Bidding Models in Internet Advertising - Alliance, RTB, RTA

In (Part 1 - Basic Bidding) and (Part 2 - Smart Bidding Model), we discussed the "four points, three rates, two controls and one enhancement" analysis model of the bidding model . The " four points " are the bidding point, billing point, bid point, and assessment point. The differences in various bidding modes are largely caused by the different positions of these four points.

For example, the following figure shows the positional differences of four points in CPC and oCPC:

The “ three rates ” are the ratios between four points, namely p (bidding point->billing point), p (billing point->bid point), and p (bid point->assessment point). In (Part 1), we analyzed one by one the impact of the true value of the "three rates" and the accuracy of the estimated values ​​on the interests of media platforms and advertisers on large media platforms. As shown in the following figure:

In (Part 2), we discussed the " two controls ", namely cost control under continuous bidding and budget control under continuous bidding. The two biggest features of the oCPX bidding model are (a) the separation of billing points and bidding points and (b) intelligent bidding control under continuous bidding.

We then explored “ one enhancement ”, which is when the amount of data is insufficient to accurately estimate the ratio after the bid point, using the only data available to enhance the value implicitly estimated by the advertiser with a fixed value, which can optimize the behavioral cost after the bid point and further improve the effect.

Finally, we used a complete “four points, three rates, two controls, and one reinforcement” analysis model to analyze the main smart bidding models of mainstream media in the industry.

In this article, we will introduce ad networks, RTB and RTA advertising , and discuss some changes in the interests of all parties.

· The article is coming ·

1. Advertising Alliance

In large media platforms, media and platforms are in the same community of interests. We can think of the large media platform as being divided into three parts: large media advertising space , traffic aggregation and delivery platform .

However, not all media have the technical capabilities to build an advertising platform, nor do all media have the sales capabilities to connect with many advertisers. These small media need to share an advertising delivery platform, that is, an advertising alliance . The various bidding models introduced above can also be implemented on the advertising alliance's delivery platform, and the "four points, three rates, two controls and one enhancement" also apply.

From the figure we can see that the biggest difference between the advertising alliance platform and the large media platform is that the media has been separated from the community of interests with the delivery platform and traffic aggregation. The interests between the media and the delivery platform are no longer completely bound. So what changes will this bring? Let’s see what changes have occurred in the “three rates” table.

In the table, we refer to advertising networks as “ platforms ” for short. All differences from the tables corresponding to the major media platforms are circled in yellow.

Because the alliance shares profits with the media in a certain ratio, the interests of the media and the platform are basically the same (although both parties have to bargain based on their own chips on the profit sharing ratio), so the above table is actually basically the same as the table in the large media platform, mainly except that "media platform" has become "media & platform", and p (bidding point -> billing point) and p (billing point -> bid point) were originally estimated by the "media platform", but are now estimated by the "platform".

Because the alliance platform does not have all the behavioral data of users in large media (users' advertising behavior data and non-advertising behavior data), but only has the advertising behavior data of users in numerous small media (no non-advertising behavior data), whether it can make more accurate estimates than the delivery system of large media depends on which of these two types of data contains more information.

So does this mean that there is no conflict of interest between the media and platforms under the alliance model? Not really.

▶ Why does a certain owned traffic only have oCPM while another Shanjia still has the oCPC model to choose from?

On a certain platform, when you choose to place an ad on a certain free traffic, you can only choose the oCPM mode. However, when you choose to place your ad on a certain ad network, you can choose oCPM or oCPC. Why not just unify them all and keep only oCPM?

In the alliance model, one difference is that when p (billing point -> bidding point) is too high, the platform will compensate the advertiser, because it sees whether the overall cost of the advertiser after multiple bids (such as activation cost) exceeds the cost of the advertiser's bid, so the platform cannot attribute the compensation to the specific media.

In addition, the platform has to consider long-term benefits, and has no incentive to overestimate p (billing point -> bidding point). However, since individual media do not have to pay compensation and do not have to consider long-term benefits, they have the motivation to actively make the platform's estimate of p (billing point -> bidding point) higher.

