In-depth analysis of bidding models in Internet advertising

In (Previous article - Basic bidding model), we discussed the "four points and three rates" in the "four points, three rates, two controls and one reinforcement" analysis model of the bidding model. The "four points" are bidding point, billing point, bidding 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 the four points in CPM and CPC.

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 figure below.

At the end of the previous article, we also introduced the oCPX bidding model. The difference between it and basic bidding models such as CPM/CPC/CPA is that it realizes the separation of billing points and bidding points, and also realizes intelligent bidding control under continuous bidding.

In this article, we will start with smart bidding control, namely the “two controls”, and then introduce the “one enhancement”. Then we can use the complete "four points, three rates, two controls and one strengthening" analysis model to analyze the industry's mainstream smart bidding modes such as oCPX, double bidding, AEO, VO, etc.

· The article is coming ·

6. “Two Controls”: Intelligent Bidding Control under Continuous Bidding

The above discussion on the "four point three rate" is based on the analysis of each bid individually. Under the most commonly used GSP bidding mechanism, for a single bid, advertisers only need to bid the highest price they can accept to maximize their profits. For example, after calculation, the advertiser finds that the maximum amount they can afford is 800 yuan per paying user. According to historical experience, the activation to payment ratio is about 0.1, so the activation bid should be 80 yuan.

However, because the placement is a continuous multiple-auction behavior, on the one hand, the price given by the advertiser is the average price of multiple bids, and the results of the early auctions will affect the maximum price that the advertiser can accept later (for example, if you buy it cheaply in the early stage, you can bid higher later). On the other hand, some advertisers also have budget constraints. Because of these two reasons, it is necessary to adjust each bid based on the results of past auctions. For each cause there is a control.

1. Cost control under continuous bidding:

Still assuming that the advertiser's activation bid is 80 yuan, if the 100 activations purchased through the early delivery only cost 50 yuan each, then you can actually bid higher later, as long as the final comprehensive cost does not exceed 80 yuan. Increasing the bid can also buy more volume.

Similarly, if the average cost of the previous campaign reaches 120 yuan, then you need to limit the bid so that the final comprehensive average cost drops to around 80 yuan. oCPX usually means that the media platform controls the cost of the bids at the bid points under continuous bidding, so that the final average cost is close to the value set by the advertiser.

2. Budget control under continuous bidding:

This control is for situations where advertisers have daily or weekly budget limits. For example, the daily budget is 10,000 yuan, and 5,000 yuan has been spent in the morning. Then in the subsequent auctions, you should lower your bid to prevent the budget from being spent too early. If you spend your budget too early, you won’t have the budget to purchase some very cheap traffic later on, and you won’t minimize your average cost.

To reduce the average cost of each behavior (such as the average cost of activation) under a limited budget, the budget must be spent at a reasonable pace. It’s not about an absolute average spend (spending 1/24 every hour), but about making the spend and traffic distribution roughly consistent (for example, buying more at night when traffic is high).

If you want to do better, you can buy more when traffic is cheaper and buy less when it is expensive. This is where our technical prowess is demonstrated. It is worth noting that in some bidding products, when budget control is set, advertisers do not need to fill in the bid points in the delivery background, but the media platform will set and adjust them entirely according to the budget.

These two are the two controls in our “four points, three rates, two controls and one strengthening”.

Now basically all media platforms’ bidding models, as long as they include words like “smart bidding”, mean that they have at least one of these two controls. It is worth mentioning that in 2014, when I was working at a startup company to build the first mobile DSP platform in China, I separated the bidding point and the billing point, and innovatively used PID control to implement these two intelligent bidding control functions, leading many large companies in bidding products (although these products have not been used by many people...)

One thing worth noting is that the name oCPC does not indicate where the bid point is. Therefore, a bidding pattern cannot be uniquely identified. In this article, we can agree to add a bid point letter at the end to uniquely identify a bidding mode. For example, oCPC-A, oCPM-I, etc. But why aren’t media platforms named this way? Because the media platform actually has no way to guarantee that the advertiser will send back action a or something else. Suppose the media platform wants to use the oCPC-A model, but the advertiser can actually return not a but paid p, and fill in a paid bid in the bid (originally an activated bid should be filled in). At this time, oCPC-A becomes oCPC-P.

