Methodology

The DMR Model

The diminishing marginal returns (DMR) model decomposes every ad impression into four multiplicative factors and parameterizes each from primary sources. This page documents the equation, the four factors, the per-channel parameterization, the projection methodology, the limitations, and the prior literature.

The Equation

Effective Inventory = Impressions × Attention × Density × Legit Share

Quality-Adjusted Price = Nominal CPM ÷ (Attention × Density × Legit Share)

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. 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 — that gap is the DMR curve made explicit.

Falsifier

This model is falsifiable per channel. A rising nominal CPM is not evidence of degradation on its own — it is only evidence of degradation when the quality multiplier A × D × L is measured to be falling underneath.

Passes the test (rising CPM + holding quality)

Direct mail (A × D × L +32%), out-of-home (+8%), host-read podcasts (roughly flat at −8%). Rising CPMs on these channels reflect genuine demand, not substrate decay. Newsletters also sit in archetype A by substrate (reader-funded scarcity), but per-impression quality has drifted ~−20% as the category expanded 7.5× in volume — a nuanced case.

Fails the test (falling quality, price rising or compressing)

Meta (−74%), Google Search (−58%), CTV (−63%), LinkedIn (−38%), TikTok (−40% from 2020 baseline), Linear TV (−66%). On Meta, Search, and LinkedIn, CPMs are rising; on CTV, TikTok, and Linear TV, nominal CPMs are actually compressing while quality falls faster. Either way, the substrate is degrading.

Anyone producing an independently-measured quality multiplier for any channel in the “fails” column that holds flat or rises from 2015 to 2025 invalidates the degradation claim for that channel. Each A, D, and L value in this model traces to a party that does not sell the ad: Nelson-Field, Adelaide, Lumen, Pew, fraud0, DoubleVerify. If those measurements are wrong, the thesis is wrong — and the correction path is specified.

The Four Factors

I

Impressions

Formal

Total monetizable ad slots delivered by the channel per year.

Measurement

Channel-native disclosed metric — paid clicks for Search, ad impressions for Meta, GRPs for TV, listens for podcasts, OAAA-reported revenue for OOH.

Why it matters

The volume term. What platforms sell. Most channels show massive growth in I — Meta has 7× more impressions in 2025 than 2015. Volume disguises decay; almost every collapsing channel looks like it's growing on this metric alone.

▸5 sources
A

Attention

Formal

The fraction of nominal impressions that receive sufficient human attention to produce a downstream behavioral effect, anchored to OOH = 1.00 (waste-adjusted reference).

Measurement

Cross-vendor synthesis: Adelaide Attention Units (composite ML score from 11M+ monthly placements), Nelson-Field biometric panels (115K ad views, 60 formats, 12 countries), Lumen eye-tracking (11.8s on a 30s TV ad), TVision panel measurement, dentsu Attention Economy.

Why it matters

Per-impression cognitive engagement. Premium video gets ~13.5s of active attention. Facebook feed gets ~2s. TikTok ~3.5s. OOH ~12s. Display under 1s. Vendors sometimes disagree by 2-3x on the same channel — sensitivity ranges flagged in the per-channel notes.

▸6 sources
D

Density

Formal

The fraction of the channel's audience in the upper-income cohort with discretionary spending capacity. Operationalized as $100K+ household income because that's the convention Pew and most audience-survey vendors publish.

Measurement

Pew Research Center demographic crosstabs by platform, GWI Global Web Index, Nielsen audience composition, channel-specific demographic studies.

Why it matters

An impression delivered to a low-income audience is worth less than the same impression delivered to a higher-income one because the recipient can't actually purchase most premium products. Pew documented Facebook's daily upper-income usage falling from 78% in 2015 to 45% in 2025 — the platform inverted from 'Harvard exclusive' to downmarket. Same dynamic on TikTok (only 12% of users earn $100K+) and CTV ad-tiers (28% $100K+ vs 39% on ad-free tier). Note that $100K+ is roughly the top 33% of US households, not the top decile — the true K-shape Group A cohort starts around $200K. We use $100K as a proxy because it's the convention Pew measures at; the trend at $100K is directionally correlated with the (steeper) trend at $200K+.

