Wikidata Statements, References, Ranks and Qualifiers

The statement as the unit of assertion

Wikidata is the best-known public knowledge graph that stores claims rather than facts. Its own help pages say so plainly: Wikidata "does not state what the population of Germany actually is; it simply provides the information on what the population of Germany is according to a specific source" ^[1]. That choice shapes everything below.

The vocabulary comes from the Wikibase data model, the software model behind Wikidata ^[2]:

• An item is a thing being described, identified by a QID such as Q64 (see the Wikidata QID reference). A property (P-number) names a kind of data and carries a datatype.
• A snak is the smallest assertion the model can make: a property with a value, or one of two special forms described below. The data model notes that editors never meet the word; it exists so developers do not confuse a snak with a statement.
• A statement has a subject, one main snak, zero or more qualifier snaks, a list of references, and a rank. The data model gives it as Statement(Entity Snak {PropertySnak} {ReferenceRecord} Rank).

Older documentation separates a claim (main snak plus qualifiers) from a statement (claim plus references); the JSON documentation calls this distinction historical ^[3]. Here, statement means the full structure; for the general concept, see the glossary entry on claims.

Three kinds of snak

A snak's snaktype is one of value, somevalue or novalue ^[3]. The editing interface calls the last two unknown value and no value.

• Value: the property has this specific value.
• Somevalue (unknown value): the property has a value, but the value is not known. The data model's example is Ambrose Bierce's date of death: unknown, yet he is certainly not among the living ^[2].
• Novalue (no value): the property has no value for this item. Help:Statements calls this "a positive statement" and gives the example of a person who had no children ^[1].

Both separate a declared absence from incompleteness, and neither is a placeholder value: if "none" were a value, a query for items sharing a value would conclude that Australia and Iceland have a common neighbour ^[4].

Qualifiers: the context of a value

A qualifier is a property-value pair attached to the statement, not to the item. The original design paper is explicit: for "population 2,651,040, as of 2010, method estimation", the extra pairs "do not refer to Rome but to the assertion that Rome has a population of 2,651,040" ^[4].

Help:Qualifiers lists four jobs ^[5]:

1. Modify what the item means ("France, excluding Adélie Land").
2. Say how the value was determined (determination method: estimation, census).
3. Constrain the validity of the value (point in time, start time, end time).
4. Add details a single pair cannot hold (a cast member's role in a film).

A statement "should still provide useful data even without a qualifier", and qualifiers modify the main value or the statement as a whole, never other qualifiers ^[5]. When a property should have one value but sources report several, keep all of them and qualify each ^[1].

References: where a claim comes from

A reference is a set of snaks describing a source, from a single URL to title, author, publisher and page ^[2]. The JSON documentation states that references "record provenance information for the data in the main Snak and qualifiers" ^[3]. It supports the whole qualified assertion.

Help:Sources sets the working conventions ^[6]:

• stated in (P248) points to an item describing a publication, database or report; reference URL (P854) points to a web page.
• For web pages, add the publication date or, failing that, retrieved (P813), plus title and an archive URL "so that the source can be tracked down if the URL changes".
• Statements supported only by imported from Wikimedia project (P143) "are not considered sourced statements", and Wikipedia articles are not acceptable sources.

The data model permits an empty reference list: editors can add statements without a reference and others may supply one later (Wikibase/DataModel; see also Hogan et al., Knowledge Graphs, §10.1.4).

Ranks: preferred, normal, deprecated

A rank is a three-level marker on each statement ^[7]:

• Normal is the default and "provides no judgement or evaluation of a value's accuracy and currency".
• Preferred marks "the most current statement or statements that best represent consensus". Several statements may be preferred at once.
• Deprecated marks statements "known to include errors" or representing outdated knowledge. It is not for correct historical values, which keep normal rank and carry start time and end time qualifiers.

The page recommends recording a reason with reason for preferred rank (P7452) or reason for deprecated rank (P2241). It also separates ranks from references: "References merely state where a data value comes from; ranks indicate what data value is considered the most correct" ^[7]. Deprecating rather than deleting stops the value being re-added, and a deprecated statement must still be sourced, since the source really did make the claim.

From ranks the data model derives the best rank for a property on an item: preferred if any statement has it, otherwise normal. The "best statements" are those with the best rank ^[2].

