NEUROSTACK
Back to blog
Engineering|8 min read|March 2026

How We Built a 14-Agent Marketing Analysis System

When we set out to build Neurostack Marketing Intelligence, we had a specific thesis: marketing analysis is too complex for a single model call. Instead of asking one AI to evaluate everything, we designed 14 specialized agents — each an expert in one dimension.

The Architecture Problem

A typical website audit touches content quality, technical SEO, conversion optimization, accessibility, security, and more. These are fundamentally different analysis domains. A content expert looks for messaging clarity and readability. A security analyst checks HTTP headers and cookie policies. Asking one model to do both produces shallow results.

Parallel Agent Design

Our solution: 14 agents running in parallel, each with its own prompt engineering, scoring rubric, and domain-specific checks. The agents are split into two tiers — 6 Core agents (accounting for 68% of the total score weight) and 8 Specialized agents (32% weight).

Core agents cover the fundamentals that every website needs: Content & Messaging (15%), Growth & Strategy (13%), Conversion Optimization (12%), Technical SEO (12%), Competitive Positioning (8%), and Core Web Vitals (8%). Specialized agents handle deeper dimensions like Accessibility, Schema Markup, AI Visibility (AEO), and Privacy Compliance.

Agent Breakdown
Core Agents — 68% Weight
edit_note
Content & Messaging
15%
trending_up
Growth & Strategy
13%
ads_click
Conversion Optimization
12%
code
Technical SEO
12%
military_tech
Competitive Positioning
8%
speed
Core Web Vitals
8%
Specialized Agents — 32% Weight
accessibility_new
Accessibility
5%
data_object
Schema Markup
5%
visibility
AI Visibility (AEO)
4%
shield
Privacy Compliance
4%
palette
UX & Design
4%
share
Social & Sharing
4%
lock
Security Headers
3%
translate
Internationalization
3%

LLM-Agnostic Gateway

We built an LLM gateway that supports multiple providers — Gemini, Groq, OpenAI, and Claude — with automatic failover. If one provider is slow or down, the system routes to the next available option. This gives us reliability without vendor lock-in, and lets us optimize for cost vs. quality per agent.

Weighted Scoring

Each agent produces a score from 0-100 for its dimension. The final site score is a weighted average where core agents contribute more than specialized ones. This ensures that fundamental issues (broken CTAs, missing meta tags) have more impact than edge cases (missing schema markup), reflecting real-world priorities.

Sample Score Breakdown
76/100
Content & Messaging
15% weight82
Growth & Strategy
13% weight65
Conversion Optimization
12% weight91
Technical SEO
12% weight74
Competitive Positioning
8% weight58
Core Web Vitals
8% weight88
76/100
Weighted Total
127
Checks Run
58s
Time Elapsed

Results

The result is an audit that completes in about 60 seconds, produces 100+ individual checks, and generates a 10+ page PDF report with prioritized, actionable recommendations. Because each agent is independent, we can add new dimensions (like AEO and Privacy) without touching existing agents.