AI can suggest clusters, but source data and human review decide the map. A keyword string alone does not reveal exact demand, competition, or intent. Build clusters from owned Search Console data, approved research tools, SERP observation, and business relevance, then validate pages—not just labels.
Prepare the Dataset
Columns:
query
source and date
impressions/clicks if owned
country/device
candidate intent
entity/topic
existing ranking/target URL
business relevance
sensitivity/seasonality notes
Remove or mask personal information and rare user queries that could identify someone before using an AI provider. Do not paste client exports without authorization and provider-data review.
Ask for a Controlled Output
Use an allowed schema:
{
"cluster_id": "service-ac-repair",
"primary_topic": "AC repair service",
"queries": ["..."],
"intent": "local_service",
"recommended_action": "merge|new_page|update|research",
"uncertainty": "..."
}
The model must not invent search volume or difficulty. Use deterministic Sheets formulas for normalization and duplicates; use AI for semantic suggestions. Compare sample queries in live search results carefully by country/language and date to see whether the same page types satisfy them.
Map One Page per Need
Several phrasings can share a page when they represent the same task. Separate pages when intent, product/service, location with real presence, or required content differs materially. Avoid thin page-per-keyword or fake city variants.
Worked Example
A Lahore home-service site has Search Console queries around AC repair, installation, and general cleaning. AI groups them, but a human sees installation results require different proof, pricing logic, and FAQs. It becomes a separate service page. “AC repair Lahore” and “air conditioner repair Lahore” share one page.
The sheet maps current URLs and flags cannibalization: two old repair posts target the same task. The plan consolidates useful material and redirects only after link/traffic review. It does not create 40 neighborhood pages without distinct service evidence.
Failure Cases to Diagnose
- Model invents volume/difficulty: retain only sourced metrics.
- Every cluster becomes new page: update/merge where appropriate.
- Queries contain private details: aggregate/redact before AI.
- One model run is final taxonomy: sample and review.
- SERP observed without date/location: record context.
- Existing URLs ignored: map consolidation and internal links.
🇵🇰 Pakistan Angle
Pakistani search phrasing can mix English, Urdu script, and Roman Urdu. Cluster by task, not merely language string. Validate whether users expect a service page, guide, marketplace, map/local result, video, or official source.
Do not create city pages for places the business does not serve. Where local terms differ—marla/kanal, COD, board/exam names—use verified terminology and useful content rather than stuffing variants.
Hands-On Exercise
- Export an authorized query sample or create synthetic data.
- normalize/deduplicate in Sheets.
- generate schema-bound cluster suggestions.
- validate a sample against results and current URLs.
- produce update/merge/new/research actions.
Completion Rubric
- Metrics retain source/date/context.
- AI does not invent numeric opportunity.
- Private queries are protected.
- Clusters map to distinct user tasks.
- Existing URLs/cannibalization are handled.
- Local pages match real service coverage.
Sources
- Search Console Help — Performance report
- Google Search Central — Helpful content
- Google Sheets Help — Data tools
Key takeaway: let AI propose semantic groups, but keep metrics sourced, queries protected, intent human-validated, and page creation tied to a genuinely distinct task.