Machine Attention · Methodology

Make the evidence inspectable.

Start with proven sponsor delivery. Preserve source boundaries, publish what can be measured, and grade causal claims separately from observed links.

Current public report: descriptive and observational. No randomised treatment/control result or causal lift is claimed here.

Metric contract: 2026-09-11.v1

Current collection boundary

Configuration is not an event.

What an exposure must establish

A known campaign and placement, qualifying machine response, actual sponsor emission, response status and UTC timestamp. Cache delivery needs appropriate origin or edge evidence.

What discovery must establish

A campaign-specific advertiser arrival after qualifying exposure, within a declared window. Direct actor matching and campaign-only evidence receive different observation grades.

Window: Rolling 30 days; see partial observation coverage. Measurement mode: origin-verified; partial observation coverage.

Counting and attribution

Keep each claim within its evidence.

VME and VMD are separate event measures. MDR uses the same campaign cohort, period and compatible event/actor units.
TTD uses direct linked timestamps. Publish sample count and median, p25 and p75 only when the declared sample rule is met.
Post-discovery depth reports reads, unique URLs, content classes, duration and strong followbacks. MCE requires a declared comparison; a missing baseline is not a zero baseline.
ACI requires separately observed AI answers and governed scoring. It is not inferred from crawler activity.
HAR-V/HAR-P and AIC-V/AIC-P stay separate. General advertiser outcomes are not silently attributed to a sponsor.
Bot family identification does not prove exposure of a specific consumer model, an answer citation or a human decision.
Unreconciled properties, raw requests, served reads and API calls are not added into the evidence chain.

Before a controlled study begins

Preregister the comparison.

The following are publication requirements, not a claim that an experiment is active. The design must be frozen and hashed before treatment/holdout collection begins.

Question and outcomes

Freeze the hypothesis and metric contract.
Primary outcome: campaign-attributed VMD rate with an explicit denominator.
Declare secondary outcomes, minimum samples and uncertainty method.
Define start, end and stopping rules.

Assignment and exclusions

Declare eligible campaign versions, properties and placements.
Choose a randomisation unit that limits crawler contamination.
Publish assignment method, holdout share and attribution windows.
Declare machine classes, test exclusions and contamination checks.

Not registered

No controlled preregistration is published in this report

Assignment unit, holdout share, stopping rule, minimum sample and frozen document hash remain unset. Causal grade stays C0 until actual evidence supports another grade.

First-party evidence

No public actor identifiers.

Public evidence uses safe aggregates and published event descriptions. Raw IPs and machine pseudonyms are excluded. Machine matching must not identify the person behind a request or create a cross-site human fingerprint.

This report reads public aggregates; it does not create a new raw-event retention policy. Any future event bridge must publish its machine-only matching fields, identifier rotation, retention windows, rollups and deletion rules before those claims appear here.

What the evidence permits

Findings, limits and corrections.

What we can say

The report publishes the measurement state of each stage and distinguishes sponsorship configuration from proven exposure and discovery. Current observation and causal grades describe the evidence available.

What remains unproven

Delivery does not prove that a machine used the sponsor in an answer. A crawler visit does not establish a human referral. Timing alone cannot establish a sale was caused by sponsorship.

Known measurement gaps

Accepted VME proves a qualifying application emission, not receipt by a model or reading.
All rolling windows preserve their exact requested bounds; collection coverage and unregistered estate properties remain explicit.
Known audit events remain in their original ledger and are excluded by a public dated correction.
No verified edge proof or campaign-linked arrival, session or sales outcome evidence is connected. TEAS receiver readiness is recorded separately from source activation and outcomes.
Late arrivals and later absolute bucket revisions can change provisional reports.
No causal lift or AI-mediated sales are inferred.

Corrections ledger

Public corrections
Date · UTC	Affected version	Correction and effect
2026-09-11	2026-09-11.v1	Initial public contract. Existing sponsor configuration and earlier API/citation ledgers are not retroactively counted as Verified Machine Exposure.
Not measured	Not measured	Not measured
Not measured	Not measured	Not measured