws that require verified identity and labelled content · Running infrastructure costs money and effort · Laws that require portability and interoperability · Distrust of platform owners · Protocols change slowly, platforms quickly · Public money for open internet software · Search engines reward markup and set its terms
Next decision
3 Nov 2026: NLnet funding deadline

HumanAdoption & incentives

Why would anyone publish or use it?

Open. Publishers added markup when search engines rewarded it with richer results. That reward is shrinking: Google retired several rich-result types in 2025, and fewer people click through when search answers directly. The W3C argues that structured data gains value as AI agents read it, and schema.org keeps growing, for instance into the product passports the EU is introducing.

Threads
Forces
AI companies harvest what is published openly · Search answers instead of sending visitors · Language models need knowledge to check against · Protocols change slowly, platforms quickly · Search engines reward markup and set its terms
Next decision
1 Nov 2026: New Wikidata query endpoints open

HumanSustainability

Who pays for it, and for how long?

Open. Shared identifiers, query services and knowledge bases are cheap to launch and costly to keep. Wikimedia sells high-volume access to AI companies, Wikidata is rebuilding its query service, and the EU funding behind much open internet software paused in mid-2026 while it moves to a new programme. Whether public funds, the largest users or commons foundations carry the upkeep is unresolved.

Threads
Forces
AI companies harvest what is published openly · Search answers instead of sending visitors · Data protection law · Running infrastructure costs money and effort · Distrust of platform owners · Public money for open internet software
Next decision
3 Nov 2026: NLnet funding deadline

HumanUsability

Can ordinary developers and people use it?

Open. The full stack of triples, URIs, ontologies and SPARQL stays hard for non-specialists, and JSON-LD hid much of it behind ordinary JSON. Language models offer another route by writing SPARQL from plain questions, and Wikidata opened a vector database reachable through MCP in 2025. Whether developers should see RDF at all is still disputed.

Threads
Forces
Language models need knowledge to check against
Next decision
1 Nov 2026: New Wikidata query endpoints open