Ad Auction Mechanisms & Smart BiddingOpen in Colab

Ad Auction Mechanisms & Smart Bidding

Online ad auctions allocate ad slots to competing advertisers in real time (< 100 ms). The auction mechanism determines who wins, what they pay, and whether the equilibrium is truthful. This page covers the major mechanisms and the ML systems advertisers use to bid optimally.

Vickrey (Second-Price) Auction

The simplest truthful single-slot auction:

  • nn advertisers submit sealed bids b1,…,bnb_1, \ldots, b_n.
  • The highest bidder wins: winner=arg⁡max⁡ibi\text{winner} = \arg\max_i b_i.
  • The winner pays the second-highest bid: p=b(2)p = b_{(2)} (the order-statistic notation for the second largest).

Dominant strategy: Bidding your true value viv_i is dominant. You can never gain by overbidding (win at a loss) or underbidding (lose an auction you'd have profited from). This makes the auction incentive-compatible.

Proof sketch: Suppose vi=10v_i = 10 and the highest opponent bid is b−i=7b_{-i} = 7.

  • Bid vi=10v_i = 10: win, pay 7, profit 10−7=310 - 7 = 3. ✅
  • Bid bi=12b_i = 12 (overbid): same win/loss decision if b−i<12b_{-i} < 12, no gain.
  • Bid bi=5b_i = 5 (underbid): lose if b−i=7>5b_{-i} = 7 > 5, leave 3 profit on the table. ❌

Generalized Second Price (GSP) Auction

Used by Google Ads, Bing Ads for multiple ranked ad slots:

  • KK slots with click-through rates α1>α2>…>αK\alpha_1 > \alpha_2 > \ldots > \alpha_K (slot 1 gets the most clicks).
  • Advertisers are ranked by bid × predicted quality score (Google's quality score incorporates expected CTR, ad relevance, landing page quality).
  • Advertiser ranked kk pays the next advertiser's bid (the bid that would just displace them from slot kk).
pk=bk+1⋅qk+1qkp_k = \frac{b_{k+1} \cdot q_{k+1}}{q_k}

where qkq_k is the quality score of the advertiser in slot kk.

Not fully truthful (unlike Vickrey): the dominant strategy is not always to bid true value in GSP. But in practice, GSP converges to a Nash equilibrium where sophisticated advertisers bid approximately truthfully after learning.

VCG Mechanism

The Vickrey-Clarke-Groves mechanism generalizes second-price truthfulness to multiple slots:

Each winner kk pays:

pkVCG=∑j>kαj(bj+1−bj)⋅(externality term)p_k^{\text{VCG}} = \sum_{j > k} \alpha_j (b_{j+1} - b_j) \cdot \text{(externality term)}

VCG charges each advertiser their externality — the harm their presence causes to other advertisers by displacing them to lower slots. This is the theoretically optimal truthful mechanism for multi-slot auctions, but it's computationally complex and Google has documented that GSP approximates VCG equilibria in practice.

Reserve prices

A reserve price rr is the minimum eCPM required to show an ad:

show ad  ⟺  eCPMi=CTRi×bi≥r\text{show ad} \iff \text{eCPM}_i = \text{CTR}_i \times b_i \geq r

Why set a reserve price?

  1. Revenue floor: prevents inventory from being sold below production cost.
  2. Quality control: low-bid, low-quality ads degrade user experience even if they "win."
  3. Incentive to bid higher: a reserve price near the second-bid equilibrium pushes advertisers to bid higher, increasing revenue.

ML-optimized reserves: platforms learn per-slot, per-advertiser reserve prices from historical auction data to maximize expected revenue (Myerson's optimal auction theory).

Smart Bidding: ML for advertisers

Smart bidding (Google Ads) / Advantage+ bidding (Meta) automates the advertiser's bid strategy:

Target CPA (Cost Per Acquisition)

The advertiser sets a target CPA. The bidding system finds the optimal bid per auction:

bauction∗=target CPA×P(conversion∣context)b^*_{\text{auction}} = \text{target CPA} \times P(\text{conversion} \mid \text{context})

The system uses a conversion prediction model (trained on the advertiser's conversion signals) and adjusts the bid to match the target CPA over a rolling 7-day window.

Target ROAS (Return on Ad Spend)

Maximize conversion value (revenue) subject to a minimum ROAS constraint:

bauction∗=E[conversion value]target ROASb^*_{\text{auction}} = \frac{\mathbb{E}[\text{conversion value}]}{\text{target ROAS}}

Portfolio bidding

Manage bids across a portfolio of campaigns, shifting budget toward campaigns with the highest marginal return:

max⁡∑cVc(xc)s.t.∑cxc≤B\max \sum_c V_c(x_c) \quad \text{s.t.} \sum_c x_c \leq B

where Vc(xc)V_c(x_c) is the value function for campaign cc as a function of budget xcx_c and BB is the total budget. This is solved with a Lagrangian approach, treating the shadow price of the budget constraint as the universal bid modifier.

Budget pacing

Without pacing, a daily budget exhausts in the first hour. Pacing adjusts the bid multiplier ρt\rho_t dynamically:

bteffective=ρt⋅b∗b_t^{\text{effective}} = \rho_t \cdot b^*

The pacing controller (typically a PI controller) computes ρt\rho_t by comparing actual spend rate vs. target spend rate:

ρt=ρt−1+Kp⋅et+Ki⋅∑τ≤teτ\rho_t = \rho_{t-1} + K_p \cdot e_t + K_i \cdot \sum_{\tau \leq t} e_\tau

where et=target_rate−actual_ratee_t = \text{target\_rate} - \text{actual\_rate} is the spend error. LinkedIn's system achieves < 5% budget under-delivery using this approach.

Related concepts

Concept neighbourhood

How this concept connects — click any node to jump there, or explore the whole concept map.

Referenced by
Ad Auction Mechanisms & Smart Bid…Ad Ranking & CTR Predicti…Diversity, Cold Start & E…