How Much Data — The Minimum Viable Dataset for Every Kind of Advertising

October 8, 2026

Almost every argument in this industry is secretly an argument about sample size. “We need to segment further” and “the segment model is noisy”; “train it daily” and “the labels aren’t mature”; “buy the segments” and “we should mine our own logs” — each is a dispute about how many positives fit in a cell before a number stops being a guess. The arguments never name themselves because the arithmetic is unfashionable, but it is the only arithmetic that decides what a bidder can promise. I want to do it here, out loud, for every kind of campaign I know how to run: the binomial precision of a rate, the censored efficiency of a price, the trajectories of a controller, and the strategies they determine — from the cheapest-to-learn (retail media, most-walled) to the empty spreadsheet of conversational ads. It fits in one page of algebra. It has humbled me more often than any paper I’ve read.

Five templates

Rates. For a proportion p estimated from N opportunities with n positive events, the relative error is sqrt((1−p)/(pN)) ≈ 1/sqrt(n). Read it again with the emphasis where it belongs: precision is set by the positive count alone. The impressions are just how you pay for positives. Ten percent relative error needs 100 positives; five percent needs 400; twenty percent — which is model-input grade when a calibration layer sits behind it — needs 25. The same math lives in classical statistics as the events-per-variable rule for logistic regression, ten to twenty events per candidate predictor [1]. Now convert to impressions at a conversion base rate of 10^-k: a 100-conversion cell costs 10^(k+2) impressions. The public Criteo click-conversion export is the honest density anchor for k: 45.8 million rows over 24 days yield 1.33 million clicks (a 2.9% click rate) and 103 thousand conversions — a click-conditional rate of 7.7%, an impression-conditional rate near 2.2×10^-4 [4]. That mid-band k≈4 says a hundred conversions per segment cell costs a million impressions per refresh cycle, and it also exposes the scarcity asymmetry that rules how these models get fed: click-conditional conversion examples run about two orders of magnitude scarcer than impression-scale examples, which is why the conversion model forever eats the entire impression log to feed one funnel [2].

Prices under censoring. The bid-landscape model — the probability of winning at each price — is a censored-data problem by construction: a win shows you the clearing price, a loss only tells you it was above your bid. The statistical machinery is older than the auction: censored regression dates to Tobin’s limited-dependent-variable estimator in 1958 [5], and the censored-likelihood formulation of this exact model is standard in the KDD literature [3]. The number that matters for planning is efficiency: censoring a fraction c of observations costs roughly sqrt(1−c) of the effective sample — at 67% censoring you keep about 57% of what an uncensored log would have taught you, which is also the Tobit efficiency of the estimator. To pin a landscape at one context bucket to 10% relative error at a 5% win rate takes (2/0.1)²·19 ≈ 7,600 effective wins with prices — call it fifteen to twenty-five thousand logged bids per bucket depending on how much loss-notice relief you have, via the loss-notice mechanism the protocol carries for exactly this purpose [9]. Multiply by the context slices you care about — twenty geography × format × exchange cells is modest — and a full nightly refresh wants three to five hundred thousand bids per slice. Nobody has that hourly. This, and not any modeling philosophy, is why landscape models are nightly-granular.

Audiences. A self-trained lookalike needs at least ten thousand seed positives under the same events-per-variable logic [1]. At a conversion base rate of 10^-4, ten thousand seeds means a hundred million impressions of campaign history — so, in practice, only mature advertisers or aggregated vertical pools qualify to train their own, and everyone else buys the vendor’s model. The vendor shelf is not small: one filing counts “more than 370 third-party data vendors” integrated into a single platform [6].

Value. If you bid on revenue per conversion rather than conversion count, the label is continuous, and a mean with a coefficient of variation σ_v needs (σ_v/ε)² conversions. Order values are heavy-tailed; σ_v ≈ 3 is a working convention, not a measurement. For 10% relative error that is 900 conversions per cell — ten times the binary requirement, and the cheapest explanation of why value-based bidding is a feature of mature accounts.

Controllers. Pacing is the strange one: it learns trajectories, not labels. A campaign-day is about twenty-four (dual, spend-rate, remaining-demand) triples at the hourly publish cadence documented in the systems literature [7, 8], and fitting a controller offline wants thousands of campaign-days — hundreds of campaigns times weeks. In label terms pacing is the cheapest component and in breadth the most expensive: it needs no new data at all, only the logs you were contractually obliged to keep.

