Semantic Triples in SEO: Subject, Predicate, Object

SUBJECT · PREDICATE · OBJECT

A semantic triple is a fact stated in three parts, a subject, a predicate and an object, and it is the unit in which search engines store what they know about entities. The model comes from RDF, the W3C data model of the semantic web, where every statement is one triple and a knowledge graph is a set of triples sharing entities. Search engines extract triples from page text, HTML tables, DOM structure and structured data, and research at Google scored 119 million webpages by how many of their extracted triples were correct. Semantic SEO writes for that extraction: one fact per sentence, the subject first, an explicit predicate and a precise object, with structured data stating the same triples the text states. A good entity definition is an is-a triple followed by the attributes that separate the entity from its siblings, and facts that change over time need a qualifier such as a date in the same sentence. Holistic Radar keeps one value per fact across holisticradar.com, and a triple audit of any page takes six steps.

What a semantic triple is

A semantic triple is a statement with three parts: the subject is the entity the statement is about, the predicate is the property or relation, and the object is the value or the other entity. The sentence “Google launched the Knowledge Graph in 2012” contains the triple ⟨Knowledge Graph, launch date, 2012⟩ and the triple ⟨Knowledge Graph, developer, Google⟩.

SENTENCEGoogle launched the Knowledge Graph in 2012.object: developerpredicatesubjectobject: date⟨Knowledge Graph, developer, Google⟩ and ⟨Knowledge Graph, launch date, 2012⟩
Figure 1 · One sentence, two triples

The object decides which kind of triple it is. When the object is a literal value, such as a date, a number or a string, the triple states an attribute. When the object is another entity, the triple states a relation, and the two entities become connected nodes. A set of relations sharing entities forms a graph; that graph is what Google calls the Knowledge Graph.

Triples are the semantic SEO name for what databases call the entity-attribute-value model. The three names map one to one: subject to entity, predicate to attribute, object to value. A topical map uses the same structure at the level of a whole site, where each entity-attribute pair owns one URL and the page under that URL states the values. The semantic SEO examples article shows one entity-attribute-value record built this way.

One sentence can hold several triples, and one triple can be written in many sentences. “Holistic Radar, founded by Bilal Sameer, runs semantic SEO engagements” holds two triples: ⟨Holistic Radar, founder, Bilal Sameer⟩ and ⟨Holistic Radar, provides, semantic SEO⟩. The triple is the fact; the sentence is one of its verbalisations, and semantic SEO chooses the verbalisation an extractor reads with the least effort.

LITERAL OBJECTAttributetriple⟨Knowledge Graph, launch date,2012⟩Object is a value: date, number,stringDescribes one entityENTITY OBJECTRelation triple⟨Holistic Radar, founder, BilalSameer⟩Object is another entityConnects two nodes into a graphvsSubject, predicate, object map one to one onto entity, attribute, value
Figure 2 · Attribute triple and relation triple

Where triples come from: RDF and the semantic web

RDF, the Resource Description Framework, defined the triple as the data model of the semantic web. The W3C published RDF as a Recommendation in 1999 and revised it as RDF 1.1 in 2014. In RDF the subject is an IRI or a blank node, the predicate is always an IRI, and the object is an IRI, a literal or a blank node. Identifying the predicate by an IRI means two publishers who use the same predicate mean the same property.

RDF1999Schema.org2011Knowledge Graph2012RDF 1.12014Knowledge Vault2014KBT2015
Figure 3 · From RDF to page-level triples

Schema.org applied the same model to ordinary web pages. Google, Bing and Yahoo launched the vocabulary in June 2011, and Yandex joined in November 2011. Every schema.org property is a predicate, and every property value written in JSON-LD, Microdata or RDFa is one triple about the item it belongs to.

The semantic web assumed publishers would mark up their facts. Most did not, so search engines learned to extract triples from unmarked text, and that shift is why sentence structure matters to semantic SEO as much as markup does.

How search engines extract triples from pages

Search engines extract triples from four kinds of source on a page: running text, HTML tables, the DOM tree, and human annotations such as schema.org markup. Knowledge Vault, a Google research system presented at KDD in 2014, combined 16 extraction systems across those sources and assigned every extracted triple a calibrated probability of being correct. A triple stated in two sources with different extractors earns a higher probability than a triple one extractor found once.

Pagesourcestext,tables, DOM,markupExtract16 systemsin KnowledgeVaultLinkentitiesto knowledgebase nodesFuseagree orconflictProbabilityper tripleA triple stated by several sources through different extractors earns a higherprobability
Figure 4 · Four sources, one fused triple

Extraction from text uses the same steps a Python parser runs: sentence splitting, named entity recognition, entity linking and relation extraction. The Python for semantic SEO article shows a dependency parser turning “Semantic SEO requires a topical map” into ⟨Semantic SEO, require, a topical map⟩. A sentence whose subject is a pronoun, whose predicate is buried in a clause, or which joins three facts with commas gives the extractor less to work with and lowers the probability of what it does extract.

