From CCTV to Google: Talk about the mechanism design of bidding ads

Author: Wei Xi, Sina Weibo advertising product manager, is committed to analyzing the basic logic, ideas and techniques of Internet advertising. Public No.

On the auction advertising, many people have two extreme views - the first point of view will think that the bidding advertising is very simple, is not CCTV each year gold advertising position bidding, who bid high advertising position to whom;

The second extreme will think that the bidding advertising is too complex, involving game theory, mechanism design, auction theory, CTR estimation and other ordinary people simply do not understand the theory of the deep, so many people are discouraged from this ... ...

But is the truth a god horse? Which view is right? The guardian, who has been trying to dissect the underlying logic of Internet commerce in simple language, will tell you:

The truth is like the end of Pony Crossing: the river is neither as shallow as the yellow cow says, nor as deep as the squirrel says - the basic logic of bidding advertising certainly involves a lot of complex theories, but most people who read this article can easily understand the origins of its core principles.

From CCTV bid king to Google's auction advertising system

Regarding the bidding advertising, many people think that Baidu is the first domestic auction advertising company, in fact, in a broader sense, CCTV's gold advertising bidding earlier to popularize the concept of "bid advertising", as early as 1994, CCTV began to auction advertising positions by bidding, the birth of Qinchi, Edo VCD and other well-known king.

A typical auction scenario is this - the auctioneer first marks the advertising position, also indicates the floor price, and then starts waiting for the bidder to bid upward, "1 million!" "1.2 million! "2 million! ",""2 million times, 2 million twice, deal! "

Such scenes are familiar to many films and tv productions, but not all auctions are conducted in this way, and this is just one of many auction mechanisms, which are called "British auctions".

Similar to British auctions, as well as Dutch auctions, dutch auctions are the opposite, with auctioneers starting with a very high price and then experimenting until a bidder is willing to accept the deal.

So the question is - traditional advertising can bid in this way, so can online advertising bid in the same way? The answer is no, because there are several important differences between online ad auctions and traditional ad auctions, and these differences can have an important impact on the design of the bidding mechanism.

First, whether it's a British or Dutch auction, everyone's bid is public, and online advertisers may be a lot reluctant to make their bidpublic, so public bidding becomes inappropriate.

Second, CCTV's auction is a single act, and online advertising is a repeated game, that is, this advertising position was robbed, advertisers can also grab the next advertising space, advertisers can constantly adjust their bids.

Third, online advertising has multiple subject matter (each ad request may be more than one auction object), a large number of real-time calculation characteristics.

So what kind of bidding mechanism should online advertising take? In fact, there are many options, we look at one by one:

The first possible option is the "sealed first price" bid, which is in fact a familiar bidding method, many of the project bidding is taken in this way, its mechanism is that each bidder does not publish their bid, sealitit it in an envelope to the seller, the auctioneer lets the highest bidder win the bid, and pay the highest bidder's bid.

Some people say that this approach looks perfect, in line with the advertiser's demand for non-public offers, but also by many practices have proved, however, this way of bidding advertisers' bid strategy depends on how others out, but with their own real valuation has little to do with, and this applies to online advertising auction will have problems.

One feature of online advertising is the repeat game, what does the god horse mean? For example, offline project bidding is a one-time, bidders will be more inclined to be cautious about their bids, because he has only one chance, but online advertising is multiple times, that is, advertisers have multiple opportunities to constantly test other people's bids, in order to achieve their best strategy.

I give an example - Google's keyword "run" below the advertising bit, Nike thinks a click is worth 10 yuan, Adi thinks it is worth 6 yuan, this time if they take the "sealed first price" auction, and Google's advertising system gives the bottom price is 2 yuan, then Adi and Nike will try to bid.

Aldi started with 2 bucks, Nike out 2.1, Aldi 2.2, Nike 2.3, both sides have a continuous price increase process, until the increase to 6 pieces when Adi stopped bidding, because it thought the ad is worth up to six, Adi quit.

At this time only Nike, Nike is not stupid, since no one and I compete, then I for what to make 6 pieces, so hastened to 2 pieces, this time Adi will enter the field, the cycle began.

