Notes on Auto Biddings

First Talk

Market Algorithms for Autobidding by Gagan Aggarwal - Google Research

Background: After a user query, a collection of relevant ads are selected. An auction then determines which ads to show (to an user).

Advertisers are buyers.
Items to sell can be:
a. Slots (where ads are displayed) on a website.
b. Advertisers often bid on keywords, and participate in multiple auctions with queries for the relevant keywords.
Company (e.g., Google, Meta) are sellers.

The common auction for selling a single item is the second price auction.

Auto-bidding

Instead of submitting the exact bids, advertisers provides goals (e.g., maximizing utility, value, number of conversions) and constraints (e.g., budget, ROI) to an automated system. The system then sets the bids on advertisers' behalf in auctions, aiming to achieve the goal (i.e., optimize objectives) s.t. the constraints.

Question I: Bidding Algorithms

One example bidding algorithm is discussed on page 5 - 6 in the survey paper by Aggarwal et al.' 24', with a more detailed explanation in this 2019 paper by ABM' 19. Here, the optimal bid for an agent in auction $j$ is of the form:

b_j = \frac{v_j + \sum_c \alpha_c w_j^c}{\lambda + \sum_c \alpha_c}

obtain from the dual program. The highlight is that, $(i)$ in a truthful auction (e.g., second-price) where $(ii)$ the prices of items are known (and fixed), by bidding the above amount, an agent gets the items that optimize its objective.

The above results is for the scenario where only one advertiser uses an auto-bidding system, where other submit their fixed bids.

Question II: Equilibria

When everyone uses autobidding - Dynamics and Equilibrium. Some work on this: APZ' 23, ABM' 19.

Question III: Auction design

This is from an auctioneer's perspective. Many factors to consider(e.g., reserve price, multi-channel design). This issue is more complicated when buyers are value-maximizing.

This multi-channel auction design is very interesting.