Talking about advertising traffic distribution strategy

Traffic is the basis of advertising monetization. How to reasonably utilize traffic and give full play to its maximum value is a problem that every advertising practitioner will face. This article explores the distribution mechanism in traffic flow from the perspective of ADX. A reasonable distribution mechanism can maximize traffic benefits. I hope readers can gain some inspiration from this article.

Traffic flow mechanism

ADX (AD Exchange), an advertising trading market, plays a connecting role in the traffic flow process. It connects to DSP upwards and is responsible to SSP/media downwards. Understanding the advertising traffic flow mechanism with the help of its workflow will help us better understand the possible optimization points in the process of traffic overdistribution. The advertising traffic flow mechanism is as follows:

When the front-end App triggers an advertising traffic opportunity, it will send this traffic to the connected ADX. The traffic attributes usually contain relevant attributes such as advertising space and user information. When ADX receives the traffic request, it will first verify the legitimacy of the traffic. The simplest one is parameter verification, and then verify the preset value of the order/DSP, and finally send the traffic to which DSPs. When the DSP receives this traffic, it decides whether to participate in the bidding based on the relevant attributes carried in the traffic. If the traffic is suitable, it returns the bidding price (or dealId) and advertising elements to ADX. ADX receives the bidding information of each DSP, and after a series of validity judgments, it sorts the bids according to the price. The highest bidder wins. The winning advertising information is sent to the media, and the DSP is notified that its advertisement has won (this step is not necessary, but recommended). After receiving the advertising information, the media renders and displays the advertisement.

When user behavior occurs, ADX and DSP related behavior data must be transmitted back through the monitoring link. The main behaviors include exposure, click, download, wake-up, etc. There are two modes for transmitting behavioral data through monitoring links: C2S (Client to Server) and S2S (Server to Server). Currently, most customers require the C2S reporting method when placing ads. The key indicators involved in ADX are mentioned in the ADX section of the previous article "A Comprehensive Inventory of the Role of Commercial Advertising". This article aims to explore the traffic distribution mechanism and will not explain the indicators in detail. Interested readers can go ahead and read it.

Through the above traffic flow process, it can be found that advertising traffic is mainly forwarded on the ADX side. If ADX is connected to multiple DSPs, a reasonable traffic distribution mechanism can improve the fill rate and ecpm, thereby maximizing traffic revenue.

waterfall

When ADX is connected to multiple DSPs, when requesting different DSPs, should it be a serial request or a parallel request? This involves different strategies. First, let’s talk about serial requests, or waterfall.

Waterfall, translated into Chinese as "waterfall flow", literally means "flowing from top to bottom", but how should we understand the four words "from top to bottom"? In the advertising industry, waterfall refers to "requesting DSPs from top to bottom to distribute traffic based on historical eCPM data when the value of each traffic cannot be evaluated in real time." This is called an ad serial request. Let’s look at the use case of waterfall through a practical example.

Assume that ADX is connected to three platforms, and the eCPM and filler materials of the three platforms are as follows:

If there are 1,000 ad requests, there are the following ad request plans:

Solution 1: All requests to DSP1

Profit = 1000 * 20 / 1000 * 30% = 6

Solution 2: Request all DSP ad sources

Profit = 1000 * 15 / 1000 * 50% = 7.5

Solution 3: All requests to DSP3

Profit = 1000 * 25 / 1000 * 20% = 5

From the above three solutions, although the eCPM of the solution is the lowest, its fill rate is the highest and the final total revenue is the highest. Is Plan 2 the best solution? The answer is definitely no, because it only utilizes 50% of the traffic, and the remaining 50% is wasted, which leads to Plan 4.

Solution 4: First, request all 1,000 ad requests to DSP3, request the unfilled part to DSP1, and finally request the unfilled part to DSP2. The specific traffic distribution flow chart is as follows.

Revenue = 1000 * 25 / 1000 * 20% + 800 * 20 / 1000 * 30% + 560 * 15 / 1000 * 50% = 14

The final profit of Plan 4 is 14 yuan, and the fill rate is 72%. Compared with the first three plans, it not only improves the profit, but also improves the fill rate.

Now that the revenue and fill rate have increased, can the use of waterfall solve the problem of traffic distribution? Reality will always slap you in the face. The waterfall solution has the following problems:

1. The core point of waterfall lies in " historical eCPM data ". So how to measure the historical eCPM data of a dsp, and how long is this history?
2. Serial requests will increase the time it takes to display ads. The average request time is at least 100ms. Multiple requests will cause front-end display delays and a poor user experience. Due to the different environments of different ad slots and different user acceptance levels, it is necessary to set the overall number of requests/timeouts for each ad slot.
3. Since the priority of waterfall requests is determined based on historical eCPM data, for a specific request, the bid of the DSP in the front may not be as high as the bid of the DSP in the back. In this way, you will miss the DSP ads with higher bids that are ranked behind, and the traffic benefits will not be maximized.
4. The eCPM data of each DSP is maintained. Due to seasonality, the eCPM value will change, which requires manual maintenance by operations staff, which is costly.

