The product promotion logic of Internet advertising!

What the advertising business sells is every one of you in front of the screen, and here you are given a high-sounding name "traffic"; the buyer of the advertisement is the client dad, and the dads are various advertisers like Nike, H&M.

What’s really amazing here is that this set of processes for Internet advertising is fully automatic. The professional term is “programmatic trading”, and advertising is the industry with the highest degree of commercialization and scale of big data applications.

Internet advertising products based on “programmatic trading” are completely data-driven and calculation-oriented.

Today we will learn about the product strategy of Internet advertising.

01. Principles of Advertising Effectiveness

The fundamental difference between Internet advertising and traditional offline advertising lies in two points:

1. Advertising effectiveness is measurable.
2. Ability to continuously optimize profits through computing technology.

Therefore, Internet advertising is also called computational advertising.

One of the most core quantitative indicators for calculating advertising is called eCPM (expected Cost per Mille). According to the information transmission process of the advertisement, eCPM can be decomposed into the product of click-through rate and click value. These are also the two key points of our computational optimization.

To understand how computing technology can optimize advertising effectiveness, we first need to understand how advertising works.

The process from when an advertisement is exposed to the user to when it finally produces an effect is called the effectiveness model of advertising information dissemination.

The model divides the process of audience acceptance of advertising information into three major stages: selection, interpretation and attitude, which can be further decomposed into six sub-stages: exposure, attention, comprehension, acceptance, retention and decision.

The earlier the stage, the greater the contribution of improved performance to the click-through rate.

The later the stage, the greater the contribution of the improved effect to the conversion rate.

Image source: Taide Production

1. Exposure phase

This stage refers to the physical display of the advertisement, which may be offline, in a subway station, or online, in an app.

The advertising effect during the exposure stage is mainly related to the physical properties of the ad space, such as position and size, and there is not much room for technical optimization.

The effectiveness of exposure often has a much greater impact on the final result than other technical factors, which is why there is a saying that "position is king" in traditional advertising, such as billboards in New York's Times Square and billboards on high-rise buildings in the city centers of first-tier cities.

In Internet advertising, the impact of position is sometimes more significant, so how to eliminate the resulting click-through rate estimation bias from an algorithmic perspective is a very important practical issue.

2. Attention stage

Refers to the process from the audience being physically exposed to the advertisement to being aware of it consciously.

When it comes to advertising, exposure does not necessarily mean attention.

Whether an advertisement can be noticed depends on several important principles:

First, try not to interrupt the user's task as much as possible.

This is the basis of the principle of context-related advertising delivery and is also one of the starting points for discussing native advertising products.

When a user clearly identifies a fixed ad slot and no longer considers it relevant to his current browsing task, he will subconsciously block the content.

Secondly, clearly communicate to users the reason for pushing ads so that customers can establish a connection with the ad itself.

This is an important direction for the content creativity of "audience-targeted advertising".

Third, the content must be in line with the user’s interests or needs. If the content pushed is advertisements that the customer does not need or is not interested in, it will only cause user disgust. This is the principle basis of “behavioral targeting”.

3. Comprehension

Just because users pay attention to the content of an advertisement does not necessarily mean they will understand the message it conveys.

What principles should be followed in the understanding stage?

First of all, the advertising content must be relevant to user interests and understandable to users, which requires precise audience targeting.

Second, set a threshold of understanding that matches the level of concern. Because users' attention when watching advertisements will definitely not be as focused as when watching movies, the content of the advertisements must be easy to understand. You can focus on one point to attract the user's attention.

The picture below is a native advertisement released by De Rucci Mattress in WeChat Moments. Considering that most WeChat users just browse WeChat Moments and don’t pay too much attention, the content of the advertisement only focuses on one demand point: “Sleep better”, which is simple and easy to understand.

Image source: Taide Production

4. Acceptance

Just because a user understands the information conveyed by an advertisement does not mean that he agrees with that information.

The context of an advertisement has a great influence on its acceptance.

When an advertisement of the same brand appears on a spam forum and on the homepage of CCTV, users will tend to think that CCTV is more convincing. This is the brand value of high-quality media.

As targeted advertising becomes more and more common, how to make the right ads appear on the right media, namely the issue of ad safety, is also an important research direction in the advertising industry.

