Semantic Entity Checklist
The page-audit skill loads this when scoring Dimension 2 (Semantic Depth & Topical Completeness). Google's NLP (BERT, MUM) builds a semantic graph of every page. If yours is missing nodes and edges that competitors have, you lose, even when keyword density is identical. This rubric identifies which nodes are missing.
The four core audit questions
Work through these in order. Each one produces a specific finding.
1. Which entities from this topic's semantic field are present, and which are missing?
Entities are the people, companies, products, concepts, locations, and events that make up a topic. For "Kubernetes operators" the core set includes: CRDs, controllers, reconciliation loops, etcd, kubectl, Helm, Prometheus, Operator SDK, KUDO. A page that never mentions CRDs isn't an expert page, it's an outline. Pull the entity list from the top 3 competitors (fetched in Step 3 of the skill) and flag anything they all mention that the audited page omits.
2. What predicates (verbs / actions) belong to this topic's semantic field, and does this page use them?
Predicates are the load-bearing signal for expertise depth. A generalist writing about coffee brewing uses "make," "prepare," "put in." Someone who actually brews uses: grind, extract, bloom, tamp, pour, steep, filter, agitate. The predicate vocabulary alone tells Google's NLP whether you're describing the thing or inhabiting it. Audit: list the verbs used in the main body. Fewer than 5 domain-specific predicates = cap Dimension 2 at 5 regardless of word count.
3. How dense are the Entity-Attribute-Value triples?
An EAV triple is a fact of the form [Entity] [has attribute] [with value]. Example: [Aeropress] [brew time] [1-2 minutes] or [Helm chart] [default timeout] [300 seconds]. Experts know specific values; generalists describe qualitatively ("Aeropress brews quickly"). That sentence has a relationship but no value, and Google's NLP can't extract a clean triple. Count EAV triples per 500 words. Strong pages hit 15+. Weak pages hit 3 or fewer.
4. What subtopics do the top 3 competitors cover that this page doesn't?
Most "content gap" tools surface keyword overlaps instead of conceptual gaps. Do it manually: list the H2s and H3s for each competitor, then compare against the audited page. If two or more competitors have a "common failure modes" section and the audited page doesn't, that's a gap. If all three cover a subtopic the audited page handles in one sentence, that's a gap.
Worked example: "best coffee brewing method"
Thin semantic profile (scores ~4)
- Entities: coffee, water, cup, filter, grinder
- Predicates: make, prepare, pour, add, wait
- EAV triples: "coffee needs hot water" (relationship, no value)
- Subtopics: types of coffee makers, instructions
This reads like a page written by someone who has never actually brewed coffee beyond a drip machine. It covers the topic but doesn't inhabit the domain. Google's NLP will parse it into a handful of weak triples and rank it below anything with real depth.
Rich semantic profile (scores ~9)
- Entities: V60, Chemex, Aeropress, French press, Moka pot, burr grinder, gooseneck kettle, scale, tamper, portafilter, TDS meter, specialty roaster, single origin, blonde roast, natural process, washed process, crema, bloom
- Predicates: grind, extract, bloom, tamp, tare, agitate, pre-infuse, steep, decant, plunge, invert, pour, swirl, filter, pre-wet
- EAV triples:
[V60] [grind size] [medium-fine],[Aeropress] [brew time] [1:30-2:30],[Chemex] [paper filter] [25% thicker than V60],[extraction] [target TDS] [1.15-1.35%],[bloom] [duration] [30-45 seconds],[water temperature] [optimal range] [90-96°C] - Subtopics: bean-to-water ratios, grind size per method, water chemistry (TDS target ranges), common extraction mistakes (channeling, under/over-extraction), equipment calibration, bloom timing, agitation techniques, method-specific troubleshooting
The second version reads like an expert because it is one. The EAV density alone (15+ specific values in this snippet) signals domain depth to NLP that no amount of keyword variation can fake.
1–10 scoring anchor
10. Full entity network. 10+ unique expert verbs. 15+ EAV triples per 500 words. Covers all subtopics the top 3 competitors cover, plus one they don't.
7–8. Most entities covered. Domain predicates present but inconsistent. 5–10 EAV triples per 500 words. Misses 1–2 competitor subtopics.
5–6. Surface-level entity coverage. Generalist predicates dominate. Few EAV triples. Missing several competitor subtopics. Reads like summary, not expertise.
3–4. Thin entities, almost no domain predicates, no EAV triples (all qualitative). Misses most competitor subtopics. Reader learns nothing beyond a Wikipedia excerpt.
1–2. Generic wrapper. Target keyword appears, but no domain content underneath. Zero expert verbs, zero specific values. Usually AI generation without grounding or a writer with no domain access.
Quick checks during the audit
- Predicate count (30 seconds): scan the main body for verbs. Fewer than 5 domain-specific ones = Dimension 2 capped at 5.
- EAV density (1 minute): pick a 500-word section and count specific values (numbers, measurements, named settings, thresholds). Fewer than 3 = cap at 5.
- Competitor subtopic diff (2 minutes): paste the audited page's H2s and each competitor's H2s side by side. Any H2 that appears in 2+ competitors but not the audit target is a named gap.
These three checks will get you 80% of the way to an accurate Dimension 2 score in under 4 minutes.
Cross-reference
This file is the scoring rubric. For actually closing the gaps (building a semantic brief, running entity extraction against competitors, generating a predicate list), invoke the semantic-gap-analysis skill. Use this checklist to find the problem; use that skill to fix it.