Measurement Methodology | BAX — Behavioral Biometrics Engine

01 · Foundational principle

Attention is not binary.

A user who scrolls past a brand mention while distracted is not equivalent to a user who pauses, re-reads, and continues reading the surrounding paragraph. Existing digital measurement frameworks treat both as equivalent: the ad was viewable, the impression was served, the metric is recorded.

BAX measures the difference between these two users. It does so through behavioral signals observable in standard web environments without panels, hardware, or self-report: scroll velocity, dwell time per content zone, cursor dynamics, pause patterns, re-read loops, and interaction depth. These signals, in combination, classify each user session into one of seven cognitive segments and produce a continuous attention score for every content unit and brand mention encountered.

The mathematical foundation is exponential decay. Attention diminishes with position in a predictable, modelable way. The rate of that diminution varies by channel. BAX quantifies the rate, calibrates it per channel, and applies it to every metric the platform reports.

02 · The decay model

A dual-exponential function, calibrated per channel.

Attention decay in BAX follows a dual-exponential function that captures two distinct behavioral phenomena: rapid initial decay in the first several positions of a content unit, and a slower, more gradual decline in deeper positions.

Single-exponential models underestimate attention in deep positions for long-form content and overestimate it for short-form social content. The dual-exponential architecture resolves this asymmetry and produces consistent results across channels with fundamentally different content formats.

Decay rates are calibrated per channel. Key finding: web editorial and YouTube long-form produce nearly identical decay rates despite being structurally different media. This cross-channel consistency is the primary validity proof that the decay framework captures a genuine property of human attention rather than a channel-specific artifact.

AI chat responses decay faster than web editorial. Social feeds decay significantly faster than both. A brand mention in the first position of an AI response commands disproportionately more attention than the same brand mentioned further down.

Channel	Decay character	Relative rate
Web editorial	Gradual, sustained	Baseline
YouTube long-form	Gradual, sustained	Near-identical to web editorial
AI responses	Steep, front-loaded	Faster than web editorial
Web news	Moderate-steep	Faster than AI
Web short post	Steep	Significantly faster
Facebook feed	Very steep	4× faster than editorial
Instagram feed	Extreme	6× faster than editorial
TikTok	Extreme	7×+ faster than editorial

Specific decay constants per channel are calibrated on the BBE dataset and available to enterprise clients under NDA.

03 · Three-component decay architecture

Position, freshness, and session depth — applied multiplicatively.

Full attention measurement requires three decay components applied multiplicatively:


        Effective Attention
        EA = φ(d) × ψ(t) × θ(s)

φ(d) — Positional decay. Attention as a function of position within the content unit. Channel-specific calibration. A brand mentioned first in an AI response reaches 77 percent of users attentively; mentioned last, 18 percent.

ψ(t) — Temporal freshness decay. Attention weighting as a function of content age. More recently published content receives higher attention weighting. Critical for AI measurement, where model training cutoffs create systematic freshness penalties for older content.

θ(s) — Session engagement decay. Attention weighting as a function of session depth. Users who have consumed multiple content units prior to encountering a brand mention allocate lower cognitive resources than users early in their session.

The product of all three components represents the actual cognitive engagement a user allocates to a specific brand mention, accounting simultaneously for where in the content the mention appears, how old that content is, and how far into their session the user is.

04 · The seven cognitive segments

Every session, classified within fifteen seconds.

BBE classifies every user session into one of seven cognitive segments within 15 seconds of session initiation.

01Deep Reader

AHI

Sustained scroll velocity, multiple re-read loops, extended dwell time, low interruption rate. Highest attention quality. A brand encountered by a Deep Reader has twice the recall probability of the same brand encountered by a Flow Scroller.

02Active Explorer

AHI

Active navigation, high interaction depth, deliberate content engagement. Together with Deep Reader, constitutes AHI (Attention High Intent), the highest-value attention audience.