▶ In the oCPX bidding model, should advertisers send back real behavioral data?

According to the above discussion, the behavioral data of the bid points are sent back by the advertisers themselves. There are several ways for advertisers not to send back all the bid point data as required by the media platform. Let’s discuss whether it is beneficial to advertisers in turn.

1) The advertiser only sends back part of the bid data, for example, only 50%

In this case, the advertiser's p (billing point -> bidding point) is estimated to be nearly half as small, causing the advertiser to bid too low and not get enough volume. To get the same amount, you have to increase your bid to twice the original amount. With this offset, the bid at the billing point remains unchanged, and the media's revenue does not decrease. On the contrary, because half of the samples are missing, the estimate of p(billing point->bidding point) becomes more inaccurate, which damages the advertiser's own ROI. In China, attribution is done by advertisers themselves. Some advertisers think that being as conservative as possible when attributing to media platforms will be beneficial to them, but this is actually not true.

2) The advertiser returns deeper (or shallower) data instead of the bid point data, such as returning payment data (or downloads) instead of the originally agreed activation data.

In this case, if the advertiser also fills in the paid bid (or download) in the bid, then for this advertiser, oCPC-A becomes oCPC-P or oCPC-I. Unfortunately, if most other advertisers still send back activation numbers, then because the media platform does not only use the advertiser's own data when training the p (billing point -> bidding point) model, but is also affected by the data of other advertisers, then doing so will have a negative impact on the estimated accuracy of the advertiser's p (billing point -> bidding point) due to inconsistent training target types, which will be detrimental to the advertiser's delivery.

Therefore, in oCPX bidding, it is actually in the best interest of advertisers to follow the agreement and fully transmit the agreed behavioral data.

▶ Why do many domestic oCPXs have over-cost compensation mechanisms?

This question needs to start with the comparison of CPC vs oCPC. We use the "four point three rate" to compare their differences.

Comparing CPC and oCPC, we only found one difference, which is the difference in who estimates p(c->a). However, this will also lead to a very different point, that is, in CPC, p(c->a) is estimated by the advertiser, and if the estimate is too high or too low, the advertiser’s interests will be damaged. However, in oCPC, p(c->a) is estimated by the media platform. If the estimate is too high, the media platform’s income will increase, and if it is too low, the media platform’s income will decrease. In theory, then, media platforms have an incentive to deliberately make higher estimates in the short term (they won’t make higher estimates in the long term because when advertising ROI drops, they will reduce investment, ultimately damaging the interests of the media platforms). This is also the question raised at the end of the previous article.

Therefore, domestic advertising platforms such as Tik Tok and Kuaishou will have an over-cost compensation mechanism in oCPC (or oCPM, the principle is the same) to self-constrain and make up for the problems in this mechanism.

For example, if an advertiser uses an activation bid of 80 yuan per conversion, if the media platform deliberately makes a high estimate in the short term, causing the advertiser's activation cost to exceed 80 yuan, the media platform will make compensation. In foreign countries, FB and Google's advertising products do not have this compensation. I personally believe that this is because the trust mechanism abroad is better. Advertisers believe that large media platforms will consider long-term interests. Even if there are occasional over-costs, they believe that it is a technical reason and is part of the cost of delivery. However, due to historical reasons, it is difficult to establish such a trust mechanism in China, so media platforms need to tie their own hands to prove their innocence.

▶ Why do domestic media platforms require that the number of accumulated behaviors (such as activations) of advertisements reach a certain number before they start paying for the over-cost compensation of oCPX?

Douyin's oCPM compensation requires that the ad unit accumulate 10 behaviors (such as activations), and Kuaishou has similar requirements. Why can’t we just compensate for the excess costs?