Here, let’s talk about “eCPM of historical data” in detail:

1. How long is this history? There is no standard answer to this question! Because the effect of each DSP is different. The only thing we can do is to try our best to predict each company’s eCPM and fill rate. This can be verified through historical data and understood through business relationships. Only when we get correct and stable values ​​can they be true and reliable for us. 3 days, 7 days, 10 days or longer is all ok, as long as you think the number is reasonable and can stand up to scrutiny.
2. For a newly connected DSP, since it has no historical data accumulation, how should its eCPM value be evaluated? a) You can understand its eCPM and fill rate through business operation channels; b) You can provide traffic support to newly connected DSPs, and after accumulating a certain amount of data, return to the normal DSP for sorting. The cycle and sample data of this traffic support have different requirements for each algorithm team, and it is sufficient as long as it meets the needs of their own business.
3. If two DSPs have the same eCPM and fill rate, how should they be ranked? At this point, you can evaluate from other dimensions, such as interface response time, material quality, etc.

Header Bidding

Since waterfall has many problems, are there any other alternatives? Readers must be thinking, if DSP can return the current bid in real time every time bidding, then there is no need to calculate and maintain "historical eCPM data". When distributing traffic, traffic can be distributed in parallel. After obtaining bids from all DSPs, the successful bidder is determined based on the bids. This is "Header Bidding".

“Header Bidding”, translated into Chinese as “header bidding”, literally means “traffic is sent to head buyers, head media bid, and then the winning reserve price is used as the reserve price to request other DSPs that do not support real-time bidding.” To achieve this, we must first have the following prerequisites:

1. When the top buyer returns the creative, they need to return the bid at the same time, so that the media/ADX can complete the bidding;
2. Although non-head buyers do not support real-time bid returns, they need to support passing in the floor price of the ad space. In this way, if an ad is returned, the price must be higher than the floor price, which will bring the highest profit to ADX and the media.

Header Bidding originated abroad and was initially used on PCs. DFP (Google Doubleclick For Publisher) is the advertising platform with the most integration in foreign PC websites. Due to its monopoly on PC advertising and Google's Ad Exchange dynamic bidding (interested friends can search Baidu for details), it is not very friendly to publishers and other DSPs. Therefore, AppNexus hopes to join hands with other ADX/DSPs to shake the monopoly of DFP through Header Bidding technology.

From the above description, we can see that header bidding has the following advantages over waterfall:

1. Fair bidding: All DSPs bid at the same time, and each evaluates the traffic value and makes a bid;
2. Maximize revenue: DSPs that were originally ranked at the bottom of the waterfall can win ad display opportunities by increasing their bids.

in the country, the development of PC has been relatively stagnant, and greater potential lies in the mobile sector. Therefore, to be more precise, domestic header bidding should be called In-App bidding. Since In-App bidding started late in China, currently only a few leading media support real-time return bidding. Therefore, for a long time, header bidding and waterfall will coexist. For media that support real-time bidding, header bidding will be given priority, and then the winning bid will be used as the floor price for the ad space to request other DSPs, and finally bidding will be based on price.

Summarize

In fact, whether serial or parallel, they are just strategies to solve the problem, and the core goal is only one "maximum traffic revenue". From the perspective of the media, of course they hope that more media will bid at the same time; from the perspective of DSP, they must hope that the traffic will be sent to themselves first, and then sent to other parties after they have selected it, or even monopolize the traffic.

Of course, the real environment is complex and intricate, and different docking methods will also affect different strategies. Only by firmly grasping the key point of "maximizing traffic revenue" and taking into account the interests of multiple parties can we respond to changes with constancy.

1. When major e-commerce companies compete for the market during e-commerce festivals, they have sufficient traffic budgets. In order to get more budgets, traffic is distributed to e-commerce DSPs first;
2. The eCPM and fill rate of some DSPs are OK, but the material is relatively low, and sometimes they may involve black five ads, or there are technical pitfalls (such as high network latency). At this time, traffic restrictions need to be imposed on these DSPs;
3. Although some DSPs have low eCPM, their fill rate is still good, so they are more suitable for guaranteed fill and need to be given a certain proportion of traffic to maintain.
4. …

Author: Baozi

Source: Daily Notes on Commercialized Products