Classification Dimensions - Source of Truth

This document is the single source of truth for all classification dimensions used in the system.

---

Overview

There are 3 classification pipelines:

Pipeline	What Gets Classified	Cost Model	Learning
Domain Classification	The business/website (e.g., spotify.com)	FREE stages first, then PAID	LLM >=80% → Vectorize
URL Classification	The specific page (e.g., spotify.com/premium)	Inherits domain, adds page-level	LLM >=65% → Vectorize
Keyword Classification	Search queries (e.g., "best project management software")	Rules → Vectorize → LLM	LLM >=80% → Vectorize

---

Domain-Level Dimensions

These describe what the business/website IS - set once per domain, cached, reused for all URLs.

Dimension	Column Name	Description	Valid Values
tier1_type	tier1_type	High-level business archetype (7 types)	platform, marketplace, commerce, service, information, community, institutional, unknown
domain_type	domain_type	Specific business type (~45 types, constrained by tier1)	See Domain Types by Tier1
ownership_type	ownership_type	Who controls this content	owned, ugc_platform, third_party_publisher, partner_site, integration_partner, unknown
channel_bucket	channel_bucket	Marketing channel classification	pr_earned_media, institutional_authority, citation_local, marketplace_listing, affiliate_partner, influencer_campaign, sponsorship_event, email_newsletter, owned_content_marketing, owned_brand_site, ugc_community, programmatic_seo, risky_blackhat, unknown
media_type	media_type	PESO model classification	owned, earned, paid, shared, hybrid, unknown
quality_tier	quality_tier	Publisher authority level	tier_1 (DR 80+), tier_2 (60-79), tier_3 (40-59), tier_4 (20-39), tier_5 (0-19), unrated
industry	industry	Vertical (optional)	sports, gaming, finance, healthcare, technology, etc.
subcategory	subcategory	Specific niche (optional)	Free text, e.g., mobile_gaming, telehealth, saas

Confidence Columns (Per-Dimension)

Each dimension has its own confidence score (0-100):

Column	Description
tier1_confidence	Confidence in tier1_type
domain_type_confidence	Confidence in domain_type
classification_confidence	Overall/legacy confidence

Metadata Columns

Column	Description
classification_source	How it was classified: rule, v2 (domain database), vectorize, llm, merged
classification_context	JSON with signals, rules applied
last_classified_at	Unix timestamp

---

Domain Types by Tier1

The domain_type is constrained by tier1_type. These are the ONLY valid combinations:

PLATFORM (tools you log into)

domain_type	Examples
saas_product	Slack, Notion, Asana, Salesforce
code_repository	GitHub, GitLab, Bitbucket
app_platform	App Store, Google Play
documentation_portal	GitBook, ReadTheDocs
messaging_platform	Slack, Discord, WhatsApp
social_network	LinkedIn, Facebook, Twitter
audio_platform	Spotify, Apple Podcasts
video_platform	YouTube, Vimeo, Twitch

MARKETPLACE (connects buyers and sellers)

domain_type	Examples
ecommerce_marketplace	Amazon, eBay, Etsy
ticket_marketplace	Ticketmaster, StubHub
real_estate_marketplace	Zillow, Realtor.com
job_marketplace	Indeed, LinkedIn Jobs
service_marketplace	Upwork, Fiverr, Thumbtack
app_marketplace	Chrome Web Store, Salesforce AppExchange
review_marketplace	G2, Capterra, Yelp
directory_citation	Yellow Pages, BBB

COMMERCE (sells products directly)

domain_type	Examples
ecommerce_store	Nike.com, Warby Parker
travel_booking	Delta, Marriott, Hertz
subscription_commerce	Netflix, Dollar Shave Club
product_manufacturer	Apple, Samsung, Ford

SERVICE (sells services)

domain_type	Examples
agency_provider	Ogilvy, McKinsey
pr_distribution	PR Newswire, Business Wire
professional_service	Deloitte, Baker McKenzie
healthcare_provider	Mayo Clinic, Teladoc
financial_service	Chase, PayPal, Stripe
legal_service	LegalZoom

