AI Visitor Intelligence
Not all traffic is equal. Averence identifies the type of every visitor — human, AI agent, or noise — and classifies their intent so you know exactly what's engaging with your business.
AI Traffic Looks Nothing Like Human Traffic
Traditional analytics tools were built for humans. They count page views, sessions, and bounce rates. They have no concept of an AI agent making 40 structured API calls in 3 seconds on behalf of a user who asked "what's the best project management tool for a 10-person team?"
Averence's AI Visitor Intelligence layer was built specifically for this new reality — identifying agentic traffic, classifying its intent, and giving you the data to act on it.
Read vs. Act: The Two Types of Agentic Traffic
Averence classifies every agentic request into one of two fundamental categories:
- Read — The agent is gathering information: understanding your offerings, comparing options, building context for a user. These interactions have high research value and should be welcomed and served well.
- Act — The agent is executing a task: booking a demo, initiating a purchase, filling a form. These interactions require appropriate authentication and, optionally, human approval gates.
Understanding which is which lets you respond appropriately — serving Read traffic generously and applying the right controls to Act traffic.
Every Visitor Gets a Grade
In addition to Read/Act classification, every request is assigned a traffic grade that flows into your governance policy:
| Grade | Type | Description | Default Treatment |
|---|---|---|---|
| A | Verified Human | Directly identified human user | Full access, highest priority |
| B | Known AI Agent | Recognised AI assistant (ChatGPT, Gemini, Claude, etc.) | Full access, standard rate limits |
| C | Unverified Bot | Automated traffic of unknown origin or intent | Reduced response fidelity, throttled |
| F | Noise / Scraper | High-volume, low-signal traffic | Rate-limited or soft-blocked |
How Averence Classifies Visitors
Averence's identity model fuses multiple signal layers — no single signal is relied upon alone:
- Protocol signals — Presence and correctness of MCP handshake headers; tool call structure conformance
- Authentication context — Registered agent identities with issued tokens; client attribution logged per call
- Behavioural signals — Query structure, session continuity, and follow-up patterns
- IP & ASN reputation — Cross-referenced against known AI provider egress ranges (OpenAI, Google, Anthropic)
- Request cadence — Human-like inter-request delays vs. machine-speed bursts