The Advertising Economy

August 25, 2026

I graduated with a B.S. in Mathematics from Caltech in June 2025 and started working as a machine learning engineer in the advertising industry. This notebook collects what I have learned working inside digital advertising — the economy first, the mechanics in later entries. This entry is the map: who the players are, how the money and the information flow, and where the interesting fault lines are.

The transaction in one line

Strip away the acronyms and programmatic advertising is one exchange: a publisher owns audience attention and has inventory to sell; an advertiser owns budget and has a commercial goal to hit. Everything else in adtech exists to intermediate that exchange — matching the right impression to the right advertiser at the right price, at machine speed.

Advertiser ──(campaign/budget)──> DSP ──(bid requests)──> Ad Exchange <──(auction)── SSP <── Publisher
                                     ^                                |
                                     └─── Data Providers / DMP, Measurement, Ad Server ───┘

I’ll walk the actors in order of how much they matter to me, because a demand-side platform is the seat I sit in.

The actors

Advertiser (demand). Sets the objective — conversions, a CPA/ROAS target, reach — plus budget and creatives, and buys media programmatically rather than negotiating an insertion order per site.

Agency / trading desk. Not a platform, but a real layer: they hold advertiser money and operate the DSPs on the advertiser’s behalf. It matters because DSP product decisions are made for agency buyers — multi-advertiser accounts, approval workflows, reporting hierarchies — more often than for direct advertisers.

Demand-Side Platform (the DSP). Software that lets advertisers and agencies buy impressions across many supply sources from one interface, deciding audience, bid price, budget pacing, and creative per impression in real time. The IAB’s own framing: real-time bidding is “a way of transacting media that allows an individual ad impression to be put up for bid in real-time.” The boundary that separates a DSP from its predecessor, the ad network, is exact: the DSP “has the technology to determine the value of an individual impression in real time (less than 100 milliseconds).” The trade literature quotes 20–400 milliseconds end-to-end from bid request to decision. Later entries take that sentence apart; here I only want the definition on the table.

Ad Exchange. Neutral-ish marketplace technology running the auction between many DSPs and many sellers — a stock exchange for impressions. Historical first mover: the Right Media Exchange, launched 2005, acquired by Yahoo in 2007 — “approximately $680 million” for the remaining equity per Yahoo’s own 8-K, closed July 11 and booked at a $526M GAAP purchase price in Yahoo’s FY2007 10-K; the two figures coexist because they count different things, and both now sit on primary filings. Google’s AdX arrived via the DoubleClick acquisition — announced at $3.1B in April 2007, closed March 2008 after FTC and EU clearance — and became the default liquidity venue of the industry.

Supply-Side Platform (SSP) / publisher ad server. Sell-side software: aggregates publisher inventory, runs the auction(s), and optimizes yield with floors, header bidding and waterfalls, and deal management. Large publishers route many demand sources through an SSP the same way a trader routes through a smart order router. SSP tech is derivative of exchange tech — the auction engine is the same one, pointed the other way.

Ad network. The predecessor architecture: contract inventory from many sites in bulk, resell it, generally without impression-level real-time pricing. Networks still matter — curated packages, guaranteed campaigns, and notoriously, most walled-garden inventory still behaves network-like even when the plumbing speaks auction.

Publisher. Origin of supply: sites, apps, CTV channels, digital-out-of-home screens. In protocol terms the bid request describes supply with a site (web) or app (non-browser app) object plus imp objects for the sellable slots — a shape I dissect in the entry on OpenRTB.

The ancillary-but-load-bearing. Four more roles that never appear in the one-line transaction but own pieces of its spine:

• The ad server — creative management and serving/decisioning. The lineage is DoubleClick’s DART (“Dynamic Advertising, Reporting, and Targeting,” 1990s), built to raise buyer efficiency and reduce publisher unsold inventory; the ad server predates and outlives the auction.
• Data providers / DMPs — audience segments attached to bid requests. The UK ICO’s 2019 investigation into adtech found real-time-bidding participants collecting and trading rich personal data — race, sexuality, health status, political affiliation — often without proper consent. The data plane is not an auxiliary; it is load-bearing, with a legal spine.
• Measurement and verification — viewability, fraud. “Made-for-advertising” sites exist specifically to extract rent through the programmatic pipe with content-farm inventory; fraud is a documented externality of the channel, not an accident of it.
• Attribution — post-view/post-click conversion measurement. The DSP’s performance claim lives or dies here, and it is the least standardized layer in the entire stack. Two companies can disagree 30% on conversions for the same campaign and both be internally consistent.

How one impression flows (OpenRTB 2.x mental model)

1. A user hits a publisher page; the publisher/SSP constructs a bid request: device, geography, user id/segments, site or app, and imp (slot, size, formats, floor).
2. The exchange/SSP fans the request out to DSPs.
3. Each DSP runs its pipeline — eligibility filters, audience match, a value model (pCTR/pCVR or a direct expected-value estimate), a bidding policy with pacing and shading — and returns a seatbid with a price in currency per mille, a creative, and callback macros, or no-bids.
4. The auction clears (first-price in open display since roughly 2019; second-price historically and still at many places). The exchange notifies; the ad renders; win/impression/click/conversion events flow back and become training data.
5. Settlement reconciles auction logs against publisher invoices and advertiser bills. Fraud and discrepancy disputes are subtracted somewhere in that chain — and who subtracts it is a business negotiation, not a constant.