▸6 sources
L

Legit Share

Formal

(1 − bot fraud) × (1 − scam ad share) × (1 − inflated metric penalty). The fraction of impressions that are real humans seeing real, brand-safe ads.

Measurement

fraud0, DoubleVerify, Adalytics, IAS bot/IVT data; FTC Consumer Sentinel social-media-originated fraud losses; Reuters investigative reporting on Meta internal scam-revenue projections; class action filings on metric inflation.

Why it matters

An impression to a bot is worth nothing. An impression where the 'ad' is a scam product is worth less than nothing — the user's eventual encounter with the scam taxes the platform's brand alongside the advertiser's. Meta's internal documents (Reuters Nov 2025) project ~10% of 2024 ad revenue came from scam ads, ~$16B. fraud0 Q2 2025 measures 21% paid social IVT. DoubleVerify saw GIVT jump 86% YoY in H2 2024.

▸7 sources

Per-Channel Sources

14 channels, every primary source linked. Click any channel to expand.

Meta (Facebook + Instagram)MED-HIGH
▸

Reference channel for the model. US/CA ARPU $41.65 (FY 2015) → $226.93 (FY 2023) per 10-K. Quality multiplier collapsed from 0.523 to 0.134 (-74%) after applying Nelson-Field attention coefficient, Pew density inversion, and Reuters scam disclosure.

Primary Sources

LinkedInMED-LOW
▸

The premium-density anchor of the B2B ad market. Pew 2015 (at $75K+ threshold) shows ~52% of LinkedIn users; Pew 2025 (at $100K+ threshold) shows ~49% — different cutoffs, but LinkedIn resisted the Meta-style downmarket inversion (Meta 78% → 45%). D factor is the highest in the model. But A and L are degrading with the rest of digital: MediaScience 2024 measures 3.7s active attention on B2B video ads (study commissioned by LinkedIn's B2B Institute, independent methodology); LinkedIn's Community Report counts 200M+ fake account takedowns in 2024 (self-reported — used as directional signal only; fraud0's 21% paid-social IVT is the third-party anchor for L). Quality multiplier fell 38% vs. Meta's 74%. Confidence MED-LOW because no independent vendor publishes a LinkedIn Attention Unit; post-2016 US Marketing Solutions revenue is not audited (Microsoft consolidated LinkedIn into Productivity & Business Processes segment after the 2016 acquisition); US revenue estimates apply the 2015 10-K's 65% US geographic share to 2020/2025 global Marketing Solutions figures — that 65% ratio is an assumption, not a disclosed post-acquisition split.

Primary Sources

Google SearchMED
▸

QAP CAGR +24%/yr on the real per-impression CPM (CPC × CTR × 1000). Effective CPM rose $44 → $158 (3.6×) while quality multiplier fell 58% — driven by AI Overviews. Seer (Sept 2025) measured 68% paid-CTR collapse on AIO SERPs. Attention coefficient is the weakest cell in the model — no attention vendor measures Search.

Primary Sources

YouTubeMED-HIGH
▸

Slowest-decaying digital giant. Per-impression quality declined to ~0.70x of 2015 baseline by 2025. Premium tier (125M subscribers by Mar 2025) functions as Tåg 2009 escape valve — high-WTP users pay to leave rather than tolerate ads.

Primary Sources

TikTokMED-HIGH
▸

Steepest QAP slope in the model: +37.3%/yr (5-yr CAGR vs. 2020 baseline, since the channel did not exist US-side in 2015). Adelaide AU 21.1 is the lowest of any major social platform; Pew shows only 12% of TikTok users earn over $100K. CPMs falling 30% YoY in Q1 2025 even as quality collapses faster.