For example, in the oCPM model, by adding a large number of false impressions, the platform model will not react for a while, and will overestimate the p (billing point -> bidding point) of the media in the short term. (The long-term model will learn that the p (billing point -> bidding point) of this media is low, thereby reducing the bid). The compensation will not necessarily be attributed to the media, which will result in the media making a lot of cheating profits due to false presentations.

However, if under oCPC, the m->c that the media can control becomes p (bidding point->billing point), whether the ratio is estimated to be too high or too low will affect the income of a single media, so a single media will have no motivation to create false impressions.

Unfortunately, from c->bid point, the media can still achieve the effect of making p (billing point->bid point) higher by creating false clicks (which is more difficult than creating false impressions), thereby obtaining cheating profits. So as we can observe, the quality of traffic on the alliance is slightly worse.

2. RTB Advertising

As mentioned earlier, the advertising alliance separates the media from the overall interest community of the delivery platform and traffic aggregation. The RTB model then separates the delivery platform from the interest community of traffic aggregation, allowing the delivery platform to be more closely aligned with the interests of advertisers. There are three main ways to combine interests, corresponding to the three profit models of DSP: arbitrage model, service fee model, and consumption sharing model . They will be introduced one by one below.

The starting point for doing this is to give advertisers better choices. For example, in the past, all ingredients (media traffic) were only supplied to one restaurant. If you wanted to eat, you could only go to this restaurant (one delivery platform), and you had to endure it no matter how bad the service was. Now the ingredients are supplied to many restaurants (multiple DSPs) at the same time, so you can choose the restaurant with the best service. Because of the competition, restaurants will be more willing to stand from the consumer's perspective and consider things for the consumer. If you feel that the food in the restaurant does not suit your taste, you can buy it back and cook it yourself in the kitchen (the advertiser operates its own DSP).

The reason why DSP is called Demand Side Platform is that it is a platform that stands more on the demand side, better represents the interests of the demand side (advertisers), and allows some advertisers to feel more at ease uploading their data for delivery.

▶ How many profit models does DSP have? What changes have occurred in the interests of the three parties under each model?

There are three profit models for DSP:

1. Arbitrage model

Advertisers and DSPs agree on a purchase price for a certain number of actions, such as 1 yuan per click. Then DSP will purchase traffic through ADX. DSP will use technical means to control the price of each click to, for example, 0.8 yuan. At this time, DSP can earn a difference of 0.2 yuan. The stronger the DSP technology is, the lower the cost can be used to purchase clicks from ADX for the same amount, for example, if the cost is reduced to 0.6 yuan, higher profits can be obtained.

Because activation or other deeper behavioral data is typically collected by advertisers, DSPs cannot prevent advertisers from concealing activations. Therefore, the settlement behavior agreed upon by advertisers and DSP is usually similar to the billing points in large media platforms and needs to be within the scope that DSP can fully control.

Advertisers and DSPs settle accounts based on clicks or impressions, but they can also implement an oCPX model similar to large media platforms, where advertisers can bid on the DSP platform based on activation or payment.

For example, an advertiser and DSP settle at 1 yuan per click, but at the same time bid 60 yuan for each activation. Like the oCPX introduced earlier, DSP will convert the activation bid to the click bid based on the estimated p(c->a), and then convert it to eCPM based on p(m->c). At the same time, through cost control, the click cost can be controlled at, for example, 0.8 yuan, so DSP can earn 0.2 yuan per click, which is why it is called arbitrage. The more accurate DSP's estimate of click-through rate is, the lower the cost can be used to buy clicks of the same quality, thus increasing profits.

In addition, the activation cost for advertisers also needs to be controlled below 60 yuan. When I designed the bidding strategy for the first mobile DSP in China in 2014, I adopted this model. Advertisers bid based on CPA, and DSP used technical means to ensure that the cost per click was lower than the settlement cost with advertisers to ensure profit margins.