Carefully you'll see that there's an obvious flaw in this approach - the root of the instability lies in the fact that there is no Nash equilibrium from a game theory perspective (as economists have proved mathematically), that it always has a state of catch-up with me, because the auction bid depends on the opponent's bid.

At the same time, on a deeper level, this mechanism has at least two flaws in the mechanism of the repeated game:

First, it doesn't match Pareto's best, that is, a good auction mechanism should be to sell the subject matter to the highest-valued bidder, in this case, the highest-rated Nike, but half the odds are taken away by Aldi.

Second, it is not in the seller's interest to maximize, Nike Aldi evaluation is well over 2 pieces, but the bid starts from 2 blocks.

From this can be seen that this way of bidding is used in online advertising auction unreasonable, then there is a better way to bid wood, the answer is yes! Let's take a look at it next:

A change in winning the Nobel Prize

The economist William Vickery tried to solve this problem, and in his 1961 book Anti-Speculation, Auction and Competitive Sealed Tendering, a classic paper that systematically discussed the auction of "second sealed prices".

William Vickery, left.

It made a small change to the "first seal price" by still sealing the bid and still winning the auction, but the winner only needed to pay the second bidder, i.e. if Nike bids 10 and Aldi bid six, it would still be Nike to win, but Nike would only pay the second-place Aldi bid. 6 pieces.

It's a small change that's counterintuitive - the first place actually only pays the second bid, but don't look down on it, it's because it's a systematic discussion of the change that gave William Vickery the 1996 Nobel Prize in economics, and the "sealed second price" auction is also known in economics. Vickery auctions, "So what about this magical little change that hides the secret of the god horse?"

The answer is that it just systematically addresses the "first seal price" of major flaws. Some one would immediately ask: Why can this change overcome this defect?

Simply put, in the "sealed second price" bidding mechanism, everyone has a fixed optimal strategy - bid equal to their own valuation, or above the price strategy, Nike valuation of 10 yuan, Adi's bid how much do not know, this time Nike's optimal strategy is God horse? The answer is 10. Why?

We consider two scenarios:

First, if Aldi bids more than 10, then Nike can't win anyway, because Nike can't bid above its own valuation, higher than their own losses.

Second, if Aldi is below 10, then Nike should come up with the maximum price it can get to increase its chances of winning, and the maximum value is 10, i.e. Nike doesn't have the incentive to lower the bid because it doesn't have the control to finally pay the price.

What does the god horse mean? If Nike makes eight, then if Aldi makes six, Nike wins the bid and only pays 6, no different from it, but if Aldi makes nine it will lose the bid, so Nike doesn't have the incentive to move the bid to eight, which could lead to a failed bid, the only best strategy is to bid 10 bucks.

In game theory, "sealed second price" has a unique Nash equilibrium, that is, everyone out of their own real valuation of goods is the best strategy, so the mechanism is a mechanism to encourage bidders to tell the truth, and has considerable stability, that is, advertisers do not frequently adjust their bid motivation.

Under this mechanism, advertisers who have the highest ratings for ad space win, while ensuring the platform's revenue, and no advertiser selling lower bids than their own. In fact, Google, Baidu, Sina Weibo and other advertising platforms are taking this auction.

Well, would someone say, is this bidding method the perfect way to bid? In fact, not necessarily, "sealed second bid" mechanism at least one flaw, that is, its anti-cheat characteristics are not strong, if there is collusion between accomplices, in this bidding mechanism, the complicity is easier to achieve. What does the god horse mean?

Still take Nike Adi to play metaphor, Nike's psychological bid is 10 yuan, Adi is 6, this time they colluded, Adi out of 1 piece, Nike out of 10, finally Nike only need to spend 1 yuan to buy this advertising space, the injured is the advertising platform. Smart people immediately say - isn't that the "first seal price" going to happen? For example, Nike and Aldi to discuss good, Nike out of 1 piece, Aldi out 0.5 yuan, or Nike to buy this advertising space for 1 yuan!

Haha, that's right, but the "first seal price" is more likely to be a case of betrayal accomplices, although the good Nike out of 1 piece, Adi out 0.5 hair, but Adi is motivated to violate the accomplices, as long as Adi out of 2 pieces, it can win the auction, this time Nike is silly eye.

