AI Keyword Classifier

The Problem

Every SEO strategy requires keyword classification—categorizing thousands of keywords by intent, topic, funnel stage, and other dimensions. At this 60-client agency:

Manual classification was consuming massive analyst time:

• Each new client required hours of keyword tagging
• 60 clients meant endless categorization work
• Inconsistency across team members created reporting problems
• High-value analysts spent time on low-value data entry

The real cost: Approximately 50 hours per month per SEO analyst spent on manual keyword classification instead of strategy work.

---

The Solution

Architecture: Rules First, AI Second

┌─────────────────────────────────────────────────────────────┐
│                    KEYWORD INPUT                            │
│              (e.g., "buy nike shoes cheap")                 │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                 DEDUPLICATION LAYER                         │
│  - Normalize whitespace, case                               │
│  - Check canonical cache (already classified?)              │
│  - Skip duplicates within batch                             │
└─────────────────────────────────────────────────────────────┘
                              │
                    ┌─────────┴─────────┐
                    │    Cache Hit?     │
                    └─────────┬─────────┘
                        Yes / │ \ No
                            ▼   ▼
            ┌───────────────┐   ┌───────────────────────────────┐
            │ Return cached │   │      RULES ENGINE             │
            │   category    │   │  - Pattern matching           │
            └───────────────┘   │  - Keyword lists              │
                                │  - Regex rules                │
                                └───────────────────────────────┘
                                              │
                                    ┌─────────┴─────────┐
                                    │   Rule Match?     │
                                    └─────────┬─────────┘
                                        Yes / │ \ No
                                            ▼   ▼
                            ┌───────────────┐   ┌───────────────────────────────┐
                            │ Apply rule    │   │      GPT CLASSIFIER           │
                            │  category     │   │  - Few-shot prompting         │
                            └───────────────┘   │  - Taxonomy from Sheets       │
                                                │  - Batch processing           │
                                                └───────────────────────────────┘
                                                              │
                                                              ▼
                                    ┌─────────────────────────────────────────┐
                                    │          UPDATE CANONICAL CACHE         │
                                    │         (for future batch runs)         │
                                    └─────────────────────────────────────────┘

Why Rules First?

Approach	Cost per 10K keywords	Consistency	Speed
100% GPT	~$5-10	Variable	Slow
Rules only	$0	High	Fast
Rules + GPT fallback	<$1	High	Fast

Most keywords fall into predictable patterns. Rules handle 70-80% of classifications; GPT handles the edge cases.

The Hybrid Approach

1. Rules handle the obvious (70-80% of keywords)

   • “buy nike shoes” → Transactional (rule: “buy *”)
   • “how to do seo” → Informational (rule: “how to *”)

2. AI handles the ambiguous (20-30% of keywords)

   • Only keywords that don’t match rules go to GPT-4
   • Dramatically reduces AI costs

3. Learning compounds

   • Once a keyword is classified, it’s cached
   • New clients benefit from prior classifications

Google Sheets Integration

Why Sheets?

1. SEO team can edit rules without code deployments
2. Taxonomy visible and auditable
3. Version history built-in
4. Collaborative editing

Sheet 1: Taxonomy

Category	Description	Parent
Transactional	Purchase intent keywords	Intent
Informational	Research/learning keywords	Intent
Navigational	Brand/site search keywords	Intent

Sheet 2: Rules

Pattern	Category	Priority
buy *	Transactional	1
how to *	Informational	1
* price	Transactional	2

---

Results

Cost Economics

Before: Manual Classification

Metric	Value
Analyst hourly cost	~$50/hour
Hours per 10,000 keywords	~5 hours
Cost per classification run	~$250

After: Automated Classification

Metric	Value
OpenAI API cost	<$1 per run
Analyst review time	~15 minutes
Total cost per run	<$15

ROI: 94% cost reduction per classification task

Production Metrics

Metric	Value
Time savings	~50 hours/month per employee
Cost per run	<$1 OpenAI API
Test scale	10,000 rows × 3 columns
Users	Entire SEO team

Operational Benefits

Before	After
4-6 hours to classify new client keywords	15 minutes to run and review
Different analysts categorized differently	Single source of truth
More clients = more analyst hours	More clients = same infrastructure
Classification rules locked in code	SEO team updates rules via Sheets

---

Technology Stack

Component	Technology
Compute	Google Cloud Run
LLM	OpenAI GPT-4
Configuration	Google Sheets API
Language	Python

---

Why Not Just Use ChatGPT?

Common question: “Can’t we just paste keywords into ChatGPT?”

Manual ChatGPT	This System
Copy-paste required	Fully automated
Inconsistent formatting	Standardized output
No memory across runs	Canonical cache
$5-10 per large batch	<$1 per batch
No audit trail	Full logging
One person at a time	Team-wide access

---

Lessons Learned

1. Rules beat AI for predictable patterns. 70-80% of keywords follow patterns that simple rules handle faster and cheaper than LLMs. This exemplifies a core principle: workflows with model-powered steps outperform pure agent approaches when the path is predictable and cost-sensitive.

2. Deduplication is a multiplier. Aggressive dedup before API calls dramatically reduces costs and improves consistency.

3. Google Sheets as config layer works. Non-technical team members can update rules without developer involvement.

4. Batch processing is essential. Per-keyword API calls are economically unviable at scale.

5. Start with cost constraints. Designing for <$1/run forced good architecture decisions (dedup, caching, rules-first).

---

Impact

This system changed how the SEO team operates:

• Analysts focus on strategy, not data entry. 50 hours per month per employee shifted from classification to client work.
• New client onboarding dropped from days to minutes. What took a week-long classification sprint now runs while you grab coffee.
• The canonical cache compounds. Every new client benefits from prior classifications—a growing advantage over competitors starting fresh each time.

The ROI paid for the development in the first month.

---

Want to discuss AI classification?

Building a similar system for keyword tagging, content categorization, or document classification? I design cost-controlled AI systems that handle the predictable with rules and the ambiguous with LLMs—all without breaking the budget. Get in touch.