There are two reasons:

The first reason is that if there is no threshold for the number of behaviors (taking activation numbers as an example below), then advertisers may hide the number of activations and not report them, and finally ask the media platform to make full compensation (because there is not a single activation). Because advertisers do not report activation numbers, the media model will estimate this ad unit lower and lower, eventually to 0, which means that the media platform will give this advertiser less and less, and eventually no amount at all. However, in the early days, advertisers were still able to get a lot of clicks or activations (but they were not reported to the media platforms).

If the media platform does not set an activation number threshold and pays full compensation, it will continue to be exploited by advertisers. Once the activation number threshold is set, the advertiser will need to pay at least the activation bid * activation number threshold. Therefore, as long as the cost of the media platform completely stops providing content to advertisers and is less than the advertiser's minimum payment cost mentioned above, the interests of the media platform will not be harmed.

The second reason is that model learning requires a certain number of samples. Before a sufficient number of activations are accumulated, it is difficult to avoid model estimates being too high or too low. This is not a deliberate behavior of the media platform, and the media platform should not unilaterally bear this part of the cost.

▶ Why should a certain owned traffic be converted from oCPC mode to oCPM?

The sale of a certain piece of owned traffic was initially based on the oCPC model, but later changed to only oCPM.

Let’s use the difference of “4.3 rate” to see what is the difference between oCPC vs oCPM?

We can see that there are two changes from oCPC to oCPM:

The first change is in the absolute value of the ratio. Advertisers who do not seek volume originally had an incentive not to increase the ratio of p (bidding point -> billing point). Under the oCPM model, the incentives of both parties are aligned, which is an improvement.

The second change is that under the oCPC model, the media platform’s revenue would decrease regardless of whether the estimate was too high or too low. However, under oCPM, there is an incentive to estimate too high, which will damage the advertiser’s ROI. However, this problem can be compensated by the compensation mechanism.

So on the whole, the transition from oCPC to oCPM does not cause any harm to the interests of advertisers. For media platforms, it can avoid the problem of advertisers who are not looking for volume actively lowering p (bidding point -> billing point) (although this problem has little impact on fully competitive traffic). This may be one of the reasons for a certain switch from oCPC to oCPM.

In addition, there are many products in a certain article, including text ads, picture ads, and video ads, so the conversion chain is also diverse. Some charge by click, some charge by 3 seconds of playback, and some charge after the playback is completed. If m billing is used uniformly, it will also be conducive to unified management. However, FB and Google's advertising products still allow advertisers to choose between oCPM and oCPC, and the difference between the two is actually not that big.

▶ What kind of bidding method is double bidding?

As shown in the figure, taking activation paid double bidding as an example, this bidding model has two different bidding points. Advertisers provide both of these behavioral data to the media platform, allowing the media platform to provide free bidding conversion services. When it is finally converted to the billing point, there will also be two billing point prices. These two prices are conflicting, and the final price must be obtained by combining the two prices. The specific implementation depends on the design of each media.

However, what is certain is that each media platform will strive to ensure that the costs of activation and payment do not exceed the set activation and payment bids. So what is the difference between this approach and paying for a single bid with guaranteed costs (such as the "activate and pay" bidding model of a certain item)? Let's analyze it.

• If the activation cost and payment cost of a batch of traffic meet the requirements, then both double bidding and single payment bidding can be used to obtain this batch of traffic.
• If for a batch of traffic, the activation cost meets the standard but the payment cost does not meet the standard. Then this batch of traffic cannot be obtained through double bidding or single paid bidding.
• If the activation cost and payment cost of a batch of traffic do not meet the requirements, then neither the double bid nor the single payment bid will be able to obtain this batch of traffic.
• If the activation cost of a batch of traffic does not meet the standard but the payment cost does, then this batch of traffic cannot be obtained by double bidding but can be obtained by single payment bidding.

Therefore, the difference between activation-paid double bidding and single-paid bidding lies in whether to include the traffic where “activation cost does not meet the standard but payment cost meets the standard”.