5. Retention phase

For advertisers who are not just pursuing short-term conversions but are looking at the long-tail effect of their brands, they hope that the information conveyed by the advertisements can leave a long-lasting memory on users to influence their long-term choices.

Therefore, brand advertisers spend a lot of energy on advertising creativity to improve the effectiveness of this stage.

for example,
Nike’s “JUST DO IT.”
Adidas’ “Impossible is nothing.”
Shi Yuzhu's "I won't give gifts this year, if I do give gifts, I'll give Melatonin."

When we hear the slogans of these brands, it is easy to associate them with the brands themselves; similarly, when we hear the brand, its slogans will come to our minds, as if they were written into our subconscious mind.

In this way, in specific consumption scenarios, we will automatically think of this brand, and such brand advertising is undoubtedly very successful.

6. Decision stage

The ultimate purpose of successful advertising is to bring about user conversion actions. Although conversion has left the business scope of advertising, good advertising can lay the foundation for improving conversion rates.

In particular, the reason why e-commerce advertisements, led by Pinduoduo, can go viral and convert is that their advertising creativity fully utilizes human nature's expectation of cheapness and aversion to losses.

02. Technical characteristics of Internet advertising

Compared with the traditional advertising industry, online advertising has the following five characteristics:

1. Technology and computing oriented.

The digital characteristics of online advertising give it the following advantages:

• Precise audience targeting
• Real-time bidding transactions
• Automatic decision making
• Extremely low marginal cost of delivery

2. Advertising effectiveness is measurable.

Online advertising can directly record various data such as ad displays, customer clicks, and customer conversion behaviors in the system, and can continuously optimize delivery strategies through data analysis.

3. Standardized creative and delivery methods.

What the advertising demanders care about is the user group that receives the advertisements, not the ad space itself.

There are a vast amount of media on the Internet. If the advertising space standards among the media are not unified, the advertiser's same advertising idea will have to be produced in different sizes and formats, which will cost more.

Similarly, media and advertisers all have their own systems. If the interfaces between the systems are not unified, communication between any two systems will require additional development, which will not only increase the complexity of the system, but also significantly increase the development and operation costs.

Thanks to the audience targeting technology and programmatic trading methods brought by digitalization, the unification of advertising creative sizes and the standardization of key interfaces have become particularly important.

As a result, IAB launched the VAST standard for video ads and the OpenRTB standard for real-time bidding, which greatly reduced the complexity of the systems of media, advertisers, demanders, and suppliers, and accelerated the liquidity of the advertising market.

4. Diversified media concepts.

With the development of Web2.0 and mobile Internet, various Internet media have continuously expanded the concept of media.

From early portal websites like Sina and Sohu, to today's search engines, e-commerce websites, social media, short videos, live broadcasts, and then to AR and VR in the 5G era, the continuous development of media forms will also continue to drive changes in advertising forms.

5. Data-driven delivery strategy.

The most fundamental driving force in the Internet era is the deep processing and use of data.

The computing technology that the development of Internet online advertising relies on essentially relies on the large-scale application of big data technology.

The online advertising system can be considered as a big data processing platform.

The basic logic of online advertising is as follows:

1. The media collects user behavior data and advertising feedback data extensively.
2. DMP platforms and DSP platforms use cloud computing infrastructure to tag users with matching tags.
3. DSP reports the targeted audience needs and quotes of each advertisement to ADX or SSP, and multiple advertisements bid for the advertising space of the media in real time, and makes decisions on which advertisement to display when users browse the advertising page.
4. Finally, DSP will feed back the statistical data of advertising delivery results to the advertising operators to adjust the delivery strategy.

03. The core issue of computational advertising

Andrei Broder, a member of the U.S. National Academy of Engineering and former vice president of Yahoo, who first proposed computational advertising, believes that the core issue of computational advertising is "Find the best match between a given user in a given context and a suitable advertisement."

Liu Peng, the author of "Computational Advertising", believes that the core problem of computational advertising is: "Find the most appropriate advertising strategy (Ad) for a series of user (User) and context (Context) combinations to optimize the profit of the overall advertising campaign."

Professor Liu Peng’s adjustment mainly emphasized that the advertising problem is to optimize a set of display effects rather than the display effect of a certain word. In addition, “given” was removed because in some advertising products, the system cannot obtain a certain user or context representation, but this does not mean that computational optimization cannot be performed in this case.

