A Brief History of Programmatic Advertising

October 6, 2026

The thesis first, before the dates: for most of its life, the hard part of a demand-side platform was not intelligence but plumbing — receiving bid requests in time, speaking the exchange’s protocol, spending a budget without overdrawing. Each era solved one bottleneck, and each solution leaked to everyone: patents expired or were designed around, protocols standardized, papers published, open source shipped, datasets released. So the frontier of competition moved one layer inward each time — first the auction mechanism itself (2005–09), then the auction-theoretic and control-theoretic machinery of optimal bidding (2010–16), then large-scale machine learning over the bid landscape (2013–20), and finally the data and identity infrastructure that lets any of it run once cookies and third-party signals are gone (2019–now). The DSP of today is an inference-and-data company that happens to speak OpenRTB — which is the same thing the ad server of 2005 was, one abstraction layer up. This entry walks the eras and, at each one, notes both what the hard problem was and how it leaked.

Era 0, before 2005: rules, not auctions

The pre-programmatic stack was a scheduling problem, not a pricing problem. DoubleClick’s DART (Dynamic Advertising, Reporting and Targeting, founded 1995, IPO 1998) served banner campaigns booked in advance against insertion orders: the unit of computation was the placement — site × size × date × frequency cap — delivered by a central ad server with a traffic-and-counting engine, measured by clicks and server-side impression logs. Batch traffic estimation, linear-count delivery, per-campaign rules engines. There is no per-impression decision, so there is no DSP yet.

Two artifacts of that era determined everything after. The third-party cookie let an ad server recognize a browser across sites, which is what made audience targeting — as opposed to inventory targeting — representable at all: it is the data substrate the current identity era is still litigating. And keyword auctions in search proved that ad inventory could be priced by auction in software: GoTo.com ran the first search keyword auction in 1998; Google AdWords launched 2000 and added quality-based (bid × CTR) ranking in

1. The formal analysis of what search actually ran — the Generalized Second Price auction, and its lack of truthfulness (local saddle-point equilibria instead of dominant strategies) — arrived afterward, in Edelman, Ostrovsky & Schwarz (2007). The gap between practitioners shipping auctions and theory naming them recurs in every era below; it is this field’s most reliable pattern.

Era 1, 2005–2009: inventing the per-impression auction

Three nearly simultaneous threads produced real-time bidding. The Right Media Exchange (2005, Brian O’Kelley) built the first modern ad exchange: inventory from multiple publishers auctioned in software, with RTB-style prediction, budgeting, and pacing at the buyer; Yahoo acquired it in 2007, buying the remaining equity for “approximately $680 million” in Yahoo’s own 8-K (filed 2007-05-02), closed July 11, booked at a $526M GAAP purchase price in Yahoo’s FY2007 10-K — the $680M headline and the $526M book number are not in conflict; they count different things. Knapp & Blanco at Strategic Data Corp filed in 2006 the patent that defines the concept — Auction For Each Individual Ad Impression, US 2008/0162329A1 — claims covering auctioning each impression in real time among bidders, published 2008-07-03; the ur-document of the field’s IP thicket, filed into by Right Media’s people, AppNexus’s Serff, and Google’s DoubleClick lineage through 2011. And DoubleClick AdX (2008) gave the practice a default liquidity venue after Google’s $3.1B acquisition closed — announced April 2007, FTC-cleared December, EU-approved and closed March 2008.

The technical problem-set of this era is still the DSP’s load-bearing engineering today, and it was all solved before any of the theory existed:

• Latency budget. An exchange allows ~100 ms per bid request; tmax is defined to include internet latency. Everything — profile lookup, model evaluation, budget check — must fit inside that.
• Throughput. Tens of thousands of queries per second per cluster; exchange QPS grows with the number of exchanges, and a DSP must hear from all of them to know the market.
• State at speed. Budget pacing — spend $X/day without overdrawing or front-loading — became a distributed-systems problem: a shared counter updated at auction rate, classic read-modify-write contention. The early solutions were probabilistic throttling, the “spend governors” — control loops shipped half a decade before their published control theory (Zhang et al., WSDM 2016).
• Protocol fragmentation. Every exchange had its own XML dialect; N exchanges × M DSPs integration cost was both the moat and the tax of the era, and the reason standardization was inevitable.