INFORMATION (content is the product)

domain_type	Examples
news_publisher	NYT, TechCrunch, The Verge
magazine_publisher	Wired, Forbes, Inc.
blog_publisher	Medium, Substack
content_publisher	BuzzFeed, HuffPost
review_site	Wirecutter, Consumer Reports
affiliate_review_site	NerdWallet, The Points Guy
reference_wiki	Wikipedia, Investopedia

COMMUNITY (UGC-dominated)

domain_type	Examples
forum_community	Reddit, Stack Overflow
gaming_community	IGN forums, Discord servers
sports_community	ESPN forums, Fantasy Pros
qna_platform	Quora, Stack Exchange
ugc_video	YouTube (as community), TikTok

INSTITUTIONAL (authority-based orgs)

domain_type	Examples
government_site	IRS.gov, CDC.gov
education_academic	Stanford.edu, MIT.edu
nonprofit_org	Red Cross, Wikipedia Foundation
healthcare_institution	Mayo Clinic, Cleveland Clinic
financial_institution	Federal Reserve, FDIC
legal_institution	Supreme Court, ABA
trade_association	IEEE, ACM, NAR

UNKNOWN/RISK

domain_type	Description
pbn_suspected	Private blog network
spam_low_quality	Spam/low quality
unknown_other	Fallback

---

URL-Level Dimensions

These describe what the specific page IS - classified per URL, inherits domain-level info.

Dimension	Column Name	Description	Valid Values
page_type	page_type	Type of page (~40 types)	See Page Types
page_category	page_category	High-level page category	editorial, commercial, ugc, programmatic, utility, asset, risky
tactic_type	tactic_type	Marketing tactic detected (~50 types)	See Tactic Types
is_money_page	is_money_page	High-value conversion page flag	0 or 1
url_pattern	url_pattern	Pattern that matched	Free text
modifiers	Stored in context	Boolean signals	See Modifiers

Page Types

Category	Page Types
Editorial	news_article, opinion_article, feature_article, blog_post, press_release, research_article, resource_guide, howto_article
Commercial	product_page, category_page, brand_page, comparison_page, buying_guide, review_page, landing_page, sales_page, checkout_page, pricing_page
UGC	forum_thread, forum_post, subreddit_index, qna_page, ugc_article, profile_page, comment_thread, video_page, playlist_page, channel_page, social_post, repository_page
Programmatic	location_page, directory_listing, auto_generated_comparison, search_results_page, tag_archive_page, category_index_page
Utility	homepage, about_page, contact_page, careers_page, documentation_page, api_reference_page, login_page, signup_page, settings_page, legal_privacy_page, legal_terms_page, faq_page, support_article
Asset	pdf_document, image_asset, video_asset, download_page
Risk	comment_spam_page, pbn_article_page, malicious_page, parked_domain_page

Tactic Types

Category	Tactic Types
PR/Earned	pr_funding_announcement, pr_product_launch, pr_partnership_announcement, pr_wire_distribution, pr_award_announcement, pr_thought_leadership_feature, pr_generic_news_coverage, pr_crisis_coverage, pr_data_report_feature
HARO/Expert	haro_expert_quote, haro_expert_roundup, haro_data_contribution, haro_case_study_feature, expert_panel_feature
Link Building	guest_post_editorial, guest_post_paid, niche_edit_insertion, resource_page_outreach, broken_link_replacement, skyscraper_outreach, scholarship_link, infographic_embed
Affiliate	affiliate_top_listicle, affiliate_single_brand_review, affiliate_comparison_review, affiliate_coupon_deals_page, affiliate_buyer_guide
UGC	ugc_forum_organic_thread, ugc_forum_seeded_promo, ugc_qna_answer_seeded, ugc_qna_organic, ugc_reddit_organic_discussion, ugc_reddit_astroturf, ugc_video_description_link, ugc_profile_link, ugc_social_organic_post, ugc_social_astroturf_post
Owned	owned_blog_editorial, owned_case_study, owned_feature_landing, owned_resource_hub, owned_doc_or_guide, owned_release_on_own_site, owned_gated_leadgen_asset, owned_webinar_page, owned_podcast_episode_page, owned_changelog_release_notes
Programmatic	programmatic_location_page, programmatic_directory_entry, programmatic_comparison_template, programmatic_longtail_template
Influencer	influencer_review_post, influencer_social_post, sponsorship_podcast_mention_page, sponsorship_event_page, sponsorship_webinar_hosted_by_partner
Marketplace	marketplace_app_listing, marketplace_vendor_profile, marketplace_review_section
Blackhat	blackhat_pbn_link, blackhat_comment_spam, blackhat_hacked_site_injection, blackhat_link_farm_page

