Counting Your Buyers' Attention

The Question

The question is the proportion of a defined buying audience's attention attributable to each voice it engages. The Influence Map ranks voices by how many audience members each one engaged; this paper prices the same graph in the audience's own currency. Each member holds a monthly attention budget; a voice's take is the proportion of that budget it collects. The deliverable is the proportion per voice, with uncertainty, on a stated population edition. The audience of record is the firm's executive audience, defined in committed code and resolved to 179,369 operational members in the 2026Q3 edition.

The Member Layer

The member layer reads dated reaction events as the transactions of a purchase record. A member's record is a count of events across an observation window in months, first counted act month to last, inclusive; exposure is each member's own observed window, so a member observed for one month and a member observed for five years carry comparable rates. A counted member holds at least one dated reaction event: 90,088 target members in the 2026Q3 edition, inside a full engaged corpus of 146,588.

Each member reacts at a steady monthly rate, the rates vary across the audience under a Gamma distribution, and the resulting count model is the negative binomial. The fit is maximum likelihood over a log-spaced parameter grid on the full engaged corpus: r = 0.8145, alpha = 0.6708. A member's estimated rate is the posterior mean (r + x) / (alpha + T), with x its counted events and T its window in months; short records shrink hardest toward the population rate. The attention budget is the sum of these shrunk rates: 105,276 expected reaction events per month for the target audience, 178,662 for the full engaged corpus.

Member layer, edition 2026Q3	Reading
Gamma heterogeneity	r = 0.8145, alpha = 0.6708
median target-member rate	0.72 events per month
target rate quartiles	0.24–1.38; 90th percentile 2.93
median observed record	7 reactions across 9 months
target attention budget	105,276 events per month

The Voice Layer

The voice layer reads the member-to-voice incidence, the record of which members reacted to which authors' posts, as the purchase record of an attention market: 662,343 counted reaction events across 402,448 distinct voices in the 2026Q3 edition. The layer allocates observed events across voices in the Dirichlet-multinomial form of the brand-share literature; a voice's budget share is its events over all target reaction events. 97.4 percent of events name their author; the remainder stays in the denominator, so every share reads conservative. Self-reactions are excluded. Intervals on the 30 largest voices come from a cluster bootstrap of 1,000 replicates at seed 2608, resampling whole members so the intervals respect the clustering of events within a member; a Dirichlet-multinomial posterior interval sits alongside each share for reference.

Ancestry

The member layer is the negative binomial at the heart of Goodhardt, Ehrenberg, and Chatfield's 1984 Dirichlet model of buyer behavior, and the voice layer applies that model's brand-share decomposition to attention. Schmittlein, Morrison, and Colombo's 1987 Pareto/NBD and Fader, Hardie, and Lee's 2005 BG/NBD extend the same count model with dropout inference; the scope section states why this record supports neither. Zhang, Bradlow, and Small's 2015 clumpiness measures cover the within-window burstiness a monthly grain hides. Webster and Ksiazek's 2012 audience-fragmentation work and the attention-centrality literature that followed it measure audience overlap at the outlet level; the voice layer observes the same structure member by member. The components are published. Three searches of the marketing-science and audience-measurement literature, run on 2026-08-11, found no published application joining member-level lifetime-value rate models to voice-level Dirichlet allocation for a defined B2B buying population. Absence of evidence across three searches is the full extent of that claim.

Validation

A purchase-record model owes its literature two empirical regularities. Both run in-model on every edition. Double jeopardy: across the 699 voices holding at least 20 distinct target engagers, log engagers and log events per engager correlate at 0.12. Larger voices also earn more events per engager; the direction matches the law and the magnitude is mild. Duplication of purchase: across the 435 pairs of the top 30 voices, shared audience tracks the product of penetrations at a duplication coefficient D of 4.23; the through-origin fit explains 57.5 percent of the variance.

Ehrenberg test	2026Q3	2026Q2 backcast
double jeopardy correlation	0.12	0.134
duplication coefficient D	4.23	4.45

The outliers carry named structure. williamhgates and kevinolearytv share 19 engagers against about four expected at the fitted coefficient (lift 4.74); williamhgates and rbranson run at 4.34, forbes-magazine and ted-conferences at 3.66. leadership-first and nvidia share zero engagers where the fit expects about six: the inspirational-quote audience and the technology-vendor audience inside this population barely touch. Both regularities reproduced on the first attempt at audience scale and held across editions. Reproducing four decades of empirical law in a new domain is the validation argument.

Scope and Limits

Truncation. Each profile's activity record is its most recent slice, soft-capped near 20 events. Every count in this method is a floor. The cap takes the most events from the highest-rate members and the highest-pull voices, so measured concentration is understated. The truncation operates per member and carries no reference to the voice engaged.

Granularity. Months are the finest grain the record carries. Rates are monthly by construction; burstiness finer than a month is invisible.

Recency. A member's recency gap is the count of months from its last counted act month to its own reference month: the profile's crawl month, or the cutoff month when the cutoff comes first. Active means a gap of three months or fewer. 55.1 percent of target members were active in the 2026Q3 edition.

Dropout. The model fits no dropout process by design. Month granularity plus the slice cap make death-time inference unsound on this record; the recency classification stands in.

The counting rule. A reaction event counts when its month parses and, under a cutoff edition, falls at or before the cutoff. The record admits dated reactions only; the uniform rule drops 100,881 undated reaction acts from every edition identically.

Editions and Audit

Every output file carries three stamps.

Stamp	Definition
editionQuarter	the quarter the edition represents
dataCutoff	the last counted date; null when the edition reads the full record
computedOn	the date the computation ran

A backcast edition is computed from the current snapshot under the counting rule, and its stamps say so: the 2026Q2 edition carries a data cutoff of 2026-06-30 and a computedOn of 2026-08-11. Both editions read the same snapshot, so slice truncation lands on them consistently; the supported comparison is the quarter-over-quarter delta; levels remain floors.

The audience definition lives in one committed module, ops/scripts/claa/audience-definition.mjs. Its term layer holds the 21 title and headline segment patterns; its operational layer resolves membership through the firm's audience Explorer with the spam screen applied, so CLAA membership matches the Influence Map exactly. Frozen segment blocks extend the module: MIDMARKET_2026Q3, frozen 2026-08-11, resolves to 19,435 profiles at companies of 201–10,000 employees and is verified against the Explorer count. The quarterly characterization regenerates from the same module; the 2026Q3 edition gates 294,135 vault profiles to 176,431 term-matched, 94,944 with dated engagement, and 51,807 active within three months of their own reference month.

Every number in a CLAA exhibit is generated by committed code into stamped JSON and SQLite outputs before any prose is written. The model scripts write, alongside their aggregates, a sidecar database of member rates and member-to-voice incidence; a new segment is a query over those tables under a frozen definition block.

Citation

The Influence Map, The Attention Budget, and the segment editions built from frozen definition blocks consume this method; the Mid-Market Edition is the first of the segment editions. An exhibit states its edition, its population, and its findings. For the model, its ancestry, its validation, and its scope, it cites this page.