Under the arbitrage model, because the way DSP makes profits is different from the original alliances and major media, the "three rates" table has also changed. In this table, we call DSP a platform, and ADX and aggregated media traffic are called media.

The only difference from the alliance is that the p (bidding point -> billing point) estimate is a bit high. If the DSP's estimate from the bidding point to the billing point is too high, the media's revenue will increase (because the bid is too high). However, because advertisers and DSPs are settled based on the number of billing points, it does not affect the ROI. This part of the loss is borne by the platform (that is, the DSP).

2. Advertiser-operated DSP/service fee model

Advertisers in some industries, such as e-commerce advertisers (such as JD.com and JD.com), cannot send the data collected by the advertiser side, such as when their detail page arrives or when users purchase products, back to the platform (whether it is large media advertising, alliances or DSP) because the value of user data is too great. Therefore, they cannot use products such as oCPX, but they are unwilling to use only CPC to purchase traffic, so that the back-end link is not optimized at all. At this time, self-operated DSP is a relatively suitable solution for them.

Some DSPs adopt a fixed service fee approach, that is, advertisers pay a sum of money, and the DSP will concentrate on serving the advertisers. The profit of the DSP has no direct relationship with any number of behaviors. This approach is very close to the model where advertisers build their own DSP from the perspective of interest, but like other models, it may not necessarily use all of the advertiser's private data.

Because no matter whether the advertiser operates its own DSP or the DSP charges a fixed service fee, the interests of the advertiser and the platform are completely aligned, and there is no distinction between the platform and the advertiser. Therefore, in the above table, "advertiser" is directly used to represent the common interests of both parties.

In addition, because there is no billing behavior between advertisers and platforms, the billing point among the four key points disappears, and the key ratio is reduced from 3 to 2.

Because p(billing point->bid point) or p(bidding point->bid point) is estimated by the advertiser himself using the model, the effectiveness of the effect is directly related to the data owned by the advertiser himself. Therefore, this model is only suitable for advertisers who have rich data (for example, they are also advertisers of major media, such as JD.com and Toutiao). This will be analyzed in detail later.

3. Consumption sharing model

Some DSPs adopt a consumption-sharing model, that is, they share the budget spent by advertisers according to an agreed ratio.

Because there is no charging based on the number of behaviors between advertisers and platforms, the billing point among the 4 key points disappears, and the key ratio is reduced from 3 to 2.

Under the consumption-sharing model, DSP's revenue is proportional to advertisers' consumption, which is proportional to the media's revenue, making DSP's interests tied to the media. However, this model still exists because this table only analyzes short-term interests. In the long run, if the advertiser's ROI decreases, they will terminate their cooperation with DSP and switch to other DSPs. As DSP needs to represent advertisers more than alliances, it is unlikely to do anything that harms advertisers' interests in the short term.

3. RTA

When introducing advertisers’ self-operated DSPs, we mentioned that some advertisers cannot send back the data collected by the advertisers, such as when their detail pages are reached or when users purchase products, to the platform (whether it is a large media platform, alliance or DSP) because the value of their own user data is too great. Therefore, they set up their own DSPs so that they can safely use their own data to train models.

So, for advertisers who buy traffic from large media platforms or alliances, can they also not send data back to the platform, but instead build their own delivery system like advertisers’ own DSPs? The answer is yes. This method is called RTA advertising in some channels and some websites. It can be considered that the advertiser's self-operated DSP under the RTB model is also a special type of RTA advertising. The following figure shows a schematic diagram of RTA advertising under large media platforms, alliances, and RTB (that is, self-operated DSP).

In the figure, the orange parts are part of the delivery system, and models are used to estimate various ratios. We can see that in large media platforms and alliances, the delivery system is divided into two parts, one part in the large media or alliance platform, and the other part in the advertiser's own RTA delivery system. It is not difficult to understand that if the behavior that a large media platform or alliance can fully collect is clicks, then p(m->c) is estimated by the delivery system of the large media or alliance platform, while p(c->bid point) is estimated by the advertiser's own RTA delivery system.

