Search Query Clustering & Negative Keyword Systems

1. The Broad-Match Expansion Dilemma

When bidding on commercial search phrases like "enterprise CRM software", modern search algorithms will routinely match broad ad impressions to search terms like:

• crm customer support phone number (Zero purchasing intent - existing user experiencing tech issues)
• free open source crm github (Zero commercial monetization intent - developer looking for source code)
• crm software login portal (Navigational intent - existing subscriber attempting to log in)
• crm jobs salary denver (Employment inquiry - job seeker)

Allowing paid traffic to flow to these intent types destroys conversion rate and burns marketing capital.

Figure 15: Real-time search query clustering engine analyzing n-gram intent patterns and applying automated negative keyword rules.

2. The Three-Tier Negative Keyword Architecture

To insulate campaigns from broad-match leakage, we maintain global negative keyword lists organized into three hierarchical tiers:

3. Algorithmic N-Gram Search Term Clustering

To discover negative keyword candidates at scale, our data pipeline executes an automated n-gram tokenization script across hundreds of thousands of search query impressions:

# Automated N-Gram Search Term Analysis Logic
from collections import Counter
import re

def extract_wasteful_ngrams(search_terms_df, max_cvr_threshold=0.005):
    # Filter terms with significant spend but near-zero conversions
    waste_terms = search_terms_df[search_terms_df['cvr'] < max_cvr_threshold]
    
    tokens = []
    for term in waste_terms['query']:
        clean_words = re.findall(r'\w+', term.lower())
        # Generate 1-gram and 2-gram combinations
        tokens.extend(clean_words)
        tokens.extend([f"{clean_words[i]} {clean_words[i+1]}" for i in range(len(clean_words)-1)])
        
    frequency = Counter(tokens)
    return frequency.most_common(50)

4. Measured ROAS Lift from Negative Keyword Shielding