What each strategy actually requires

Retargeting is a coverage problem wearing a volume costume. The warm pool converts at the high end of the base-rate band, so 100-conversion cells are cheap — ten to fifty thousand impressions. What binds is identity: an unmatched ID is simultaneously unaddressable and unlabeled, and match rates across ID spaces run conventionally 30–70%. The minimum viable dataset is a pool export on at most a day-old cadence, a pre-launch match-rate dry run, and a hundred-plus conversions per segment you intend to price separately. Every retargeting failure I have diagnosed was a coverage failure misreported as a modeling failure.

Broad prospecting is where the base rate hits full force. At 10^-4, each targeting cell wanting its own conversion term eats a million impressions per refresh, and 10^-6 long-tail niches want a hundred million — beyond almost every advertiser, which is why sane systems share value hierarchically (fit globally, bias-correct per segment) instead of estimating per cell. The volume of a mid-size bidder funds this: one to five million peak queries per second, dayparting and edge filters shedding most of it, still leaves a hundred million to a billion bid opportunities a day, a few million impressions at a one-percent win [12], and a few hundred conversions — which is ten to twenty fresh 100-conversion cells per day, no more. That is the real budget line of the whole enterprise: a mid-size DSP funds about fifteen cells a day. Anything you can say about segment granularity follows from fifteen.

Context targeting is the cold-start escape hatch, because it requires zero user history: the features ride the request itself — site, app, category, the publisher’s taxonomy. Zero positives to start, ten to a hundred thousand labeled impressions to be good (click labels at 10^-3–10^-2 are an order of magnitude cheaper to earn than conversions). The price is a ceiling: without user state the click model saturates, and its real economic role is covering the traffic your own profiles can’t resolve.

Frequency caps need exposure-level truth. Detecting a 20% decay in conversion rate between cap-3 and cap-6 arms at a 5×10^-4 base rate is a two-proportion test demanding roughly 8×10^5 users per arm — a million to ten million users in flight before the response curve resolves any knee at all. Hence the cheaper ladder: hold out a randomly throttled five-percent cohort, and let cohort size, not model sophistication, set your error bars. The data minimum is unforgiving: every serve logged at exposure level, joined to conversions through the same keys, retained past the cap window. Flights that log at flight level cannot answer the only question capping poses.

CTV inverts the volume profile. Far fewer, longer, pricier impressions; clicks mostly absent, so engagement and completion outcomes replace the click model entirely; identity better than display’s — deterministic publisher-account and household graphs — but labels arriving as store-lift or clean-room purchase matches on a day-batch cadence with weeks-long attribution windows. The minimum viable dataset is not a dataset, it is a contract: ten-thousand-plus exposed households, a matched holdout, and a join agreement before launch. Pacing against day-level budgets on hour-class ticks is sufficient here precisely because event velocity is low.

Retail media has the best data on the planet and it never leaves. Impression, detail page, cart, purchase, all inside one platform’s first-party ledgers, labels in hours, base rates at or above the good end of the band — one filing even recognizes the revenue “on a net basis, as we act as an agent,” which tells you the platform is booking the fee and keeping the data [10]. A hundred-conversion cell costs a thousand to ten thousand impressions: the cheapest strategy to learn, the most constrained to port, because cross-platform use means a capped, day-old cohort export or a purchased lookalike. The walled garden’s moat is not the shopper. It is the join.

Conversational/LLM ads are the empty-spreadsheet strategy: no auction logs, no attribution convention, no landscape history, base rate unknown. The minimum viable program is procedural — log everything in the bid-opportunity shape from day one so labels become joinable retroactively; bootstrap with human evaluation and relevance rating; import priors from search query-intent models, the nearest labeled neighbor; seed one landscape bucket with about ten thousand exploratory bids, which suffices when you have one context class. The data is the moat at this phase, exactly as it was in the 2009–2012 open auctions, and every shortcut that skips logging is a loan against the next five years.

The ladder, the refresh, and the boutique

Across all strategies the cold-start order is the same and it is all sample size. Rung zero: rules and category benchmarks, zero labels. Rung one: a global model plus a per-segment multiplier calibrated on ~100 impressions per segment — against the 100 conversions needed to train a segment from scratch; that asymmetry, two orders of magnitude for the price of a calibration layer, is why every platform in history starts there. Rung two: coarse cross-segment models at about a million labeled impressions. Rung three: personalization, needing ten-plus events per user, which day one only retargeting-adjacent pools can afford because they bring their own features. Rung four: landscape and learning-based bidding, needing the fifteen-to-twenty thousand bids per bucket plus censoring-relief plumbing — historically the last rung a market climbs, and it is no accident that the bid-shading literature flourished only after first-price auctions returned in 2019.