Advertisers are more concerned about the payment cost. Advertisers need the traffic where “activation cost does not meet the standard but payment cost meets the standard”, so why is double bidding necessary?

The author believes that the reason is that there is less paid data in the early stages, and the paid cost will experience greater fluctuations before converging to the paid bid. At some point it will even far exceed the paid bid, resulting in failure in the learning period or advertisers unable to bear it and shutting down themselves.

Double bidding, by giving up the traffic with "activation cost not meeting the standard, but payment cost meeting the standard" (actually giving up all the traffic with "activation cost not meeting the standard", because in the early stage it was impossible to distinguish the traffic with "activation cost not meeting the standard, but payment cost meeting the standard" from the traffic with "activation cost not meeting the standard, but payment cost also not meeting the standard", so we had to give up all of them at once), can reduce the fluctuation of the initial payment cost. In the end, it may lead to more stable learning, higher success rate and lower comprehensive cost for the whole time period.

However, in the later stage, enough payment data has been accumulated to distinguish between the traffic with "activation cost not meeting the standard but payment cost meeting the standard" and the traffic with "activation cost not meeting the standard and payment cost not meeting the standard". Then giving up the traffic with “activation cost not meeting the standard but payment cost meeting the standard” will limit the amount that advertisers can buy. Therefore, for advertisements that have accumulated a large amount of payment data and have small fluctuations in initial payment costs, single-payment bidding may be a more appropriate option.

7. “One Strengthening”

The “four points, three rates and two controls” have all been discussed, and the only thing left is “one strengthening”. Let’s take a Facebook product as an example.

On the Facebook app promotion delivery settings page, we can see that the billing point can be selected as impression, the bidding point can be selected as conversion (for example, it can be activation a), and you can also choose to optimize for a certain in-app event (Optimiaztion for Ad Delivery App Events). This optimization can be a certain behavior after activation, such as payment p. Then the delivery system will try to optimize to obtain a higher payment number based on oCPM-A. Let's call this the oCPM-A-enhancedP model.

In this way, a new point, the strengthening point, is added between the bidding point and the assessment point. Assume the bid point is at a and the activation point is at p. So the question is, why don’t media platforms directly let advertisers bid at the enhanced point p at the back end, and then convert the price of the enhanced point p into eCPM by estimating p (a->p) and several other rates? If so, it is equivalent to bidding at p, without the enhanced point. Why not do this?

To review the previous content, separating the bidding point from the billing point and placing it closer to the back end of the link can bring better results, provided that the media platform's p(billing point->bidding point) is more accurate than the p(billing point->bidding point) implicitly estimated by the advertiser using a fixed value. But sometimes, due to less data or other reasons, the ratio p(a->p) estimated by the media platform using data is more inaccurate than the value implicitly estimated by the advertiser using a fixed value. Then this section cannot be placed in the billing point -> bidding point. This is why the bid point cannot be moved back to the strengthening point.

However, the data from the media platform cannot be wasted. It can be used to enhance the value implicitly estimated by the advertiser with a fixed value, that is:

p_adj(bid point->enhancement point) = p_advertiser estimate(bid point->enhancement point) * adj(x) — (Formula 1)

Where x is the feature vector of the advertisement.

Intuitively, the media platform determines based on the data that the p (bid point -> enhancement point) of the advertisement is higher than that of ordinary advertisements, so it will raise the bid. If the p (bid point -> enhancement point) is lower than that of ordinary advertisements, it will lower the bid. This price adjustment usually has a range. For example, the public document states that the "automatic optimization" of a certain item is controlled between -30% and +30%, that is, the range of adj(x) is between 0.7 and 1.3. As long as the accuracy of p_adj (bid point->enhancement point) is greater than p_advertiser estimate (bid point->enhancement point), then adding this enhancement point will be profitable.