Extracted triples are fused with triples already in the knowledge base. When the web agrees with the knowledge base, the fact is confirmed. When a page states a value no other source states, the triple stays low-confidence until other sources repeat it.

How triples decide trust: Knowledge-Based Trust

Knowledge-Based Trust (KBT) is a method Google researchers published in 2015 to score a web source by the correctness of the triples it states, instead of by the links pointing to it. The paper applied the method to 2.8 billion triples extracted from the web and estimated the trustworthiness of 119 million webpages and 5.6 million websites.

Triples extracted2.8 billionWebpages scored119 millionWebsites scored5.6 millionBar lengths are ordinal, not to scale. Trust = probability a source states thecorrect value
Figure 5 · The scale of the Knowledge-Based Trust study

KBT defines the trustworthiness of a source as the probability that it states the correct value for a fact, given that it states a value at all. A source with few false triples is trustworthy; a source that repeats popular errors is not, however many links it has. The paper separates two kinds of mistake: an extraction error, where the system misread a correct page, and a source error, where the page itself was wrong. Only source errors count against the page.

KBT is a research method, and Google has not stated that it runs as a ranking signal. The principle it demonstrates holds in every retrieval system that fuses facts: a page whose triples agree with the confirmed consensus costs less to verify than a page whose triples contradict it, and a page whose triples contradict each other costs the most.

Triples in sentences: writing for extraction

Writing for extraction means each sentence states one triple in subject, predicate, object order with no inference left to the reader. The rules follow from how parsers fail:

  • One fact per sentence. A sentence with three facts produces three chances for a misattached object.
  • The subject named, not referred to. “It” and “this” require coreference resolution, an extra step with its own error rate.
  • An explicit predicate. “Founded”, “costs” and “is located in” name the property; “is about” and “relates to” name nothing.
  • A precise object. “216 pages” is a value; “a lot of pages” is not.
01One factper sentence02Named subjectno it or this03Explicitpredicatefounded, costs,located in04Precise object216 pages, not alot
Figure 6 · Four rules that follow from how parsers fail

The rewrite below shows the difference on one sentence about a content network.

VersionSentenceExtractable triple
BeforeIt ended up pretty big once the team rebuilt things, a couple of hundred pages or so.None: no named subject, no value
AfterThe PropXpro content network runs on 216 pages.⟨PropXpro content network, size, 216 pages⟩

Sentence-level rules for modality, word order and numeric precision belong to microsemantics; the triple is the unit those rules protect.

Triples in structured data

Structured data states triples directly, without extraction. In the JSON-LD below, every property line is one triple whose subject is the Organization node.

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Holistic Radar",
  "url": "https://holisticradar.com/",
  "founder": { "@type": "Person", "name": "Bilal Sameer" },
  "knowsAbout": ["Semantic SEO", "Topical maps", "AI search optimization"]
}

The block states six triples: the type, the name, the URL, the founder relation, and three knowsAbout values. The founder property is a relation because its object is another entity, a Person node, which can carry its own triples.

Structured data adds confidence only when the text states the same triples. Google’s structured data guidelines require markup to describe content visible on the page, and a mismatch between markup and text is itself a contradiction an extractor records. The structured data article covers which schema types each page role carries.

Holistic RadarOrganizationholisticradar.comurlBilal Sameerfounder · PersonSemantic SEOknowsAboutTopical mapsknowsAboutAI searchoptimizationknowsAboutThe text must state the same triples; markup that disagrees with the text is acontradiction
Figure 7 · The triples inside one JSON-LD block

Triples and entity definitions

A good entity definition is an is-a triple followed by the triples that separate the entity from its siblings. The is-a triple places the entity in a class; the distinguishing triples state the attributes no other member of the class shares. Logic calls the two parts genus and differentia.

DefinitionIs-a tripleDistinguishing triple
A topical map is a content plan that assigns every entity-attribute pair of a domain to one URL.⟨Topical map, is a, content plan⟩⟨Topical map, assigns, entity-attribute pairs to URLs⟩
A topic cluster is a content plan that links one pillar page to pages about subtopics.⟨Topic cluster, is a, content plan⟩⟨Topic cluster, links, pillar page to subtopic pages⟩
Content planis-a classTopical mapassigns pairs to URLsTopic clusterlinks pillar to subtopicsThe is-a triple places the entity; the distinguishing triple identifies it
Figure 8 · Genus and differentia as triples

Both definitions share the class, so the distinguishing triple carries the whole difference. A definition that names only the class (“a topical map is an SEO strategy”) states a triple true of hundreds of things and identifies none. A definition that skips the class (“a topical map helps you rank”) states an effect and leaves the extractor without a type. The definition sentence is the first sentence under the heading that names the entity, so the question and its answer sit next to each other and no other sentence competes as the definition.