The goal of calculating advertising optimization is the total profit from these T impressions, that is, the difference between total revenue and total cost.

However, for a specific advertiser, the placement of advertisements is often subject to constraints, such as budget limitations and the need to ensure the amount of advertisements. We call this constraint the demand-side constraint.

Based on the existence of this constraint and according to ROI considerations, we can see that the cost or revenue of advertising is not only related to the advertisement, the user, and the context, but also to a specific advertiser.

At the same time, because the costs of most advertising products other than DSP also correspond to constants or are proportional to revenue, the cost part can be removed from the optimization formula in the figure below. What we can actively optimize is often the revenue part, so the main focus of computational advertising is revenue optimization.

After an ad impression occurs, we further decompose the composition of the ad revenue r according to its subsequent behavior.

1. After seeing an advertisement in the media, the user becomes interested and clicks on the advertisement, thus generating our first conversion data: click-through rate (CTR), which is the ratio of ad clicks to ad impressions.

2. After clicking, the advertiser’s landing page will be opened. The number of successful openings of the landing page divided by the number of clicks will give the second conversion indicator: reach rate.

3. If the scenario is e-commerce, the user starts from the landing page and further completes operations such as placing an order, which becomes a conversion. The ratio of the number of conversions to the number of successful openings is called the Conversion Rate (CVR).

We can decompose the revenue r into the product of click-through rate and click value. The product of these two parts quantitatively represents the eCPM (expected Cost per Mille) of a certain display. This is also the most frequently mentioned and most critical indicator for quantitatively evaluating revenue in computational advertising.

According to the above behavior decomposition after ad display, the click value function v() can be further decomposed into: reach rate * conversion rate * average order value.

Image production: Taide

For most advertisements, it is necessary to calculate the eCPM of the (a,u,c) triplet to make a decision.

According to the decomposition of eCPM, deciding which part will be estimated by whom is the fundamental reason for the emergence of various billing models in the advertising market. It is also an important link between business logic and product architecture in the advertising market.

The billing model I mentioned in "Do you know how Kuaishou and Douyin make money through advertising? 》 has been discussed in the article, so I will not repeat it here.

We just need to understand that there are two major categories of billing models:

1. Brand advertising is generally settled based on CPT or CPM, with CPM being the mainstream.
2. Performance-based advertising is generally calculated based on CPC, CPS, and CPA, among which CPC occupies a monopoly position.

04. Product strategy for Internet advertising

Driven by the demand for data utilization and monetization, the development of online advertising has gone through four stages.

1. Contract advertising products are mainly divided into CPT advertising sold by time period and CPM advertising sold by agreed display volume, and are mainly used for brand advertising.

2. The most important form of bidding advertising products is search advertising, and its product form is bidding on search keywords.

When this type of advertising expands to off-site display advertising traffic, it evolves into a product form that bids on page keywords and user tags, namely ADN.

Bidding advertising mainly serves performance-based advertising.

3. Programmatic advertising. The core of programmatic trading is real-time bidding and machine decision-making. Its core development driving force is the demand for the use and monetization of advertiser data and third-party data in the market, which has also given rise to the development of the data trading industry.

4. Native advertising products. With the development of mobile Internet, how to deal with the relationship between advertising and non-advertising content to make advertising less annoying to users has become a hot topic. As a result, "native advertising" that looks more similar to page content has emerged. Native advertising also puts forward new technical requirements: "native" requires more precise contextual data labeling information.

From the development stage of online advertising, we can see that the optimization of its product strategy mainly includes the following three aspects:

1. Focus on the business itself. As a type of commercial product (B-side product), advertising products require product managers to have a deep understanding of the business model and continuously practice and optimize based on actual business, including how to design the bidding mechanism, data search during cold start, and how to define the audience targeting tag system.

2. Pay special attention to data. Data is the blood of the Internet. The optimization and iteration of product functions and strategies must strictly adhere to the "closed loop of product design, R&D, and operations that starts with data analysis and ends with data feedback."

3. Try to provide users with efficient and convenient UI without pursuing flashy and cool designs.

05. Main technologies of computational advertising system

In computational advertising, both products and strategies are closely related to technology because this market is essentially computationally driven.