Primary Sources

Linear TV (Broadcast + Cable)HIGH
▸

Approaching bimodal (QM at 34% of 2015 baseline in 2025 — projected cross 2028 on central-case decay). Total TV ad viewing share fell to 12.6% in Q1 2025 (Wieser). Sports concentration: 95 of top 100 broadcasts, NFL alone 84. The non-sports tier is approaching zero-utility.

Primary Sources

CTV / Streaming Ad TiersHIGH
▸

Marketed as the premium alternative to social. Looks healthy on volume (Amazon Prime Video added ~50B impressions in Jan 2024). But Pixalate Q4 2024: CTV IVT rose to 24% (Samsung 31%). TVision: per-ad attention dropped 29.6% — first decline ever measured. Netflix CPMs $54 → $31 (-43%).

Primary Sources

Podcasts (Host-Read)MED-HIGH
▸

The ONLY channel in the model where the attention coefficient hasn't moved at all in a decade. 71% brand recall (Nielsen), 86% among heavy listeners (Sounds Profitable n=5,000), 84-94% listen-through rate. Host-read still 62% of podcast ad revenue per IAB FY2023.

Primary Sources

Podcasts (DAI / Programmatic)MED
▸

Quality multiplier collapsed 41% — comparable to Meta. Volume-build is masking the quality decay (impressions grew ~55x from near-zero 2015 base). Magellan AI Q3 2025: ad load 8.34%, approaching the 10% behavioral breaking point.

Primary Sources

Newsletters / Substack SponsorshipsMED
▸

Reader-funded. The only channel in the model where CPM inflation is genuinely demand-driven, not quality-collapse-driven. 1.95x top-decile audience density. Substack paid subscriptions 250K (2020) → 5M+ (March 2025). 5-year basis is more honest — 2015 was effectively pre-channel.

Primary Sources

Out of HomeHIGH
▸

The model's reference channel (A = 1.00). Physical substrate cannot be contaminated by prior impression. QMS × Amplified Sydney study (1.23M observations): 12 seconds total attention per impression, 90% of sites exceed the 2.5s memory-formation threshold. Near-zero fraud by design (no programmatic vector). 19 consecutive quarters of YoY revenue growth.

Primary Sources

Direct Mail + CatalogsMED-HIGH
▸

Physical substrate. USPS Marketing Mail volume 80.1B (FY2015) → 57.5B (FY2024) as First-Class postage rose ~60% — the volume decline is pricing, not per-piece quality decay. JICMAIL (PwC-vetted UK panel) measures 108s of attention per DM item over 28 days. ANA 2025 Response Rate Report: average 4.4% response, 161% ROI on house-list DM — highest of any paid channel measured. Winterberry 2024: $37.3B US DM spend, +2.6% YoY; 80% of brands plan to grow DM in 2025. Catalog renaissance — IKEA, Sephora, Amazon, Old Navy relaunching or expanding print.

Primary Sources

Retail Media NetworksMED
▸

Currently the fastest-growing channel (+41.7%/yr EI). Volume-build phase. ACOS crossed 30% → 32.5% (Jan 2026, highest recorded). 70% of Amazon sellers now advertise (up from 40%). Slope inversion projected ~2028-2029 (85-90% confidence) when sponsored slot density hits ceiling. Walmart Connect +51% YoY CPC in Q4 2024 was the early signal.

Primary Sources

Display ProgrammaticHIGH
▸

The floor of the model. Adelaide AU 22.09 (vs CTV 69.53). Lumen: '30 display impressions = 1 TV ad' on attention basis. Profit Ability 2: £3 ROI per £ spent (lowest of all channels measured). aCPM $9.70 — most expensive per unit of attention. fraud0 21% onsite IVT.

Primary Sources

Projections

Forward projections (2025-2040) use constant-slope extrapolation per quality factor. For each channel and each factor (A, D, L, I), the 2015-2025 annualized rate of change is computed and applied recursively to project the 2025 value forward.