In steady state the refresh cadences differ per component and each one is the delay law speaking. Click models stream — hourly increments, daily full, billions of predictions a day in the published accounts [11]. Conversion models would go stale waiting thirty days for labels when roughly 35% of conversions arrive within an hour, half by a day, but a persistent 13% tail arrives after two weeks [2]; the delayed-feedback reweighting of immature labels makes a daily retrain admissible once conversion flow reaches about a thousand a day, and weekly-with-pooling below a hundred. Landscapes stay nightly and let the pacing dual absorb intraday drift, because refitting a censored model hourly costs precision it cannot spend. Lookalikes refresh weekly on clean-room cadence; pacing publishes hourly as published systems have since the 90-minute feedback controllers [7, 8].

And then the boutique — the case the arithmetic is really for. At ten queries per second: 8.6×10^5 requests a day, on the order of 10^5 bids after edge shedding, a couple thousand impressions at a 2% win, ten clicks, and 0.2 to 2 conversions a day. No per-segment click model. No per-bucket landscape — one bucket would take a year. No self-trained conversion or value model. The viable stack is rungs zero and one, everywhere: rules, context, purchased segments, vendor lookalikes, one global calibrated model, and at most one pooled landscape at campaign granularity. This is the junk-shed factor of the exchange business translated into economic dress: 80–90% of queries are shed at the edge by the big bidders precisely because their data economics don’t need them, while the small bidder needs everything and gets nothing.

One budget note closes the loop to the money. Purchased data and identity are booked — in the only filing that lets us see — inside platform operations expense: hosting, traffic, compute, data, staff, ≈4% of gross spend against a ≈21% take [6]. Any input whose cost exceeds a tenth of the take is structurally uncompetitive, and the marginal data purchase competes against free label fabrication from your own event plane. So the rational ordering as you scale is always: fabricate before you buy, and the build-vs-buy crossover is exactly the day a purchased segment stops filling 100-positive cells faster than your own pipeline does.

The industry argues about models. The models argue about labels. The labels argue about counts. One hundred positives, fifteen thousand bids, ten thousand seeds, eight hundred thousand users, a year of one bucket: write the five numbers on an index card, and most strategy disputes resolve themselves before the first meeting ends.

References

1. Peduzzi, M., Concato, J., Kemsky, E., Silbersnitz, H. & Feinstein, A. R. (1996). “A Simulation Study of the Number of Events Per Variable in Logistic Regression Analysis.” Journal of Clinical Epidemiology 49(12). doi:10.1016/S0895-4356(96)00236-3
2. Chapelle, O. (2014). “Modeling Delayed Feedback in Display Advertising.” KDD. doi:10.1145/2623330.2623634
3. Zhang, W., Yuan, S. & Wang, J. (2014). “Optimal Real-Time Bidding for Display Advertising.” KDD. doi:10.1145/2623330.2623633
4. Criteo. Display Advertising Challenge dataset (45.8M rows, 24 days), archived at UCI. https://archive.ics.uci.edu/dataset/477/criteo+display+advertising+challenge
5. Tobin, J. (1958). “Estimation of Relationships for Limited Dependent Variables.” Econometrica 26(1). doi:10.2307/1907382
6. The Trade Desk, Inc. (2026). Annual Report (Form 10-K) for fiscal year 2025 (370+ data vendors; data-related costs in platform operations expense), filed 2026-02-27. https://www.sec.gov/Archives/edgar/data/1671933/000167193326000014/
7. Zhang, W., Rong, Y., Wang, J., Zhu, T. & Wang, X. (2016). “Feedback Control of Real-Time Display Advertising.” WSDM. doi:10.1145/2835776.2835843
8. Yang, X., Li, Y., Wang, H., Wu, D., Tan, Q., Xu, J. & Gai, K. (2019). “Bid Optimization by Multivariable Control in Display Advertising.” KDD. doi:10.1145/3292500.3330681
9. IAB Tech Lab, OpenRTB Version 2.6 (lurl loss-notice URL and price macros). https://github.com/InteractiveAdvertisingBureau/openrtb2.x/blob/main/2.6.md
10. Criteo S.A. (2026). Annual Report (Form 10-K) for fiscal year 2025 (retail media revenue recognized net, as agent), filed 2026-02-26. https://www.sec.gov/Archives/edgar/data/1576427/000157642726000014/
11. McMahan, H.B. et al. (2013). “Ad Click Prediction: a View from the Trenches.” KDD. doi:10.1145/2487575.2488200
12. 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