This division is not based on the billing point, so sometimes large media or alliance platforms estimate p (bidding point -> billing point) (such as oCPC), and sometimes it may be part of p (billing point -> bid point) (such as oCPM), while the advertiser's RTA system estimates all or part of p (billing point -> bid point).

So for advertisers, there are six advertising modes (3 ordinary ones plus 3 RTA). Which of these six modes has a better delivery effect? The implementation of models and algorithms depends on the technical capabilities of each company. Let’s take a look at the essential differences between these six models in terms of data sources.

In the above table, assuming that the behavior that large media platforms or alliances can fully collect is clicks, then there are two models that need to be estimated, one is p(m->c) and the other is p(c->bid point). The data sources of each model are divided into features and labels. Only when both the feature data and label data are good can the final model make better estimates.

In addition, the blue background indicates that this model is estimated by large media, alliances or non-self-operated DSPs, and the orange background indicates that this model is estimated by the advertiser's own RTA delivery system or self-operated DSP. In addition, all participants in the above table can purchase third-party data, so third-party data is no longer listed separately in the table.

Big media vs alliances: The data that the alliance can obtain is limited to users' advertising behavior, but there are data from more media. Therefore, if the alliance is not large enough and the advantage of the number of media cannot make up for the lack of variety, then the estimates of the two models will not be as accurate as big media.

DSP vs Alliance: First of all, ADX will only send local requests to DSP, so in terms of feature dimensions, DSP can only obtain user advertising behaviors of some small media. In addition, for the p(m->c) model, DSP can only get feedback on whether the successful bidding impressions were subsequently clicked, so the label is much worse than that of the alliance.

Therefore, if DSP wants to play its own advantages, it must rely on advertisers' willingness to send it more in-depth data (such as activation, payment, etc.). Either it is the advertiser itself (self-operated DSP), or it is also a large media (such as Baidu's DSP, Tencent's DSP). It is difficult to maintain a third-party DSP without data.

"Big media and alliances" vs. "Big media + RTA and alliances + RTA": In RTA advertising, big media and alliances transmit user characteristics to advertisers' RTA delivery systems on a purely voluntary basis, and advertisers have no control over how much is transmitted. Some big media or alliances do not transmit characteristic data at all. Although label data has an absolute advantage for RTA in the p(c->bid point) model, this model cannot be trained if the advertiser does not have a large number of its own user features.

If an advertiser has the characteristics of a current user, it means that the user is already a user of a product of the advertiser. Therefore, RTA is not suitable for acquiring new users, but is more suitable for placing advertisements of the advertiser's other product B to users of a certain product A under the advertiser. Therefore, RTA is particularly suitable for e-commerce advertising, because most users have actually purchased things from e-commerce advertisers. E-commerce advertisers already have the characteristic data of these users and can recommend advertisements for new products that they have not purchased to these users. RTA is not very suitable for game advertising because game advertisers usually need to buy new users who have never used any of their games. For these new users, advertisers do not have any characteristic data.

To summarize: advertisers who have accumulated a large amount of data and repeatedly recommend different products to users (such as e-commerce advertisers), or advertisers who absolutely cannot transmit data to the platform due to some special considerations, can consider buying traffic from large media or alliances through RTA, or buying traffic from ADX through self-operated DSP. In fact, these two methods are more commonly used by e-commerce companies (such as JD.com) on the market. For other advertisers, it may be best to send the data back to large media or alliances for buying traffic.

【Summarize】

This series of articles introduces the "four points, three rates, two controls and one reinforcement" analysis model, and analyzes the interests of multiple parties under RTB advertising (and RTA) under three different DSP profit models: big media, alliances, and RTB advertising. This analysis model is also used to analyze the mainstream bidding patterns in the market.

The three articles are quite long, and I would like to express my gratitude to the readers who are truly passionate about computational advertising and have the patience to read them all. This article also received help from several friends to review it, thank you again. Some readers also added WeChat and sent me feedback and affirmation, and I would like to thank them here.

Author: Shentanshe

Come to: Shentanshe