Computational advertising is a typical big data-driven personalization system, that is, a system that dynamically delivers personalized content based on individual user and contextual information. Its essence is the combination of an offline distributed computing platform and an online stream computing platform.

There are four main parts to a computational advertising system:

1. An online serving engine that responds to requests in real time and completes decisions. Its main functions include: retrieval, sorting, and global optimization of overall revenue.

The main modules include:

The main task of the ad server is to connect various functional modules and complete online advertising decisions.

The most important indicators are query per second (QPS) and advertising decision latency.

Ad retrieval is mainly used to find qualified ad candidates from the ad index based on user attributes and page attributes when online. The main technology is inverted index.

Ad ranking is used to efficiently calculate the eCPM of ads online and rank them. The calculation of eCPM mainly depends on the click-through rate estimation, which requires obtaining the CTR model and features through offline calculation, and sometimes also requires obtaining the real-time click-through rate features through stream calculation. In advertising products that need to estimate click value (such as DSPs that settle based on performance), a click value estimation model is also required.

Yield management (yield mgmt) mainly adjusts the local advertising ranking results under the premise of optimizing the global yield, such as online allocation in the GD system and bidding strategy in the DSP. This part generally requires the use of some kind of allocation plan calculated offline to complete online decision-making.

The ad request interface is essentially an HTTP request, which is used for data communication between different systems.

Customized audience segmentation: Since advertising is the media that replaces the advertiser to achieve user contact, it is sometimes necessary to divide the user group according to the advertiser's logic. The function of this module is to serve as a product interface for collecting user information from advertisers. If the collected data requires more complex processing, it will be imported into the audience targeting module via the data highway for completion.

2. Offline distributed computing data processing platform.

The most challenging algorithmic problems in computational advertising are concentrated in the online data processing part.

There are two output goals for discrete data processing:

One is to obtain reports, dashboards, etc. from statistical logs for reference when people make decisions.

The second is to use data mining and machine learning technology to conduct audience targeting, click-through rate estimation, distribution strategy planning, etc., to provide support for online machine decision-making.

In order to perform distributed processing on large-scale data, distributed storage and MapReduce computing frameworks such as Hadoop are generally chosen.

The main modules of offline data processing are:

User session log (session log) generation. Logs collected from various channels need to be organized into a unified storage format with user ID as the key. Such logs are called user session logs. The purpose of organization is to make the subsequent audience targeting process simpler and more efficient.

Behavior orientation. The function is to mine user logs, label users with certain tags in the structured label base according to the behavior in the logs, and store the results in the online cache of user tags for use by the advertising delivery machine.

This part is the raw material processing plant for computing advertising, so it has a very critical position in the entire system.

Context-oriented. It includes online page crawling and caching of page tags. This part cooperates with behavioral targeting and is responsible for labeling context pages for online advertising.

Click-through rate modeling. The function is to train the model parameters and corresponding features of click rate on a distributed computing platform, load them into the cache, and use them in online delivery system decision-making.

Allocation planning. It serves the online revenue management module. It uses offline log data to plan according to the specific needs of the global optimization of the advertising system and obtains an allocation plan suitable for online execution.

Business intelligence (BI) systems. Includes ETL (extract-transform-load) process, dashboard and Cube. These are the summaries of all data processing and analysis processes with human-terminal interfaces.

It is responsible for the task of external information exchange, and requires operators to make timely adjustments to some system settings based on data feedback.

Ad management system. This part is the user-oriented product. The user here refers to the advertising operator, that is, the interface between the client execution (AE) and the advertising system.

AE customizes and adjusts advertising through the advertising management system, and interacts with the data warehouse to obtain advertising statistics to support decision-making.

3. Stream computing platform for online real-time feedback.

Online data processing can basically be considered as a mirror function of offline data processing. It is designed to meet the advertising system's requirements for real-time data feedback and solve computing problems that offline distributed computing platforms cannot respond to quickly. Streaming management platforms are often used as infrastructure.

The main modules of online data processing include:

Online anti-cheating. Determine in real time whether there is fraudulent traffic in the traffic source and remove it from subsequent pricing and statistics.

Billing. In addition to real-time settlement of advertising fees, the billing function also needs to manage deductions. After deducting cheating traffic, for advertisements that have exhausted their deduction budgets, the system needs to notify the advertising indexing system to take them offline.