But in the case of the "second seal price", Nike out of 10, Aldi out of 1 piece, this alliance is very powerful, because Aldi any way to betray, it is unlikely to win the auction (it can not bid more than 6), so it has no incentive to betray, so the accomplice is easier to achieve, the probability of cheating will be greater.

Well, since there is this flaw, then now online advertising Google, Baidu why use it? An important reason is that, unlike the single auction offline, online advertising is a large-scale repeat game, large-scale means that the number of advertisers participating in the bidding is numerous, objectively increased the difficulty of colluding to cheat, in a sense to cover up the shortcomings of this mechanism.

VCG, a multi-advertising auction mechanism

Above we discussed the basic principles of the bidding mechanism, all the examples are an ad bit situation, and in the real advertising system, a request for advertising is often more than one, such as search engines have multiple advertising places, how should we set up the bidding mechanism?

Google, Baidu on the "sealed second price" auction carried out an extension, that is, if there are multiple advertising space, the first by the second plus a minimum bid unit deduction (such as 0.01 yuan), the second by the third charge, the third by the fourth charge, and so on, this kind of bidding method is called "broad second price auction" Generalized Second Price Auction, or GSP for short.

GSP Broad Second Price Auction

This approach maximizes the advantage of "sealing the second price", i.e. it creates a stable equilibrium, but it also has a disadvantage that it is not a way to maximize the benefits of all bidders.

Thus, three economists Vickrey, Clarke, and Groves in three papers put forward a multi-item auction mechanism, referred to as "VCG auction", this complex bidding mechanism from the overall interest of the entire bidder, it is still the high price, However, the deduction is the total loss to the other bidders that the high bidder participates in, i.e. the total benefits of the non-high bidder's participation are calculated first, and then the total benefits of the others after the participation of the high-priced person is calculated, the difference between the benefits is the loss of the other participants.

In short, you participate in the bidding, you cause losses to other bidders, you need to pay for the overall benefit of the system to reduce the cost to ensure that the overall benefits of the maximization.

This bidding method is relatively complex, and I use a simple example to illustrate - if there are now two ad bits, that is, search for the "run" keyword first and second place, the first can bring 20 clicks, the second can bring 10 clicks.

There are three advertisers bidding, Nike pays six per click, Aldi's 4 and Reebok's 2. Nike won the first ad spot, Aldi won the second ad position, and Reebok failed to bid.

The question is: How much should Nike be charged? According to the VCG Bidding Law:

  • First calculate the earnings of Aldi and Reebok when there is no Nike participation in the bidding, i.e. Adi wins the first ad position, Reebok wins the second ad position, the revenue is 4 x 20 x 10 x 2 x 100

  • Then calculate the proceeds of Nike's participation in the bidding of Aldi and Reebok, i.e. Aldi won the second ad position, Reebok out, the return of 4 x 10 plus 2 x 0 x 40.

  • The difference between the benefits of the two is 100-40 x 60

  • Then Nike should pay 60/20 for each click.

Facebook's advertising system is taking this bidding method, which maximizes the benefits of the bidding participants, but we can see that the interests of the auctioneer is not maximized, in the above case Nike's deduction of 3 pieces than the GSP bid rules under the second-place charge of 4 less, it can be said that, Facebook is sacrificing short-term interests and taking a longer-term view, because bid ads are not short-term behavior, and Facebook believes it is in its long-term interest to ensure that advertisers' overall interests are in the long term.

So why doesn't Google do the same VCG bid, partly because while VCG maximizes the benefits of bidders, it's very difficult to explain to advertisers the huge educational costs; Varian has made it clear that one of the reasons Google was reluctant to move from GSP to VCG in 2002 was that the cost of user education was too high.

On the other hand, economists have shown that VCG's revenues are no higher than GSP's, as it runs the risk of falling revenues directly from GSPs to VCGs.

This article mainly introduces the story of the bidding ad mechanism design, in fact, the bidding mechanism selection is only a small aspect of the bidding advertising, about other aspects of the bidding advertising, I will be introduced to you in other articles in the following.

You should also note that Weixi's article has been committed not only to tell you "what" but more importantly to tell you "why", knowledge itself is very important, the way to express knowledge is also very important, pay attention to "Weixi chat advertising", more easily understand the underlying logic of Internet commerce.