Triples that change over time: qualifiers

A triple without a time is only true while its value holds, so facts that change need a qualifier: a date, a period, a place or a condition attached to the triple. Wikidata stores qualifiers on each statement, such as a start time on a position held or a point in time on a population figure. A knowledge graph that receives “the price is 49 dollars” with no date cannot tell whether the page is current or stale.

STALENo qualifierKnowledge Graph holds 500 millionobjectsTrue once, false laterTwo values look like acontradictionDURABLEWith qualifier500 million objects at launch, May2012True permanentlyTwo dated values form a time seriesvs
Figure 9 · An unqualified triple goes stale

The qualifier belongs in the same sentence as the triple it qualifies. “The Knowledge Graph held 500 million objects at launch in May 2012” states the value and the moment it was true; “The Knowledge Graph holds 500 million objects” states a value that stopped being true years ago. A page that separates the figure from its date across two paragraphs forces the extractor to join them, and the join fails more often than either sentence alone.

Qualifiers also resolve apparent contradictions. Two pages stating two employee counts for one company contradict each other until each carries its year, at which point they form a time series. Semantic SEO dates every figure that can change: prices, counts, rankings, team sizes and versions of software.

Triples on holisticradar.com

Holistic Radar keeps one value per fact across holisticradar.com, because two pages stating two values for one predicate create a contradiction no extractor can resolve. The site keeps a register of facts, each with its value, the page that owns it, and the wrong wordings to search for. The central entity, the predicate pair and the quality-node count are three entries, and the Map Template page owns all three.

SubjectPredicateObjectOwner page
Holisticradar.comcentral entitySemantic SEOMap Template
Holisticradar.compredicate pairKnow and RankMap Template
Holisticradar.comquality nodes at launch17Map Template

Every other page that mentions one of these facts either states the identical value or links to the owner page instead of restating it. A sentence that repeats a fact owned elsewhere is a second source for the same triple, which helps only when the value is identical.

How to audit the triples of a page

A triple audit lists the facts a page states and checks each against the text, the markup and the rest of the site. Six steps cover it:

  1. Extract every factual sentence from the main content.
  2. Rewrite each as ⟨subject, predicate, object⟩; a sentence that will not rewrite states no fact.
  3. Mark sentences with a pronoun subject, a vague predicate or an imprecise object.
  4. Compare each triple with the page’s structured data and record mismatches.
  5. Search the site for the same subject and predicate and record conflicting values.
  6. Rewrite the marked sentences and correct the conflicts at the owner page first.
Extractfactual sentencesRewriteas triplesMarkweak subject, predicate,objectMarkupcompare JSON-LDSiteconflicting valuesCorrectowner page firstSteps 1 and 2 scale with a parser; steps 3 to 6 need a person who knows the correctvalue
Figure 10 · Six steps of a triple audit

Steps 1 and 2 run in Python with a dependency parser on large sites; steps 3 to 6 need a person who knows which value is correct.

Semantic triples questions

Are semantic triples a Google ranking factor?

Google has not named triples as a ranking factor. Triples are the form in which Google stores facts in the Knowledge Graph, so a page whose facts extract cleanly gives Google more confirmed facts to associate with the page and its source.

What is the difference between a triple and a keyword?

A keyword is a string a searcher types, and a triple is a fact with a subject, a predicate and an object. “Topical map definition” is a keyword; ⟨Topical map, is a, content plan⟩ is a triple that answers it.

How many triples should a page contain?

A page should contain every triple its query network asks for and no triple that belongs to another page’s entity-attribute pair. The count follows from the map, not from a target number.

Semantic Triples in the semantic SEO topical map

Semantic Triples is the fact-unit attribute of the page semantics entity inside the Holistic Radar topical map, and it sits in the outer section under the Know predicate.

The outer section holds the page because it explains how facts are stored and written without selling a service, and its bridge passes session signals toward Semantic Copy, where every client sentence is written as an extractable triple.

→ Semantic copy: client pages written one extractable fact per sentence.

→ What are microsemantics: the sentence rules that protect each triple.

→ What is entity SEO: entities, attributes and relations across a site.

→ Python for semantic SEO: extracting triples from sentences with a parser.

Get my free Radar ScanFree Radar Scan