Why lower win rates are better

In the world of the ad exchange, your "win rate" is the ratio of bids to wins. This seems like an intuitive metric: if you want to win more, bid higher. Even with all of the crazy multi-tier pseudo-auctions we see these days, this is still generally the case.

I recently heard a client say "DSP XXX has higher win rates than DSP YYY... so therefore, DSP XXX is better". Counterintuitively, it's probably the opposite.

Let's assume a very simple scenario. You want to spend $100 on a simple audience-targeted campaign across some defined set of inventory, and you're flexible on the CPM that you pay. For this example, let's assume that the audience is fairly large, so there's plenty of inventory available. You log into DSP XXX and DSP YYY and set up the campaign, run it for a day, and look at the results.

DSPSpendCPMWin Rate
XXX$100$2.503.1%
YYY$100$1.100.6%

DSP XXX had a higher win rate at a higher CPM. But wait - I thought a higher bid would mean winning more impressions?

Pacing and Win Rate

Inside every DSP is a pacing algorithm that decides how to spend a budget over the course of some time interval. The basic idea is that if you want to spend $100 a day, you want to spend $4.16 per hour or about $0.07 per minute. For a typical audience-targeted campaign, a bidder sees around 10,000 eligible impressions a minute. If you buy every impression at a $2.50 CPM, you'll spend $25 per minute. So bidders bid on only a fraction of the impressions - a fraction determined by the win rate. If your win rate is 3.1%, you need to bid 32 times for every impression you win; to win 28 impressions you need to bid 900 times a minute.

So if you need to bid 900 times a minute to spend your budget at a $2.50 CPM, what does the DSP do for the other 9100 bid requests? It ignores them, or, "goes to sleep".

Let's look at DSP YYY and see what it's doing. With the exact same inventory, this DSP bids $1.10 and wins only 0.6% of the impressions. Let's do the math. At a lower CPM, you need to win 63 impressions a minute to hit our budget. At a win rate of 0.6%, you need to bid 150 times for every impression you win. This means you need to bid on around 9500 impressions to hit our budget - effectively, all of them. There's no need to go to sleep.

Gross Win Rate

Let's put these results in a new table to see them in a different light.

DSPCPMImpressions PurchasedImpressions TargetableGross Win Rate
XXX$2.5040,0009,000,0000.4%
YYY$1.1090,9009,000,0001.1%

If you calculate the win rate based on the total targetable volume instead of the impressions where DSP decided to submit a bid, the story is completely different. DSP YYY actually won more impressions on the same traffic.

Recommendations

  1. Require DSPs to provide Gross Win Rate instead of Net Win Rate in their reporting and analytics (AppNexus just added this in the "Bid Analyzer" tool)
  2. Don't adjust bidding strategy based on Net Win Rate; you may be lowering your actual win rate by raising your bids! (AppNexus v8 introduces automatic gross win rate optimization to find the optimal ratio)