Era 2, 2010–2013: the plumbing standardizes, measurement becomes science

The OpenRTB Consortium assembled November 2010, convened by buy- and sell-side vendors, when RTB was about 4% of display. OpenRTB 2.0 (2012-01) unified display, mobile, and video into one JSON API; the version train ran through 2.x (VAST video and geo in 2.1, PMP deal ids and bot flags in 2.2, native in 2.3, audio in 2.4, header-bidding signals and viewability fields in 2.5), and a 3.0 (Sept 2017) added signed bid requests and supply-path transparency before being effectively absorbed back into the 2.x lineage. By 2017, one trade estimate had programmatic near 29% of the digital mix.

Protocol commoditization is the event of this era: once connectivity was a checkbox — bidder kits, SDKs, and eventually Prebid — differentiation had to move upstream, into decision quality. The market data tracks the intuition: US RTB spend $23.5B (2018) to roughly $57B (2023).

At the same time, the auction became legible. Yuan, Wang & Zhao (2013) was the first public measurement study of RTB from a DSP’s own logs: winning-price distributions, the censoring problem — you only observe the clearing price when you win — and a distributional result that aged better for being negative: the log-normal price hypothesis fails when you test it where it’s supposed to live. Splitting 50 placements × 160 days into ~192k ⟨placement, hour⟩ segments, under 1% passed Shapiro–Wilk or Anderson–Darling at segment level, no matter the bidder or the hour’s volume. Aggregate pools can still look log-normal-ish, which is why the assumption kept haunting the parametric literature; per-segment, it was dead on arrival, and that is the field-measured reason the nonparametric lines (histogram leaves, mixtures, distribution-free ladders) won. Cui et al. (KDD 2011) had already named the object: the bid landscape, the win-rate-versus-bid curve, and the censored-data estimation problem it poses. And in 2014 the iPinYou dataset — 64.7M bid request logs published with a simulation harness (“truthfully bid on the impression logs” to simulate the full-volume auction) — turned the whole field into a reproducible empirical science. Everything in Era 3 is legible because iPinYou exists.

Era 3, 2013–2016: the algorithmic core goes public

Three research threads, all visible from outside a company, became table stakes:

(a) Optimal bidding. Zhang, Yuan & Wang (KDD 2014) gave the canonical recipe — model the winning-price distribution per segment, choose the bid to maximize value subject to budget via one Lagrangian — with provable better spend-efficiency than fixed-CPB bidding, and the refinements stacked fast: censored-Gaussian regression (Wu et al. 2015), survival-tree functional landscapes (Wang et al. 2016), deep censored learning (Wu et al. 2018), mixture-density networks at Adobe (Ghosh et al. 2019), LSTM price ladders (Ren et al., KDD 2019). Each paper relaxed an assumption of the last; this lineage is the subject of a whole entry.

(b) RL and control for pacing. Spend-governor folklore got published and generalized: PID feedback control of delivery (Zhang et al., WSDM 2016), multivariable MPC (Yang et al., KDD 2019), RL bidders formulating the day as an MDP (Cai et al., WSDM 2017; multi-advertiser games, Jin et al., CIKM 2018). And the theory arrived for what the heuristics had been doing: Balseiro, Besbes & Weintraub (Management Science 2015) showed that in repeated auctions with budgets, bidding one’s true value is not dominant — pace-adjusted, uniformly-discounted bidding is the equilibrium behavior. In other words, every DSP’s λ-throttling had been reconstructing an equilibrium strategy since 2006, empirically, before anyone proved that was what it was.