Worked example: Berlin's population

Figures are from Berlin's live entity JSON ^[8], read 28 October 2026; they will change.

Q64 carried 47 statements for population (P1082): 46 normal, 1 preferred, none deprecated. 29 of the 47 had no reference. Two statements describe the same reference date:

Read with their qualifiers, the figures measure different things. Wikidata keeps both; rank picks the default for queries and infoboxes, without recording why.

Statement A in JSON, abridged:

{
  "id": "Q64$C5506726-07D8-4E52-8AB4-BB5FB9317501",
  "type": "statement",
  "rank": "preferred",
  "mainsnak": {"snaktype": "value", "property": "P1082",
    "datavalue": {"type": "quantity", "value": {"amount": "+3700577", "unit": "1"}}},
  "qualifiers": {
    "P585": [{"snaktype": "value", "datavalue": {"type": "time",
      "value": {"time": "+2025-12-31T00:00:00Z", "precision": 11}}}],
    "P459": [{"snaktype": "value", "datavalue": {"value": {"id": "Q52679562"}}}]
  },
  "references": [{
    "hash": "3384faddfcb22f1d6f76da385283f5e62ee583e1",
    "snaks": {"P248": [{"datavalue": {"value": {"id": "Q141265959"}}}],
              "P813": [{"datavalue": {"value": {"time": "+2026-09-03T00:00:00Z"}}}]}
  }]
}

Male and female population qualifiers are omitted. Precision 11 means significant to the day ^[3].

Export to RDF

The RDF dump writes every statement twice ^[9]:

• Truthy statements link the item directly to a value with a wdt: predicate. They cover "the best non-deprecated rank for a given property", and "the qualifiers are ignored".
• Full statements are separate nodes (wds:) linked from the item with p:. The node holds the rank (wikibase:rank), the value (ps:), qualifiers (pq:) and references via prov:wasDerivedFrom. Statements with the best rank also have type wikibase:BestRank.
• Reference nodes (wdref:) are named by a hash of their content, so identical references share a node; their values use pr:.

The same Berlin statements, from the entity's Turtle export ^[10], abridged:

wd:Q64 wdt:P1082 "+3700577"^^xsd:decimal .        # truthy: date and method gone

wd:Q64 p:P1082 wds:Q64-C5506726-07D8-4E52-8AB4-BB5FB9317501 .
wds:Q64-C5506726-07D8-4E52-8AB4-BB5FB9317501 a wikibase:Statement, wikibase:BestRank ;
    wikibase:rank wikibase:PreferredRank ;
    ps:P1082 "+3700577"^^xsd:decimal ;
    pq:P585 "2025-12-31T00:00:00Z"^^xsd:dateTime ;
    pq:P459 wd:Q52679562 ;
    prov:wasDerivedFrom wdref:3384faddfcb22f1d6f76da385283f5e62ee583e1 .

wdref:3384faddfcb22f1d6f76da385283f5e62ee583e1 a wikibase:Reference ;
    pr:P248 wd:Q141265959 ;
    pr:P813 "2026-09-03T00:00:00Z"^^xsd:dateTime .

wd:Q64 p:P1082 wds:Q64-d789393f-445d-f810-b5ce-84543997a922 .
wds:Q64-d789393f-445d-f810-b5ce-84543997a922 a wikibase:Statement ;
    wikibase:rank wikibase:NormalRank ;
    ps:P1082 "+3913644"^^xsd:decimal ;
    pq:P585 "2025-12-31T00:00:00Z"^^xsd:dateTime ;
    pq:P459 wd:Q15194024 .

Statement IDs have "no guaranteed format or meaning" (note $ in JSON, - in RDF). An unknown value is a blank node, which the query service replaces with skolem IRIs tested via wikibase:isSomeValue. A no-value statement is a wdno: class on the statement or entity, defined in OWL as the complement of having any value for that property ^[9]. Giving the edge its own node so context can attach to it is reification (Hogan et al., §3.3.2).

Mapping Wikidata constructs to provenance concepts

The export uses one W3C provenance term, prov:wasDerivedFrom. PROV-O defines derivation as "a transformation of an entity into another, an update of an entity resulting in a new one, or the construction of a new entity based on a pre-existing entity" ^[11]. The other rows are analogies, not part of the export.