Note the loop in step 4: the auction is also the labeling function. Every auction is simultaneously a transaction and an experiment whose result is labeled late, partly, and by counterparties.

The economic asymmetry

The sentence I keep using: supply is abundant and near-perishable — an impression must be sold within ~100 ms or it’s gone — and demand is consolidated and slow-moving — budgets, targets, brand constraints move on timescales of quarters. The DSP’s function is to convert perishable supply into durable advertiser ROI, and its fee is earned only if that conversion is real.

Four systems wearing one trench coat

The mental model that has survived contact with everything else I know: a DSP is four systems wearing one trench coat.

1. A bid engine. OpenRTB JSON arrives; a price (or a 204) comes out within tmax.
2. A campaign object model. Advertisers think in campaigns, audiences, budgets, KPIs (eCPC/eCPA); the DSP translates that into per-impression arithmetic and reports back in their language.
3. A data pipeline. Every bid, win, impression, click, conversion becomes a log line, then a feature, then a training row, then an invoice line.
4. A financial engine. DSPs front the media cost and bill advertisers: margin = advertiser price − clearing price − fees − fraud refunds. Pacing, fraud deductions, and reconciliation with publisher invoices all live here. This is the part I am most confident is structurally right from public filings — The Trade Desk’s 10-K describes take-rate-on-spend almost in these terms — while the exact fee-stack conventions remain the least transparent numbers in the industry.

Of these, only the first two are visible to a buyer. The moat is three and four.

Why the DSP seat is the valuable one

Connectivity got commoditized. Once OpenRTB standardized (2010 onward), any DSP can connect to any exchange, and integration count stopped being an advantage. What cannot be commoditized is knowing what an impression is worth. Value estimation is machine learning plus proprietary feedback loops — your own conversions training your own models — so the modern moat is data and inference, not integrations. Google is the proof case: it turned a serving company (DoubleClick, $3.1B, 2008) into the AdX-exchange + DV360-DSP stack and then owned both sides of the trade; the DSP-side Invite Media acquisition (2010) was folded in, renamed DoubleClick Bid Manager, then Display & Video

1. Contrast an independent performance DSP whose whole game is the open internet’s remaining traceable data: bidstream, its own pixels, data partnerships, consent-granted first-party graphs.

Under a “performance DSP” business model, revenue ≈ take-rate × spend (plus sometimes data or service fees), and profitability is the accuracy of the value model. Overbid and the advertiser’s CPA target breaks — churn. Underbid and share is lost while fixed costs sit on less revenue. Which gives the compressed version of the economics: a DSP’s product is the error distribution of its value model; a DSP’s moat is the feedback data that narrows it.

Where the open questions are

Three things I do not yet know well enough to state as settled, and intend to close:

• The fee stack. Who takes what cut of a dollar of advertiser spend is the least transparent number in the industry — DSP fee vs. take vs. SSP cut vs. exchange fee, netted and rebated across contracts. The best near-primary sources are public filings: Trade Desk 10-Ks, AppLovin, Magnite, PubMatic, Criteo. I intend to read them closely (they also contain the honest versions of the industry stats the trade press floats).
• Open vs. walled. How much of the chain is still open RTB versus Google/Amazon/Meta, where — for the first time in this economy’s history — the same firm owns demand, supply, and identity, and its “auction” prices against first-party information no one else sees.
• Where money and information diverge. Fraud, attribution windows, and publisher arbitrage all live in the gap between what the auction log says and what the invoice says. The economics is downstream of the plumbing, and most of its hard problems are really disagreements about the plumbing’s feedback channel.

Those are the seams worth prying open, and I’ll keep prying. The auction mechanics themselves, the latency engineering, the bidding and pacing theory, and the machine-learning stack all have their own entries now.

References

1. IAB Tech Lab, “OpenRTB (Real-Time Bidding).” https://iabtechlab.com/standards/openrtb/
2. Wikipedia, “Real-time bidding.” https://en.wikipedia.org/wiki/Real-time_bidding
3. Wikipedia, “Demand-side platform.” https://en.wikipedia.org/wiki/Demand-side_platform
4. Wikipedia, “Ad exchange.” https://en.wikipedia.org/wiki/Ad_exchange
5. Wikipedia, “DoubleClick.” https://en.wikipedia.org/wiki/DoubleClick
6. Wikipedia, “Online advertising.” https://en.wikipedia.org/wiki/Online_advertising
7. ICO, Update report into adtech and real time bidding, 2019-06-20.
8. 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
9. Diaz Ruiz, C. (2024). “Disinformation and fake news as externalities of digital advertising.” Journal of Marketing Management. doi:10.1080/0267257X.2024.2421860
10. 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
11. Balseiro, S., Besbes, O. & Weintraub, G. (2015). “Repeated Auctions with Budgets in Ad Exchanges.” Management Science 61(4):864–884. doi:10.1287/mnsc.2014.2022
12. The Trade Desk, Form 10-K (annual reports; revenue-recognition and take-rate framing). https://investors.thetradedesk.com
13. 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/
14. 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/