03Targeted Scanner

Purposeful movement through content, scanning for specific information. Particularly relevant to query-driven consumption including AI chat responses.

04Flow Scroller

Consistent scroll velocity without significant pauses or re-reads. Moderate attention depth. Benchmark segment for recall comparisons.

05Headline Skimmer

Primary engagement with headlines and lead paragraphs. Brand mentions below the first paragraph have near-zero recall probability for this segment.

06Distracted Browser

Irregular scroll patterns, frequent interruptions, low dwell time. Near-zero attention quality.

07Content Binger

High volume, low depth. Predominant in high-decay social feed environments.

AHI constitutes approximately 18 to 22 percent of measured sessions across editorial environments.

05 · The BAX Index

One score, four dimensions, comparable across channels.


        BAX Index formula
        BAX Index = 0.45 × Decay-weighted Reach
           + 0.25 × Sentiment
           + 0.20 × Exposure Quality (BEI)
           + 0.10 × Accuracy

45%

Decay-weighted Reach

The proportion of users who encountered the brand mention, weighted by position-adjusted attention score. Volume without positional weighting systematically overstates attention value.

25%

Sentiment

Contextual sentiment weighted by cognitive load of surrounding text. Positive sentiment in low-cognitive-load context scores higher than in dense or caveated context.

20%

Exposure Quality (BEI)

The Brand Exposure Index measures context quality: text clarity, source authority, framing relevance, structural prominence. A brand cited from its own authoritative domain scores higher than the same brand cited from a third-party aggregator.

10%

Accuracy

Factual correctness of information AI models associate with the brand. Inaccurate positive mentions score lower than accurate neutral mentions. BAX identifies hallucinations and produces correction packages.

A BAX Index of 72 on AI means the same quality of attention as a BAX Index of 72 on web or video. Cross-channel comparability is a design property, not post-hoc normalization.

06 · Scientific foundation

Four anchors in the published literature.

Reference 01 · Behavioral biometrics ↔ eye-tracking

Behavioral biometric signals as measured by BBE correlate with eye-tracking ground truth at ρ = 0.92 (Anwyl-Irvine et al., 2021). Validates use of behavioral signals as proxy for visual attention without panel recruitment or hardware.

Reference 02 · Attention ↔ recall

Deep Reader encounters produce twice the brand recall probability of Flow Scroller encounters for equivalent content (Dentsu/Lumen, 2023). BAX weights all metrics by segment accordingly.

Reference 03 · Decay model foundation

The dual-exponential decay model is grounded in Ebbinghaus forgetting curve research and extended to positional decay in digital content through the BBE calibration dataset. Kahneman dual-process theory informs segment classification: AHI segments exhibit System 2 engagement; lower segments exhibit System 1 or disengaged characteristics.

Reference 04 · IAB/MRC alignment

BAX methodology aligns with the IAB/MRC Attention Measurement Guidelines (November 2025). BAX operates as a data signal-based measurement system using behavioral biometric signals without physiological hardware, panel recruitment, or survey instruments.

07 · Calibration dataset

Built on nine years of behavioral data.

5M+

URLs across editorial, news, and commercial web content

9 years

Continuous data collection (2017 to 2026)

10,000+

YouTube videos across long-form and short-form formats

300+

AI model variants across 6 major platforms

Fortune 500 campaign validation across financial services and pharmaceutical industries.

08 · Versioning and auditability

Documented, weighted, and reproducible.

This document describes BBE v1.0. Methodology updates are versioned and published. All BAX Index scores reference the methodology version under which they were computed.

Every dimension of the BAX Index is documented, weighted, and auditable. Weighting rationale, segment classification criteria, and decay parameters are available to enterprise clients under the BAX data transparency framework.

NDA notice Specific calibration constants, decay parameters, and proprietary model weights are not published publicly. They are available under NDA to qualified enterprise partners and institutional research collaborators. Contact: [email protected]

The BAX Measurement Methodology