Conflicts and constraints

Wikidata does not prevent conflicting statements. Property constraints are statements on a property, using property constraint (P2302), that describe expected usage, and they are "hints, not firm restrictions" with allowed exceptions ^[12]. Four bear on provenance:

• Single value: items should have no more than one statement with this property.
• Single best value: items should have one "best" statement, for properties that change over time. Population (P1082) carries it; Berlin satisfies it with one preferred statement.
• Citation needed: statements should have at least one reference.
• Mandatory qualifiers: statements must carry the listed qualifiers.

Disputes are settled socially: other views may be added "as long as they include a source and appropriate qualifiers", and consensus "should be indicated by a preferred rank" ^[1]. A statement can also name its opponents with the qualifier statement disputed by (P1310) ^[7].

Lessons for enterprise knowledge graphs

Coexistence and election are different layers. Wikidata keeps every sourced claim and elects a default with rank. A system of record must name one current value per subject, property and date, because a filing needs one answer. Keep both layers: the claims with their references, and the election as a separate dated act. The Knowledge Vault article covers the automated version of the same problem, fusing extracted claims at web scale.

Record why a value was elected. Wikidata recommends but does not require a reason for a rank; Berlin has none. Store the rule or reviewer, date and evidence behind each election.

Qualify before you compare. Berlin's figures conflict only until method is read. Compare statements only when subject, property, scope, time and method align.

Declare unknowns explicitly. Somevalue and novalue separate "unknown" and "none" from "not yet entered"; enterprise records need the same, with a dated search boundary (declaring unknowns).

Do not publish only the truthy view. A wdt:-style projection drops dates, methods and sources; consumers, including AI agents, need the qualified statement.

Limits

• Rank is editorial. It records community consensus, any editor may change it, and it may have no stated reason ^[7].
• References are optional. The model allows empty reference lists, and 29 of Berlin's 47 population statements had none on the day read.
• A reference is a pointer. It names a source; it does not hold a copy. Archive URLs are recommended for web pages, not enforced ^[6].
• Revision history is not valid time. The entity's revision counter and last-modified timestamp ^[9] record when the record changed, not when a value held in the world. Valid time lives in qualifiers, where present.

Checklist for borrowing the pattern

This article is educational and is not legal or regulatory advice.

Reliable Context applies these lessons inside the customer's perimeter: one canonical record per real thing; a cited, dated source behind every value; unknowns declared; any past state reproducible.

Sources

1. Help:Statements, Wikidata (read 28 Oct 2026) — "according to" a source, unknown and no value, plurality and consensus.
2. Help:Qualifiers, Wikidata (read 28 Oct 2026) — four uses of qualifiers and the rules on qualifying.
3. Help:Sources, Wikidata (read 28 Oct 2026) — reference properties, archive URLs, P143 not a source.
4. Help:Ranking, Wikidata (read 28 Oct 2026) — preferred, normal and deprecated rank; reasons; ranks vs references; statement disputed by.
5. Help:Property constraints portal, Wikidata (read 28 Oct 2026) — constraints as hints; single value, single best value, citation needed, mandatory qualifiers.
6. Wikibase/DataModel, MediaWiki (read 28 Oct 2026) — snaks, statement structure, ReferenceRecords, empty reference lists, ranks and best rank.
7. Wikibase JSON format, Wikimedia documentation (read 28 Oct 2026) — snaktypes, claim vs statement, reference scope, time precision.
8. Wikibase/Indexing/RDF Dump Format, MediaWiki (read 28 Oct 2026) — truthy and full statements, prefixes, special values, revision metadata.
9. Berlin (Q64) entity data, JSON, Wikidata (read 28 Oct 2026) — population statements, ranks, qualifiers and references in the worked example.
10. Berlin (Q64) entity data, Turtle, Wikidata (read 28 Oct 2026) — RDF triples in the worked example.
11. Vrandečić & Krötzsch, "Wikidata: A Free Collaborative Knowledgebase", Communications of the ACM 57(10), October 2014 — design rationale for qualifiers, special values, plurality and ranks.
12. Hogan et al., "Knowledge Graphs", arXiv:2003.02320v6, September 2021 — reification (§3.3.2); Wikidata as a secondary source with optional references and contradictory claims (§10.1.4).
13. PROV-O: The PROV Ontology, W3C Recommendation, 30 April 2013 — terms in the mapping table.