The media platform cannot directly obtain the value of p_advertiser estimate (bid point -> enhancement point) because this value is implicit in the bid. But fortunately, it is not necessary. We record the ecpm adjusted for the reinforcement point as ecpm_adj, then

ecpm_adj = p(bidding point->offer point)*p_adj(offer point->enhancement point)*enhancement point bid

(Formula 1) and

Enhanced point bid = Bid ​​point bid / p_advertiser estimate (Bid point->Enhancement point)

Substituting into the above ecpm_adj formula, we have

ecpm_adj = p(bidding point->bid point)*p_advertiser estimate(bid point->enhancement point)

*adj(x) *(bid point bid / p_advertiser estimate (bid point -> enhancement point))

After the appointment

ecpm_adj= p(bidding point->offer point)*adj(x)*offer point bid

Just use this value as the ecpm for bidding point sorting.

From this we also know why in (Equation 1) adj(x) is simply multiplied by p_advertiser estimate (bid point -> enhancement point), because if it is a more complex function, the above derivation may not hold.

In addition to Facebook, Google's enhanced CPC is also a product of this type based on the document description. In addition, the "automatic optimization" function of a certain article is also a product that uses payment as an enhancement point.

8. Summary of “Four Points, Three Rates, Two Controls, and One Strengthening”

After analysis, we will find that most bidding models are composed of "four points, three rates, two controls and one strengthening". When encountering a new bidding strategy, we can split them according to these different parts for easy understanding. We summarize by analyzing the bidding patterns of several mainstream companies on the market. Because the assessment points are different for each advertiser and have little to do with the delivery platform, they are not listed.

(The information in the following table is from the public backend or help documents of each delivery platform)

In the table, app events can be various app events such as download, installation, activation, payment, etc. but cannot be the amount of money collected. The "v" in Facebook's billing point means that the video has been watched for 10 seconds or completed. In addition, in Facebook's products, when app events select different behaviors, the selection of billing points and enhancement points is limited, and not all can be selected.

There is no confirmed evidence about the m/c of ​​Google's billing point, it's just my guess. Tencent uses oCPA to represent the bidding model with the bidding point at a and the billing point at m or c. The naming method is different, so it is easy to be confused with the models of other companies, so be careful to distinguish them.

In addition, for all the bidding modes with budget control in the table, advertisers do not need to fill in the bid points in the delivery background, but are completely set and adjusted by the media platform according to the budget. Minimize the cost of your bid while trying to spend your budget.

In each company's product, each model can also be matched with several different consumption speeds. For example, one of them is (balanced delivery, priority on volume, priority on low cost). The specific strategies of each company are different. I guess that different pacing strategies and parameters are matched in the specific implementation of cost control and budget control.

Facebook and Google also have two consumption speeds to choose from (standard and accelerated), Kuaishou has two levels (standard delivery and balanced delivery), Tencent has two levels (standard delivery and accelerated delivery), and Baidu has three levels (standard, uniform speed, and accelerated).

The recovery amount in the table represents the advertiser's revenue generated through in-app purchases or advertising within a certain period of time (e.g., one week). When the bid point is the recovery amount, it means the price required for every 1 yuan recovered, which is the ROI bid.

For example, an advertiser can bid 0.4 yuan for a recovery amount of 1 yuan within a week, and the corresponding ROI target is 0.4. The ROI optimization product in Facebook is what everyone knows as VO (value optimization). According to the difference in cost/budget control, there are two corresponding bidding modes: Value optimisation with min. ROAS and highest value.

The product for optimizing app events is AEO (app events optimization). Based on the difference in cost/budget control, there are two corresponding bidding modes: lowest cost (in fact, there is also a target cost, which is slightly different in control) and cost cap.

【Summarize】

This article introduces the "two controls and one enhancement" in the "four points, three rates, two controls and one enhancement", and then analyzes in detail various intelligent bidding models in large media platforms, such as oCPX, double bidding, activation and payment, AEO, VO, etc. This concludes the analysis of the bidding patterns of major media platforms.

In the next article, we will turn our attention to affiliate advertising and RTB advertising to see how the three-party game has changed in affiliate advertising and RTB advertising as the interest binding relationship changes.

Author: Shentanshe

Source: Shentanshe