(c) CTR/CVR as big ML. Google’s “Ad Click Prediction: a View from the Trenches” (McMahan et al., KDD 2013) popularized FTRL and the discipline of calibrated prediction at billions-of-impressions scale; Factorization Machines (Rendle 2010) became the standard sparse-interaction substrate; Wide & Deep (2016) and Deep Interest Network (KDD 2018) moved the industry toward attention-style models. CVR’s special pathology — conversions arrive after the auction, so labels are censored by time — was formalized by Chapelle (KDD 2014), the delayed-feedback model.

Why this era matters for the thesis: it is the publication diffusion. The buy side’s edge from 2006–2012 had been undisclosed heuristics. From 2013–16 the same machinery arrived as KDD papers, an open dataset, and (soon) open-source implementations. Pacing and landscape estimation were no longer differentiators by about 2017 — visible in the fact that by 2019 the papers had moved on to auction design itself.

Era 4, 2016–2021: the mechanism changes under everyone

The most under-documented pivot in the field. Beginning in 2019 the dominant exchanges — Google AdX (announced March 2019) and the open exchanges, Amazon TAS among them — moved display from second-price to first-price auctions, meaning every winner pays its own bid. Header bidding (Prebid, launched 2015, 300+ adapters, pages calling 5–15 bidders in parallel) concurrently flattened the waterfall, so the clearing price became the max of parallel bids rather than the runner-up of a serial one. The two changes compounded:

1. Bid shading went from edge trick to core subsystem — the full story in its own entry. Yahoo’s CIKM 2020 paper and its KDD 2021 successor documented the industry fix so thoroughly that by 2022 it too was commodity.
2. Latency engineering again: header bidding multiplies QPS, and the timeout chain of everyone’s reference implementation now sets the real per-bidder budget.
3. Autobidding became the product. Advertisers stopped setting CPMs; platforms’ tCPA/tROAS buyers became the default buyer. Theory caught up fast: uniform value-maximizing pacing multipliers dominate in truthfully-solvable mechanisms (Aggarwal et al., 2019); the pacing equilibrium in first-price markets exists and is unique (Conitzer et al., EC 2019 / Management Science 2022); the first-vs-second welfare gap under autobidding got bounded (Deng et al., NeurIPS 2024); the migration itself got measured with difference-in-differences on staggered publisher switches — revenue per sold impression jumped 25–70% of pre-treatment levels at the switch, then reverted toward second-price levels as bidders learned to shade (Goke, Weintraub, Mastromonaco & Seljan, EC 2022), with soft floors and supply-side competition as the adoption channel (Despotakis, Ravi & Sayedi, JMR 2021).
4. Supply-path forensics: signed bid requests, schain/sellers.json/ ads.txt — an Era-2 deliverable implemented as an Era-4 emergency — because the stack had grown four to six intermediaries deep and the buyer wanted each hop attested at bid time.

Era 5, 2019–now: the data substrate taken away

The mechanism and algorithm layers are now open; the contest is over what they run on.

• Law arrives. The ICO’s 2019 report found RTB broadcasting user-level data processing special-category data (race, sexuality, health, politics) without lawful basis. The Belgian DPA’s Decision 21/2022 voided core parts of IAB Europe’s Transparency & Consent Framework under GDPR (the trade called it an “atomic bomb” for good reason); the Dutch DPA ordered an RTB profiling halt. GDPR/CCPA consent banners measurably prune addressable inventory. The legality of a feature is now a product requirement.
• Identity deprecation. Apple ITP (2017 →) capped third-party cookies; ATT (iOS 14.5, April 2021) made the IDFA opt-in; Google’s cookie-deprecation saga announced, wavered, and reversed through 2024 with the Privacy Sandbox (cohort-based APIs like Turtledove) in the interregnum. The technical response: probabilistic identity graphs and stitching, unified-ID schemes (UID2 et al.), and contextual/first-party re-weighting — the feature-availability shift itself becoming an ML problem.
• Compliance rides in the protocol. GPP and consent strings inside OpenRTB regs make the bid request itself carry legal metadata — compliance evaluated per auction, a genuinely novel protocol fact.
• Auctions move server-side (Prebid Server, ad-server-side, CTV server-side rendering), which shifts the DSP’s systems profile from edge-latency engineering toward log-scale stateful joins: event → identity → conversion. Increasingly a data-engineering company.
• MFA fraud becomes measurable: content-farm inventory arbitraging the pipe, quantified from exchange logs (Kochalski et al., 2021, at the best available confidence); detectable by policy — ads.txt/sellers.json — more than by ML.
• And the LLM/agent era (2023– ), early and evidence-thin: LLMs as text encoders in CTR models; “agentic” framing on DSP marketing pages (evidence: marketing, flag only); academically, autobidding now written as multi-agent mechanism design (Aggarwal, Perlroth, Zhao, EC 2023). Whether a bidder-as-agent changes the economics or just the UI is the live question.

