EXP-011: Category Traffic Density — AI Frameworks Is a Dead Zone, AI Infrastructure Returns 4.7x More Traffic Per Article Than AI Tools

Experiment ID: EXP-011
Status: COMPLETE
Date: 2026-06-04
Data Window: 2026-04-03 to 2026-06-04 (62 days, full site lifetime)
Owner: Effloow Experiment Lab
Builds on: EXP-009 (word count lever), EXP-010 (search demand variable)

1. Hypothesis

Primary: Content categories are not equal in traffic efficiency. A category with fewer total articles but higher search demand per topic should show measurably higher views-per-article (traffic density) than a high-volume category covering niche or low-demand topics.

Secondary: The current pipeline systematically over-produces low-density categories (AI Tools, AI Frameworks) while under-producing the highest-density category (AI Infrastructure).

Business question: Which category should receive a higher share of pipeline production budget to maximize traffic-per-article published?

2. Data Sources

Measurement date: 2026-06-04
Reference for age calculation: 2026-06-04 (today)

3. Methodology

3.1 Category Corpus Count

Total articles per category were counted by scanning category: frontmatter across all 252 article files:

3.2 Traffic Density Formula

For each category, traffic density was computed as:

traffic_density = sum(top_page_views for articles in category) / total_articles_in_category

This metric answers: "On average, how many monthly views does each article in this category generate, given the entire corpus?"

3.3 Age-Adjusted Views Per Day

To control for indexing time differences across articles, a secondary metric was computed for articles that appear in top_pages:

views_per_day = total_views / days_since_publication

Both metrics are reported to capture different dimensions (absolute ROI vs velocity).

4. Results

4.1 Top-Pages Articles Detail

All 8 articles appearing in top_pages (GA4 data, excluding home and section index pages):

Observation: All 8 top-page articles are from the "legacy" cohort (published before content_track tagging began, 2026-04-04 to 2026-04-21). Zero pipeline articles (sandbox-poc, paper-poc, tool-scout) appear in the top 10 — consistent with EXP-010's finding.

4.2 Category Traffic Density

The key table — views-per-article across the full corpus:

Key finding: AI Infrastructure (2.21 v/article) delivers 4.7x more traffic per article than AI Tools (0.47 v/article). AI Frameworks, with 31 articles, has zero top-page presence.

4.3 Age-Adjusted Performance (Articles in Top-Pages Only)

Note: AI Development's 1.066 avg is significantly skewed by llm-fine-tuning (1.396 v/day), the site's single best performer. Without it, AI Development's average drops to 0.736 v/day — comparable to AI Infrastructure's 0.600.

4.4 Category Representation Rate

"Representation rate" = articles making it into top_pages / total articles in category:

AI Infrastructure's 6.25% representation rate is 4x higher than AI Tools (1.6%).

5. Key Findings

Finding 1: AI Frameworks is a Traffic Dead Zone

31 articles published, zero traffic from search. These articles — primarily covering new framework releases (model updates, SDK announcements) — have near-zero organic search demand. Users searching for these topics either go directly to official docs or don't search at all.

Estimated traffic cost: 31 articles × ~0 views = the same traffic as 0 published articles.

Finding 2: AI Infrastructure Has 4.7x Higher Traffic Density Than AI Tools

With 48 articles, AI Infrastructure generates 106 top-page views (2.21/article). AI Tools has 62 articles but only 29 top-page views (0.47/article). The pipeline currently generates more AI Tools content (tool-scout track) than any other format, but this is the lowest-return category besides AI Frameworks.

Why: AI Infrastructure topics (self-hosting, GPU setup, deployment guides) have high, evergreen search demand. Developers searching for "Ollama setup" or "Hetzner GPU server" have immediate intent. AI Tool reviews typically compete with the tool's own documentation and 100+ other review sites.