Modifiers (Boolean Signals)

Category	Modifiers
Content	has_video, has_audio, has_gallery, has_infographic, has_interactive, user_reviews_present, ugc_present, high_engagement, paywalled
Commercial	affiliate_links_present, sponsored_disclosure_present, nofollow_link, dofollow_link, ugc_rel_tag, sponsored_rel_tag
Quality/Risk	thin_content, auto_generated, low_quality_template_match, keyword_stuffed, excessive_ads, popup_heavy, broken_layout
Foreign/Spam	foreign_language_primary, foreign_language_mixed, non_latin_script, foreign_tld_english_content, machine_translated, gibberish_detected
Technical	http_only, redirect_chain, parked_domain, expired_domain
Expert/HARO	expert_byline, multiple_expert_quotes, data_citation, case_study_format

---

Keyword-Level Dimensions

These describe search queries - classified per keyword.

v2 Dimensions (Recommended)

Dimension	Column Name	Description	Valid Values
journey_moment	classification_journey_moment	Point in buyer journey (12 values)	problem_unaware, problem_aware, solution_curious, category_exploring, feature_comparing, option_evaluating, provider_selecting, purchase_ready, purchase_completing, onboarding, optimizing, advocating
journey_direction	classification_journey_direction	Movement through journey	entering, progressing, stalled, regressing, exiting
expertise_level	classification_expertise_level	Knowledge level	novice, intermediate, advanced, expert
buyer_behavior	classification_buyer_behavior	Purchase approach	impulsive, methodical, price_sensitive, quality_focused, brand_loyal, research_heavy
role_context	classification_role_context	Who is searching	individual_consumer, small_business_owner, enterprise_buyer, developer, marketer, student, professional, hobbyist
demand_pattern	classification_demand_pattern	When demand occurs	evergreen, seasonal_recurring, event_driven, trending_spike, declining
content_decay	classification_content_decay	How quickly content goes stale	timeless, slow_decay, moderate_decay, fast_decay, real_time
query_specificity	classification_query_specificity	Query length/specificity	head, torso, long_tail, ultra_long_tail
has_brand_mention	classification_has_brand_mention	Contains brand name	0 or 1
brand_mentioned	classification_brand_mentioned	The brand name if detected	Free text

v1 Dimensions (Backwards Compatible)

Dimension	Column Name	Valid Values
funnel_stage	classification_funnel_stage	awareness, consideration, decision, retention, advocacy
intent_type	classification_intent_type	informational, navigational, commercial_investigation, transactional, local, support
keyword_pattern	classification_keyword_pattern	question, comparison, superlative, problem, use_case, etc.
ranking_difficulty	classification_ranking_difficulty	very_easy, easy, medium, hard, very_hard
trend_velocity	classification_trend_velocity	rising_fast, rising_slow, stable, declining, volatile
seasonality	classification_seasonality	evergreen, seasonal, event_driven
competitive_density	classification_competitive_density	low, medium, high, dominated
commercial_value_index	classification_commercial_value_index	Numeric (volume x CPC)
sentiment_intent	classification_sentiment_intent	positive, neutral, negative, urgent
content_format_intent	classification_content_format_intent	listicle, comparison_review, how_to_guide, video_preferred, etc.

---

Classification Pipeline Stages

Domain Classification (7 Stages)

Stage	Name	Cost	What It Does
0	Cache	FREE	Check D1 for existing classification
1	Rules	FREE	Known domains, TLDs (.gov/.edu), platform patterns
1.5	Google Ads Categories	FREE	Use cached DFS category data → tier1_type hint
2	Vectorize	FREE	Semantic similarity to labeled domains
3	Low-Noise Crawl	FREE	HEAD + partial GET (8KB), CMS/og:type detection
4	Instant Pages	$0.000125	DataForSEO full page fetch
4.5	Domain Patterns	FREE	Fallback rules for placeholder pages
5	LLM	~$0.0001	Workers AI for ambiguous cases

Early exit: If confidence >= 70% at stages 3 or 4, skip remaining stages.

