The Retreat

Advertising technology spent the 2010s optimizing one side of the ledger. Every platform built its measurement story around the buy side: impressions delivered, clicks captured, conversions attributed. The terms on earnings calls — Average Revenue Per User (ARPU), CPM, cost per click (CPC), return on ad spend (ROAS) — are all supply-side metrics. Platforms know what they sold. What they decline to measure is what was bought.

What was bought is human attention. For most of the 2010s the industry assumed that attention was roughly stationary — that a thousand impressions delivered to human eyeballs in 2020 bought approximately the same product as a thousand impressions in 2015. Nelson-Field's biometric work, Adelaide's Attention Units, Lumen's eye-tracking panels, TVision's household measurement, and dentsu's Attention Economy series all challenged that assumption by instrumenting the demand side. Their finding was consistent across methodology: per-impression active attention has fallen on nearly every digital channel, with substantial divergence between platforms.

The industry's response has been to treat volume growth as the primary indicator of platform health. But volume and attention are multiplicatively related in the production of an effective impression, not additive. 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 been operating in since at least 2020.

This paper formalizes the gap. It treats advertising inventory as a production function — the IADL framework — and parameterizes every term from primary sources over a decade. The contribution is not theoretical. Tim Hwang's Subprime Attention Crisis (2020), Karen Nelson-Field's The Attention Economy and How Media Works(2020), and Matthew Syrett's “Tragedy of the Advertising Commons” (2004) all anticipated the structural argument. What the model adds is the per-channel decomposition, a decade of longitudinal parameterization, and a forward projection through 2040 with explicit crossing points.

Effective Inventory (EI) is the total useful supply a channel delivers — not just how many ad slots, but how many effective ad slots after accounting for who saw them, whether they paid attention, and whether they were even real. Quality-Adjusted Price (QAP) is the inverse formulation: what advertisers actually pay per unit of effective attention. When QAP rises faster than nominal CPM, the substrate is degrading — and that gap is the diminishing-marginal-returns curve made explicit.

Two methodological notes. The reference anchor for A is out-of-home, set to 1.00. A physical billboard delivers one message per surface; a social feed interleaves many messages on the same unit of attention, and per-message attention falls accordingly. Every digital A coefficient expresses per-impression attention as a fraction of the OOH baseline. D is operationalized at $100K+ household income because that is the convention Pew and most vendor audience surveys publish; the true top-decile spending cohort is directionally steeper than what $100K+ implies.

Every factor in this model is measured by a party that does not sell the ad. That is not a stylistic choice. It is the only epistemically defensible input to a framework that evaluates a platform's product against that platform's own claims about that product.

Supply-side self-reporting is the dominant failure mode in advertising measurement. Platforms define what counts as an impression, what counts as a view, what counts as a conversion, and what the attribution window should be. They also bear the direct financial consequence of those definitions — tighter definitions shrink revenue, looser definitions expand it. Meta settled the DZ Reserve v. Facebookadvertiser class action (N.D. Cal. 4:18-cv-04978) for $37.5M in October 2022, covering a class period of 2014–2019 in which Meta's “Potential Reach” figures to advertisers were alleged to overstate by 200–400%. Inflating reach was the rational response to the incentive structure, not the exception to it. The model therefore treats Meta's audited ARPU as data — it is a SEC-reported 10-K disclosure, cross-checked by auditors with no skin in the advertising game — but treats Meta's “impression delivered” count as an ungroundable claim until a third-party vendor measures it independently.

Retail media is the same trap in a different costume. Amazon, Walmart Connect, Target Roundel and Kroger Precision Marketing pitch closed-loop attribution as a measurement advantage: the retailer sees the impression and the purchase on the same system. That is true, and insufficient. The retailer sees thatthe purchase happened. It does not see whether its ad caused the purchase. A customer who saw a TV spot, received a catalog, read a product review on a newsletter, and then closed the loop on a sponsored listing at the top of an Amazon search page generated demand that every upstream channel contributed to — but the retail media platform, which sees only the last touch, claims the full conversion. This is the same structural failure as branded Google Search credited by GA4: the channel closest to the transaction harvests credit for demand manufactured everywhere else. The measurement is self-reported by a party with financial incentive to overstate causation. That is why this model places retail media in archetype B (volume masking decay) despite the platforms' measurement-advantage marketing.

The generalizable principle: when a supply-side platform reports a metric that shapes pricing of its own product, that number cannot serve as an input to a model that evaluates the product. It can establish that a platform's nominal ARPU grew (audited in 10-Ks). It cannot establish that the impressions were real, that the audience was reachable, that the ads were legitimate, or that the shown ad caused the conversion. Every term in A × D × L in this paper traces to a measurement the platform could not fabricate without cross-vendor collusion.

Meta is the model's reference channel because it is the cleanest longitudinal case: a decade of 10-K disclosure, primary-source biometric attention measurement, a leaked internal scam-revenue projection, and a regulator-published class action on inflated reach metrics. Every other channel's parameterization is calibrated relative to Meta's.

The nominal story: US/Canada ARPU grew from $41.65 in 2015 to $270 in 2025, a 6.5× increase. Meta's North-American user base grew only modestly over the same period; the rest came from extracting more revenue from each existing user — more ads per user, priced higher each.