Three scenarios are computed but only the central case is shown by default:

  • Central: exact observed 2015-2025 rate.
  • Accelerating: 1.5× the observed rate (captures DMR convexity as substrate thins).
  • Decelerating: 0.5× the observed rate (captures AI-targeting "save" or regulatory intervention).

All factors have a hard floor of 0.05 to prevent compounding to zero. Newsletters and Retail Media are excluded from trajectory plots because their near-zero 2015 baselines create unstable ratios.

Limitations

Google Search attention coefficient is the model's weakest cell. No attention vendor publishes a Google Search AU. The 0.50x value (2025) is derived from CTR collapse data (Seer Interactive 68% AIO drop), historical SERP eye-tracking (Mediative 1.17s/listing baseline), and counter-evidence from Sahni & Zhang 2024 (users prefer more search ads). Sensitivity range 0.40-0.60. Confidence: LOW.

Attention vendor disagreement. Adelaide, Lumen, Nelson-Field, and dentsu sometimes diverge by 2-3× on the same channel. The model uses a normalized synthesis but flagged disagreements remain in the underlying notes. CTV is the largest disagreement: TVision per-impression measurement vs Adelaide channel-level aggregate.

2015 baselines for newer channels. TikTok did not exist US-side in 2015. Newsletters and Retail Media had near-zero 2015 ad markets. These channels use 2020 baselines or are excluded from trajectory comparisons.

Pre-2020 fraud and scam data. No channel-native bot or scam time series exists pre-2020. The Reuters Meta scam disclosure (Nov 2025) is the only hard internal number; pre-2022 values for Meta scam share are interpolated from FTC Consumer Sentinel proxy data.

Constant-slope projections. The forward projection assumes per-factor decay rates remain constant. The actual DMR curve is likely nonlinear (accelerating as substrate thins). Central scenario is conservative; accelerating scenario better captures convexity.

Influences

The thinkers and research this model builds on or argues with.

Karen Nelson-Field — Amplified Intelligence

'Cost of Dull Media' (Cannes 2025), 'The Attention Economy and How Media Works' (2020), 'The Attention Playbook' (2023)

The biometric attention measurement methodology that the model's A coefficient ultimately rests on. 'Slow-decay vs fast-decay' environment classification across 60 formats in 12 countries.

https://www.amplified.co/

Tim Hwang — Subprime Attention Crisis (FSG, 2020)

Argued programmatic ad inventory is structurally like subprime mortgages: inflated, fraudulent, about to collapse.

The structural-collapse thesis ancestor. The model parameterizes Hwang's qualitative argument with post-2023 data he didn't have access to.

https://us.macmillan.com/books/9780374538651/subprimeattentioncrisis/

Matthew Syrett — 'The Tragedy of the Advertising Commons' (MarketingProfs, Sept 2004)

Applied Hardin's tragedy of the commons to advertising attention 22 years before this model.

The rhetorical ancestor of the structural argument. Cited for intellectual honesty.

https://www.marketingprofs.com/4/syrett5.asp

Brian Wieser — Madison and Wall

Quarterly TV inventory and ad market commentary.

The Linear TV 'permanently declining inventory' framework. The Q1 2025 12.6% TV viewing share figure that anchors the Linear TV row in the model.

https://madisonandwall.com/

John Burn-Murdoch — Financial Times

FT social media decay series, GWI time-spent analysis (2024-2025).

The audience-side withdrawal evidence (250K adults across 50+ countries). The visual design inspiration for the trajectory chart's small-multiples treatment.

https://www.ft.com/john-burn-murdoch

Brynjolfsson, Collis, Liaqat, Kutzman, Garro, Deisenroth, Wernerfelt — NBER WP 32846 (Aug 2024)

'The Consumer Welfare Effects of Online Ads: Evidence from a 9-Year Experiment'

The paper this model's framing argues with — its 'no disutility from ads' finding rests on a sample of users who survived 9 years on Facebook, missing the audience that already left. Selection-effect critique parameterized by Pew demographic data.

https://www.nber.org/papers/w32846