Online behavioral feedback. Includes real-time audience targeting and real-time click feedback. User behaviors and advertising logs occurring in a short period of time are promptly processed into real-time user tags and real-time click-through rate model features. This part is of great significance to improving the effect.

Real-time indexing. Receive advertising data in real time and build an inverted index.

4. A data highway that connects and runs the above three data flows.

The above four parts cooperate with each other to complete the system's data mining and online decision-making.

First, the logs of the online delivery system are connected to the data highway (DH), through which the logs are transmitted to the offline data processing platform (DC) and the online stream computing platform (SC).

Afterwards, the offline data processing platform (DC) periodically processes the data from the past period in batch mode to obtain clustering labels and other model parameters, and puts them into the cache for use in the online delivery system decision-making.

At the same time, the online stream computing platform (SC) is responsible for processing data from a short period of time recently, obtaining quasi-real-time user labels and other model parameters, which are also cached for use by the online system in making decisions. These are timely supplements and adjustments to the offline processing results.

As a result, the entire system forms a closed-loop decision-making process. Once this closed loop is built, it basically relies on machine calculations to operate, and the role of humans is only to make strategic adjustments and controls.

Such a closed-loop system is the key to effectively and fully utilizing big data.

It should be emphasized that since personalization requires the most accurate understanding of the user, in addition to the logs of the personalization system itself, other business line data or purchased data are generally used. These data will enter the data highway (DH) and subsequent processing flow.

Therefore, in the same enterprise, we need to share both offline and online computing platforms, as well as all user behavior data, among different businesses as much as possible.

Next, we will look at the main technologies of computational advertising systems from the perspective of algorithm optimization and system architecture.

From the perspective of algorithm optimization, the main technologies are:

1. Extracting user, advertising, and context features in audience targeting problems.
2. A click-through rate prediction model that implements local optimization for each impression.
3. Online allocation technology that takes into account both delivery volume constraints and timely decision-making during delivery.
4. "Market mechanism design and pricing strategy technology" to maximize profits when achieving dynamic equilibrium in a multi-party game market.
5. Use the Explore & Exploit method in reinforcement learning to more comprehensively sample user, advertising, and context feature data.
6. Personalized recommendation technology.

From the perspective of system architecture, the main technologies are:

1. After the advertiser's budget, targeting conditions and other information are set, they need to take effect online quickly. Real-time indexing technology is needed to retrieve advertising candidates.
2. Use a NOSQL database to provide users, contextual tags, and other features for delivery.
3. MapReduce distributed computing platforms such as Hadoop are widely used for large-scale data mining and modeling, and stream computing platforms are also used to implement short-term user behavior and click feedback.
4. A high-concurrency, fast-response real-time bidding interface is implemented in the advertising trading market.

Image source: Douban Reading (deleted if infringement)

06. Summary

Finally, let me summarize a little bit. Today, Taide will share with you five main parts:

The principles of advertising effectiveness.

Among them, the effectiveness of the first stage of "exposure" has a much greater impact on the final result than other technical factors.

The core value of computational advertising lies in the use of audience targeting, behavioral targeting, and contextual targeting to continuously optimize the effectiveness of the second stage of "attention".

The effectiveness of other stages such as "understanding", "acceptance", "retention" and "decision-making" mainly depends on the content of the advertising creative itself.

The core problem of Internet advertising is how to find the most appropriate advertising delivery strategy for a series of user and context combinations to optimize the profit of the overall advertising campaign, that is, to calculate the optimal eCPM of the (a,u,c) ternary function.

The product strategy of Internet advertising has three core points:

In-depth understanding of business models;

Data-driven, data-operated, and data-driven decision-making;

The core of user experience is speed rather than coolness.

The core systems of computational advertising mainly include an online delivery engine that responds and makes decisions in real time, a distributed computing data processing platform, and a stream computing platform that provides online real-time feedback.

Among the main technologies of computational advertising, I personally believe that the most core ones are the user, advertising, and context feature extraction technologies for audience targeting, as well as online allocation technologies that take into account both delivery volume constraints and timely decision-making during delivery.

The above is Taide’s superficial understanding of the Internet advertising industry. I hope it will be helpful to everyone. You are welcome to send me private messages to correct me and discuss!

Author: Daily life of B-end products

Source: Daily B-end Products