Finding 3: AI Development Has the Highest Velocity When Articles Hit

The 1.066 v/day average for AI Development articles in top-pages is the highest of any category — but the LLM fine-tuning article (1.396 v/day) is an outlier. The MCP ecosystem article (0.736 v/day) suggests that broad developer-interest topics within AI Development also perform well at velocity, but AI Development has the most competition (72 articles) and lowest representation rate at that volume.

Finding 4: Automation Category Punches Above Its Weight

7 articles, 1 in top-pages at 4.43 views/article — the highest density. However, the sample is too small (n=7) for generalization. Worth monitoring as more Automation articles are published.

6. Structural Diagnosis

Mapping content_track to likely category:

The structural problem: The three highest-volume tracks (tool-scout, sandbox-poc, paper-poc) map to the categories with the lowest traffic density. Legacy articles — which had no content_track because they were written as standalone, search-demand-first pieces — map to the highest-density categories.

7. Recommendations

R1: Replace tool-scout format with infrastructure guides (HIGHEST IMPACT)

Stop producing generic tool review articles for AI Tools category. Replace the tool-scout track's production quota with AI Infrastructure content: self-hosting guides, GPU setup walkthroughs, deployment tutorials. Target 5 AI Infrastructure articles per week instead of 3 tool reviews.

Expected lift: From 0.47 v/article to 2.21 v/article — a 4.7x improvement in traffic per article published.

R2: Pause AI Frameworks production immediately

AI Frameworks shows 0.00 views/article across 31 published articles. Continue producing AI Frameworks articles only if they include a self-hosting or hands-on component that transforms them into an AI Infrastructure-type piece (e.g., "How to run X framework locally" rather than "X framework announced Y feature").

R3: Increase AI Infrastructure production share to 40%+

Current distribution (estimated pipeline output):

• AI Tools (tool-scout): ~30%
• AI Frameworks (sandbox-poc, paper-poc): ~40%
• AI Development / Infrastructure: ~30%

Target distribution:

• AI Infrastructure: 40% (self-hosting, deployment, GPU setup)
• AI Development (practical guides): 35%
• AI Tools (only when genuine hands-on review): 15%
• AI Frameworks: 10% (only when self-hosting component exists)

R4: Use AI Infrastructure density as the pipeline quality gate

Before publishing any article, ask: "Does this article belong in AI Infrastructure?" If yes, publish immediately. If it belongs in AI Frameworks, only publish if it has a working local setup guide included. If it belongs in AI Tools without a hands-on component, reconsider or expand to infrastructure level.

8. Null Hypothesis Test

The null hypothesis (all categories perform equally when controlling for publication age) is rejected. After controlling for article age using views-per-day:

• AI Infrastructure articles in top-pages average 0.600 v/day
• AI Tools articles in top-pages average 0.659 v/day (single data point)
• AI Frameworks has zero representation in top-pages despite 54+ days of indexing time

The category difference persists after age correction, confirming it is not driven purely by publication timing.

9. Limitations

1. Top-pages sample size: Only 8 articles appear in GA4 top_pages. Category-level conclusions draw from small counts (particularly Automation at n=1, AI Tools at n=1).
2. Confound with legacy cohort: All top-performing articles are from the "legacy" pre-pipeline cohort. It is not possible to fully separate "category effect" from "pipeline vs. legacy" effect — they are correlated by design.
3. Missing sub-category data: "AI Infrastructure" covers diverse sub-topics (cloud setup, local inference, Docker deployment, GPU provisioning). Not all AI Infrastructure topics may perform equally.
4. Single GA4 snapshot: metrics.json top_pages represents a point-in-time snapshot, not a rolling 30-day window normalized per article.

10. Next Experiment Recommendations

• EXP-012: AI Infrastructure sub-topic analysis — which sub-types (self-hosting, cloud GPU, deployment) drive the most traffic within the winning category?
• EXP-013: Controlled pipeline experiment — publish 10 new AI Infrastructure articles using current pipeline agents and measure 30-day traffic vs. matched AI Tools articles published in the same window.
• EXP-014: Title template analysis — within the AI Infrastructure category, do titles containing cost hooks ("$5/month", "free", "zero cost") outperform setup-only titles?