Blind spot 01
Referrals get scattered
Traffic sent by AI assistants lands inconsistently across “referral” and “direct” — the same journey classified differently across sites, times and tools.
The problem
Three blind spots
The analytics stack — referrer headers, UTM parameters, session models — assumes a human operating a browser. AI-mediated discovery breaks that assumption in three ways, and current conventions misclassify all three.
Blind spot 02
Agents look like people
AI agents now read pricing and product pages on a buyer’s behalf. These sessions are neither ordinary bots nor humans — and get miscounted as human traffic.
Blind spot 03
Influence hides in “direct”
The biggest effect: a buyer gets an AI recommendation, then arrives via branded search or direct — and the revenue is credited to brand strength, not the AI.
The consequence is that any attribution limited to deterministic signals measures a visible minority of AI’s effect and omits the majority.
The convention
A shared vocabulary: three classes
ATAC does three deliberately limited things: it defines a common vocabulary so “AI traffic” figures are comparable, a way to declare provenance rather than infer it, and honesty rules for reporting. The vocabulary is three traffic classes.
Class A
AI-Referred
A human followed a link from an AI assistant. Deterministic — identifiable from the request itself (a declared tag or a classified referrer).
Class B
Agentic
An automated AI system visited directly — reading content for a live user or acting on their behalf. Detected, ideally cryptographically verified.
Class C
AI-Influenced
Revenue shaped by an AI recommendation with no click trail — the large, currently-invisible one. An estimate, never an observation.
the hidden majorityEvery class assignment carries an explicit confidence value, and you can’t blend them without labelling the mix:
verified · cryptographically authenticateddeclared · stated by the platformhigh · exact rule matchmedium · tested inferencelow · heuristic only
How provenance gets declared
Make it explicit, not inferred
The heart of the convention is a way for AI platforms to declare where traffic came from, instead of leaving sites to guess. It’s a single lightweight query parameter, ai_ref, that survives referrer stripping and needs no header cooperation:
https://yourbrand.com/pricing?ai_ref=chatgpt:chat:assistedIt carries the platform, the surface (chat, search, agent, api, voice) and the session class (assisted = a human following an AI link; delegated = an agent acting for them) — with an open, free platform registry. It declares provenance, never identity: no tracking, no fingerprinting, no personal data.
ATAC also extends brand.context with an agentAccess object, so a site can tell agents where its canonical, machine-readable commercial facts live — before an agent mis-parses a JavaScript-rendered pricing page and the recommendation is shaped by absent data.
The honesty rules
The real risk isn’t bad data — it’s false certainty
The inferred class (AI-Influenced Revenue) is where numbers are easiest to manufacture, so ATAC is strictest there. A Class C figure is a modelled quantity, and the convention requires it to be published as a range with a confidence statement — never a single point estimate — derived using at least one input independent of the traffic it’s explaining, with the model, window and assumptions all disclosed.
“The central integrity risk in this domain is not false data but false certainty.”
That independent input is a genuine, multi-turn, replicated, multi-system recommendation measurement — not a citation count or a share-of-voice score, which measure mention, not decision. It’s a deliberately high bar, set exactly where the temptation to fabricate precision is greatest.
An open invitation
A draft, offered to the industry
ATAC is published as a v0.1 draft for public comment — free to adopt, requiring no licence, product or relationship with AIVO. It’s positioned as a layer within the stack, not a competitor to any part of it: it complements robots.txt, consumes Web Bot Auth signatures as its highest confidence tier, and draws its lineage from the Media Rating Council’s invalid-traffic framework and the IAB’s agentic-advertising work — addressing the organic discovery-and-attribution layer those don’t yet cover.
Comments are invited from all parties. The convention, its platform registry and a public change log are maintained openly, with each version deposited on Zenodo — this is a convening move, not a closed standard.
The full draft
Read the complete convention
This article is an overview. The full v0.1 draft defines the taxonomy and subclasses, the ai_ref syntax and processing rules, the agent-access extension, the conformance classes, and the security and integrity considerations — published open-access on Zenodo, with public comment invited.
Citation: de Rosen, T. & Sheals, P. (2026). AI Traffic Attribution Convention (ATAC), Version 0.1 — Draft for Public Comment. AIVO Standard. Zenodo. https://doi.org/10.5281/zenodo.21399490 · Licensed CC-BY-4.0.