The quality story runs the other way. The quality multiplier Q = A × D × L fell from 0.523 to 0.135 — a 0.26× contraction. Each term contributed:

The effective ARPU trajectory that results: $21.78 in 2015, $36.45 in 2025. Nominal ARPU rose 6.5×. Real ARPU rose 1.67×. Meta is running 6.5× harder to stand 1.67× in place — hence Figure 1's name.

The per-channel narrative differs, but the template is general. Google Search followed a different mechanism (AI Overviews triggering a 2024–2025 CTR inflection; Seer Interactive measured a 68% paid-CTR collapse on AIO SERPs) but produced the same pattern: nominal CPC rising as effective click quality collapses. TikTok did not exist at 2015 baseline, but its 2020–2025 QAP slope is the steepest in the model at +37.3%/yr. Linear TV is on track to cross the bimodal threshold (QM < 25% of 2015 baseline) around 2028. Every channel tells a version of the same story; Meta is simply the one with the fullest disclosure record.

Two findings sit in those three numbers. First, the status-quo 2025 mix delivers only 1.46× more attention per dollar than the 2015 mix, measured against today's attention factors. That gap reflects diversification away from the Meta-and-Search duopoly rather than any improvement in channel quality; the 2015 mix is worse today mainly because it concentrated 55% of budget in what are now the two most-degraded digital channels.

Second, and more consequentially, the optimal-routing allocation delivers roughly 2× more effective attention per dollar than the 2025 median mix. The gap is not marginal. The single-channel extremes tighten the point: QAP on Meta (2025) is $82; QAP on OOH (2025) is $6.71. Every dollar routed from the first to the second buys roughly 12× more attention, despite the nominal OOH CPM appearing higher on a rate card.

Model this on your own mix via the effective CPM calculator.

Brynjolfsson et al., NBER WP 32846 (August 2024). The highest-profile counter-finding is a nine-year Facebook experiment reporting that long-tenured users derive positive consumer welfare from advertising. The critique this paper makes is selection-based. The sample is by construction the population that stayed on Facebook for nine years. It misses the third of upper-income users Pew documents having left between 2015 and 2025. The population whose ad disutility became large enough to cause exit is absent from a longitudinal study of the stayers. The model's Density-factor collapse (1.00 → 0.58 on Meta) is the shape of the missing population.

Attention vendor disagreement. Adelaide, Lumen, Nelson-Field, and dentsu occasionally diverge by 2–3× on the same channel. CTV is the largest case — TVision per-impression panel measurement vs. Adelaide channel-level aggregate. The model uses a normalized synthesis and flags per-channel disagreements in the methodology page. No single vendor is load-bearing.

Google Search attention is the weakest cell. No attention vendor publishes a Google Search Attention Unit. The 2025 value (0.50) is derived from CTR-collapse data (Seer Interactive's 68% AIO SERP paid-CTR drop), historical SERP eye-tracking baselines (Mediative 2015), and counter-evidence from Sahni & Zhang 2024. Sensitivity range 0.40–0.60. Confidence is graded LOW on this cell specifically.

Pre-2020 fraud and scam data gaps. No channel-native time series exists before 2020 for bot and scam share. The Reuters Meta scam disclosure (November 2025) is the only hard internal number; pre-2022 values are interpolated from FTC Consumer Sentinel proxy data. The direction of travel is not in question; the slope before 2020 is.

Constant-slope projections. Forward projections assume per-factor decay rates remain constant through 2040. The actual DMR curve is likely nonlinear — accelerating as substrate thins. The central scenario is conservative; the accelerating scenario (1.5× observed rate) better captures convexity, and the decelerating scenario (0.5×) captures a hypothetical AI-targeting “save” or regulatory intervention.

Hard factor floors. All factors have a hard floor of 0.05 to prevent compounding to zero in long projections. The floor affects tail-scenario outputs but not the central-case findings.

The retreat this paper proposes is not a rejection of digital advertising. It is capital discipline.

Most channels in the 2025 media plan are delivering less effective attention per dollar than they did a decade ago, even as their nominal prices have risen. The growth-masking-decay pattern is structural to platform economics: a publicly-traded attention business must grow inventory. It cannot choose to deliver fewer, better impressions and still clear quarterly guidance. The degradation is not the failure of any individual platform's management. It is the predictable behavior of the asset class.

The exceptions are the channels whose substrate cannot be manufactured. Physical out-of-home inventory grows at the rate of physical space. Direct mail and catalog inventory grows at the rate of postage economics and deliverable addresses. Host-read podcast inventory grows at the rate of show-host trust. Reader-funded newsletter inventory grows at the rate of paid subscriber relationships. In each case, adding impressions requires something other than software — capital, time, audience relationship. That non-manufacturability is why their quality-adjusted price tracks genuine demand rather than racing to offset substrate decay.

The recommendation follows. A portfolio routing budget away from mature-decay and accelerated-collapse channels and toward archetype A substrates buys several times more effective attention per nominal dollar. Advertisers do not need to predict a platform's collapse to benefit from the repricing; they only need to stop paying the degradation tax on the channels where it is visible now.

That is the retreat. The rest is noise.