The synthesis: a commoditization treadmill, layer by layer

Every row is the same move: an information-processing advantage → a publication, standard, or patent expiration → the advantage relocating into private data or private contracts. The compute was never the moat; the per-request cost curve is a constraint DSPs share. The history is of differentiation retreating from shared constraints into non-shareable places: first patents, then algorithms, then logs of proprietary interactions, now consent-granted first-party graphs. The last layer that hasn’t leaked is data — which is exactly why the era we’re in looks less like an algorithms race and more like a legal-and-identity one.

Where the primary sources are (for the curious)

IAB Tech Lab’s OpenRTB version history and the spec PDFs on GitHub; US patent 2008/0162329A1 from USPTO/Google Patents; DoubleClick’s 2004 10-K on SEC EDGAR for DART and ASP-model economics; Yahoo’s own 8-K of 2007-05-02 (the Right Media press release, “approximately $680 million”) and its FY2007 10-K Note 3 (the $526M GAAP figure) on the same EDGAR box; FTC docket P074102 and EU case COMP.52577 for the 2008 deal; the Belgian DPA’s Decision 21/2022 itself, with Veale et al.’s analyses. The remaining trade numbers (spend series, percentages) still travel through Wikipedia and Statista chains and deserve primary confirmation before I repeat them at anyone important — which is written here mostly so that future me doesn’t forget that promise.

