Theory No. 02
working framework · 2026
The Retreat: the economics of attention
Digital ads have never cost more. Or been worth less. A cross-channel model of the diminishing marginal returns of advertising attention: fourteen channels, a decade of primary-source data, projected to 2040. Published as a full interactive model and whitepaper.
the claim
Volume is masking substrate decay
Between 2015 and 2025 the price of advertising on every major digital channel rose while the underlying product, effective human attention, collapsed. Platforms report the supply side: impressions delivered, clicks captured, conversions attributed. What they decline to measure is what was actually bought. Per-impression attention has fallen on nearly every digital channel, the audiences seeing the ads have less purchasing power than they did, and a growing share of impressions go to bots or are themselves scams.
A platform that doubles its impression volume while halving attention per impression delivers the same effective inventory at twice the nominal price. That is the regime nearly every major digital channel has operated in since at least 2020. Most channels that appear to be growing are retreating; volume growth is hiding it.
the model
Every impression decomposes into four factors
The model treats advertising inventory as a production function. Volume is what platforms sell; the other three factors are what advertisers actually buy.
Effective Inventory = Impressions × Attention × Density × Legit Share
Quality-Adjusted Price = Nominal CPM ÷ (Attention × Density × Legit Share)
I — Impressions
The volume term, and the only one platforms headline. Meta delivers 7x more impressions than a decade ago. Volume disguises decay.
A — Attention
Seconds of active human attention per impression, anchored to out-of-home = 1.00 via independent biometric and eye-tracking vendors. Premium video ~13.5s. A social feed ad ~2s.
D — Density
The fraction of the audience with real purchasing power ($100K+ household income, via Pew). An impression shown to someone who cannot buy the product is worth less.
L — Legit share
(1 − bot fraud) × (1 − scam share). Independent vendors measured 21% invalid traffic on paid social; Meta internally projected ~10% of 2024 ad revenue came from scam ads.
A methodological rule runs through the whole model: every quality factor is measured by a party that does not sell the ad. Independent attention panels, census-weighted demographics, adversarial fraud vendors, audited financial filings. A thesis about whether advertisers get what they pay for cannot be built on the seller's own claims about what was delivered. It's the same discipline as the attribution warehouse: don't trust the number reported by the party it pays.
the reference case
Meta: running to stand still
Meta is the model's cleanest longitudinal case: a decade of audited 10-K disclosure, independent biometric attention measurement, and a leaked internal scam-revenue projection. Nominal revenue per US/Canada user grew 6.5x over the decade ($41.65 to $270). Over the same period the quality multiplier collapsed to 0.26x of its 2015 value: per-ad attention fell to ~2 seconds (below the memory-formation threshold), upper-income daily usage fell from 78% to 45%, and the legitimate share of impressions dropped as bot and scam load grew.
Multiply it out and real, quality-adjusted revenue per user grew just 1.67x against the 6.5x nominal rise. Meta is running 6.5x harder to stand 1.67x in place, and advertisers are paying the difference. The pattern generalizes: Google Search followed the same shape through a different mechanism (AI Overviews collapsed paid click-through ~68% on affected results), and TikTok's real price is inflating at +37% per year, the steepest in the model.
the taxonomy
Four archetypes, one retreat
Classified by trajectory rather than category, all fourteen channels land in one of four states. The only resilient class is the one whose substrate cannot be manufactured by adding more impressions: physical out-of-home, direct mail and catalogs, host-read podcasts, reader-funded newsletters. Surface growth there is constrained by capital, postage, or trust, so price tracks genuine demand instead of substrate decay.
A · Structural resilience
Substrate cannot be manufactured. OOH, direct mail, host-read podcasts, newsletters. Not projected to cross any decay threshold through 2040.
B · Volume masking decay
Inventory expanding faster than quality degrades; the slope inverts when surface expansion exhausts. CTV and retail media.
C · Mature decay
Every quality factor trending negative while nominal price holds through advertiser path-dependence. Meta, Google Search, LinkedIn.
D · Accelerated collapse
Structurally finished or collapsing before maturity. Linear TV (projected to cross the bimodal threshold ~2028), display, TikTok.
The portfolio implication is the point. Measured against today's attention factors, an allocation concentrated in archetype-A channels delivers roughly an order of magnitude lower quality-adjusted price than the median 2025 media mix. The conclusion is capital discipline, and the name of the paper: not a rejection of digital advertising, but a retreat toward channels where the quantity advertisers pay for is still the quantity delivered. A public attention platform cannot make that choice for you; it must grow inventory to clear quarterly guidance, even when every marginal impression is worth less.
the full model
Explore the live model: five interactive figures, an effective-CPM calculator, the methodology, and the whitepaper