Self-learning: LLM classifications with >= 80% confidence are upserted to Vectorize.

URL Classification (5 Stages)

Stage	Name	Cost	What It Does
0	Domain Cache	FREE	Use cached domain classification
1	Rules	FREE	URL patterns, known sites
2	Vectorize	FREE	Similarity to labeled URLs
3	Content Parse	$0.000125	DataForSEO Instant Pages
4	LLM	~$0.0001	Workers AI fallback

Self-learning: LLM classifications with >= 65% confidence are upserted to Vectorize.

Keyword Classification (4 Stages)

Stage	Name	Cost	Coverage
1	Rules	FREE	~60% of dimensions
2	Vectorize	~$0.00001	~25% of dimensions
3	Brand Lookup	FREE	1 dimension
4	LLM	~$0.0001	~15% (fallback)

Self-learning: LLM classifications with >= 80% confidence are upserted to Vectorize.

---

Source Files

File	Purpose
src/lib/classification-constants.js	All enum definitions, mappings, helpers
src/data/domain-database.js	Generated from master CSV (4,400+ domains)
classification-data/domains/_master.csv	Curated domain classifications
src/lib/classifier-rules-engine.js	Stage 1: Rules engine
src/lib/classifier-vectorize.js	Stage 2: Vectorize similarity
src/lib/low-noise-crawler.js	Stage 3: FREE homepage metadata extraction
src/lib/classifier-content-parser.js	Stage 4: DataForSEO Instant Pages
src/lib/classifier-llm.js	Stage 5: Workers AI LLM
src/lib/domain-classifier.js	Domain classification orchestrator
src/lib/backlink-classifier.js	URL classification orchestrator
src/lib/target-url-classifier.js	Target URL classification (rules only, FREE)
src/lib/keyword-classifier.js	Keyword classification orchestrator

---

Database Tables

domains

domain TEXT PRIMARY KEY,
tier1_type TEXT,
domain_type TEXT,
ownership_type TEXT,
channel_bucket TEXT,
media_type TEXT,
quality_tier TEXT,
industry TEXT,
subcategory TEXT,
tier1_confidence REAL,
domain_type_confidence REAL,
classification_source TEXT,
classification_confidence INTEGER,
classification_context TEXT,
last_classified_at INTEGER

urls

url TEXT PRIMARY KEY,
domain TEXT,
domain_id INTEGER,
page_type TEXT,
page_category TEXT,
tactic_type TEXT,
is_money_page INTEGER DEFAULT 0,
url_pattern TEXT,
classification_source TEXT,
classification_confidence INTEGER,
last_classified_at INTEGER

keywords

keyword TEXT,
classification_funnel_stage TEXT,
classification_intent_type TEXT,
classification_journey_moment TEXT,
classification_expertise_level TEXT,
classification_buyer_behavior TEXT,
classification_query_specificity TEXT,
classification_has_brand_mention INTEGER,
classification_brand_mentioned TEXT,
classification_source TEXT,
classification_version INTEGER DEFAULT 2,
classification_completed_at INTEGER

---

Current State Issues

Database vs Master CSV Mismatch

The database contains ~50 domain_type values from LLM classifications using an older taxonomy. The master CSV contains ~25 domain_type values using the V3 taxonomy.

Database domain_types (sample of ~8,000 domains):

• saas_product (4441) - Valid V3
• service_business (852) - NOT V3, should be professional_service or agency_provider
• forum_community_board (413) - NOT V3, should be forum_community
• blog_content_site (278) - NOT V3, should be blog_publisher or content_publisher
• entertainment_streaming (235) - NOT V3, should be video_platform or audio_platform
• telehealth_provider (216) - NOT V3, should be healthcare_provider
• personal_site_portfolio (175) - NOT V3, no direct equivalent
• etc.

Master CSV domain_types (sample of ~4,400 domains):

• saas_product (1639)
• ecommerce_store (536)
• government_site (257)
• travel_booking (256)
• financial_institution (245)
• etc.

Fix Required

1. Create mapping from old LLM types → V3 types
2. Update database with corrected types
3. Re-seed Vectorize with V3 types
4. Update LLM prompts to output V3 types only