References

1. IAB Tech Lab, “OpenRTB (Real-Time Bidding),” incl. full version history. https://iabtechlab.com/standards/openrtb/
2. Knapp, R. & Blanco, M. (2006). “Auction For Each Individual Ad Impression.” US 2008/0162329A1, published 2008-07-03. https://patents.google.com/patent/US20080162329A1
3. Wikipedia, “DoubleClick.” https://en.wikipedia.org/wiki/DoubleClick
4. Wikipedia, “Ad exchange” (History: Right Media 2005, O’Kelley; Knapp/Blanco filing). https://en.wikipedia.org/wiki/Ad_exchange
5. Yahoo! Inc., Form 8-K (filed 2007-05-02; exhibit: Right Media press release, ‘approximately $680 million’ for remaining equity). https://www.sec.gov/Archives/edgar/data/1011006/000115752307003677/
6. Yahoo! Inc., Form 10-K FY2007 (filed 2008-02-27; Note 3: Right Media GAAP purchase price $526M). https://www.sec.gov/Archives/edgar/data/1011006/000089161807000108/
7. Wikipedia, “Online advertising” (early banner/search milestones). https://en.wikipedia.org/wiki/Online_advertising
8. Wikipedia, “Real-time bidding” (spend series via Statista; privacy & fraud sections). https://en.wikipedia.org/wiki/Real-time_bidding
9. Edelman, B., Ostrovsky, M. & Schwarz, M. (2007). “Internet Advertising and the Generalized Second-Price Auction.” American Economic Review 97(1):242–59. doi:10.1257/aer.97.1.242
10. Yuan, S., Wang, J. & Zhao, X. (2013). “Real-time Bidding for Online Advertising: Measurement and Analysis.” arXiv:1306.6542. doi:10.48550/arXiv.1306.6542
11. Cui, Y., Zhang, R., Li, W. & Mao, J. (2011). “Bid Landscape Forecasting in Online Ad Exchange Marketplace.” KDD. doi:10.1145/2020408.2020454
12. Zhang, W., Yuan, S., Wang, J. & Shen, X. (2014). “Real-Time Bidding Benchmarking with iPinYou Dataset.” arXiv:1407.7073
13. Zhang, W., Yuan, S. & Wang, J. (2014). “Optimal Real-Time Bidding for Display Advertising.” KDD. doi:10.1145/2623330.2623633
14. Balseiro, S., Besbes, O. & Weintraub, G. (2015). “Repeated Auctions with Budgets in Ad Exchanges.” Management Science 61(4). doi:10.1287/mnsc.2014.2022
15. Zhang, W., Rong, Y., Wang, J., Zhu, T. & Wang, X. (2016). “Feedback Control of Real-Time Display Advertising.” WSDM. doi:10.1145/2835776.2835843, arXiv:1603.01055
16. Cai, H., Ren, K., Zhang, W. & Wang, K. (2017). “Real-Time Bidding by Reinforcement Learning in Display Advertising.” WSDM. doi:10.1145/3018661.3018702, arXiv:1701.02490
17. McMahan, H. B. et al. (2013). “Ad Click Prediction: a View from the Trenches.” KDD. doi:10.1145/2487575.2488200
18. Chapelle, O. (2014). “Modeling Delayed Feedback in Display Advertising.” KDD. doi:10.1145/2623330.2623634
19. Rendle, S. (2010). “Factorization Machines.” IEEE ICDM. doi:10.1109/ICDM.2010.127
20. Ren, Q. et al. (2019). “Deep Landscape Forecasting for Real-time Bidding Advertising.” KDD. doi:10.1145/3292500.3330870, arXiv:1905.03028
21. Aggarwal, G., Badanidiyuru, A. & Mehta, A. (2019). “Autobidding with Constraints.” WINE. doi:10.1007/978-3-030-35389-6_2
22. Conitzer, V., Kroer, C., Sodomka, E. & Stier-Moses, N. (2019/2022). “Pacing Equilibrium in First-Price Auction Markets.” EC 2019, doi:10.1145/3328526.3329600; journal version Management Science 70(2), 2024. doi:10.1287/mnsc.2022.4310, arXiv:1811.07166
23. ICO, Update report into adtech and real time bidding, 2019-06-20.
24. Belgian DPA, Decision 21/2022 re IAB Europe TCF, 2022-02-02.
25. Veale, M. & Zuiderveen Borgesius, F. (2022). “AdTech and Real-Time Bidding under European Data Protection Law.” German Law Journal. doi:10.31235/osf.io/wg8fq
26. Veale, M., Nouwens, K. & Santos, C.T. (2022). “Impossible Asks: Can the Transparency and Consent Framework Ever Authorise Real-Time Bidding After the Belgian DPA Decision?” TechReg. doi:10.26116/techreg.2022.002
27. Prebid, “Introduction to Prebid” (header bidding launch, adapters, parallel bidders). https://docs.prebid.org/overview/intro.html
28. Kochalski et al. (2021). “Detecting Ad Fraud and Financial Losses in Digital Advertising.” (AdKDD; citation chain via secondary sources — arXiv ID unresolvable this pass, flagged rather than fabricated.)
29. Aggarwal, G., Perlroth, M. & Zhao, J. (2023). “Multi-Channel Auction Design in the Autobidding World.” EC. doi:10.1145/3580507.3597707
30. Goke, I.J., Weintraub, G., Mastromonaco, A. & Seljan, M. (2022). “Bidders’ Responses to Auction Format Change in Internet Display Advertising Auctions.” EC. doi:10.1145/3490486.3538249, arXiv:2110.13814
31. Despotakis, S., Ravi, R. & Sayedi, A. (2021). “First-Price Auctions in Online Display Advertising.” Journal of Marketing Research 58(5). doi:10.1177/00222437211030201