In the front, I introduced the development history, transaction mode and common advertising style of online advertising. This article introduced the core function module and system architecture of online advertising transaction system.

1、 Core functions of online advertising trading system

The online advertising trading system (adexchange) is responsible for connecting traffic owners and advertisers, and for completing billing and trading. Because of the more and more advanced programmed advertising technology, sometimes someone will use RTB (real-time bidding) to refer to adexchange.

Before introducing the advertising trading system, it is necessary to briefly introduce how an advertisement is delivered to the user’s eyes when the user views the web page:

When the user clicks the website / APP / video and other media to generate a network request, the media will give the request to the SSP responsible for integrating traffic, and the SSP will give the information with the user’s profile to adexchange for transaction. Adexchange will notify all DSP responsible for integrating the advertiser to make a bid, and inform the size of the advertising space, the URL and category of the advertising space, and the cookie of the user ID, etc.; select an advertiser with the highest bid in each DSP platform, and charge the second highest bid + 0.1 yuan, and then return the original advertising idea of the advertiser with the highest bid to the user, so that the basic process of an advertising transaction is finished, and the whole process is generally controlled within 100ms, and all the data involved will be guaranteed in DMP Deposit.

With the above flow chart, you should be able to understand the basic operation mode of an adexchange (advertising trading system). The main core modules are divided into the following parts:

1. Online delivery module

Help the advertiser to complete the advertising action: including budget setting, frequency control flow control, advertising management, precise orientation, launch time, competitive bidding and other related operations; complete the relevant functions of receiving requests, timely decision-making, advertising, etc;

(1) Budget setting: before each advertisement is sent to the user, it will first query whether the budget is sufficient. In general, large domestic platforms do not support overdraft;

(2) Frequency control flow control: in order to prevent users from seeing too many advertisements and causing antipathy, frequency control and flow control will be carried out when advertising is launched to ensure that most of the content in a page is what readers want to see, and the most intensive level that general users can accept is that one advertisement appears in eight information flow contents;

(3) Advertising management: advertising management here refers to the management of advertising space, including the specification, style, schedule, status, revenue, etc. of each advertising space in the website or app;

Advertising space is the resource of advertising system. It is an important thing for commercial product managers to monitor and control the content of advertising space and ensure the maximum benefit of advertising space. Generally, advertising space will be contracted to large contracted brands in a whole period of time. Such brands have a strong strength ratio. It is a high-quality choice to sell resources to them first, and then complete the small pieces The traffic of is sold to small contract advertisers or some high-quality competitive advertisers. At last, the rest will conduct procedural real-time bidding. For details, please refer to the previous article (commercial product manager | calculate advertising (2): transaction mode and billing mode of online advertising)

In order to prevent the advertising space from opening empty windows, generally when designing the advertising space, a layer of bottom guarantee advertising will be added to it, generally from other advertising leagues. The main reason why advertising retrieval system is funnel model is to find a compromise between algorithm effect and performance. Funnel model from top to bottom, the upper algorithm is relatively simple, can handle a large number of ads, the lower the level, the more refined the algorithm.

(4) Precise orientation: the basic problem to be solved is who the user is, where the user comes from, and where the user wants to go. In the delivery module, it is mainly used by advertisers and records the target groups that advertisers want to launch. In the real operation of the system, there are mainly the following ways:

Group attribute orientation: mainly refers to the orientation based on the basic attributes of users, including age, gender, etc. For example, advertising cosmetics to female users.

Behavior orientation: it refers to mining users’ interests based on users’ historical behavior data, such as the data of users’ comments, likes and so on.

Geographic location orientation: refers to the use of mobile devices to record the geographical location of users and put corresponding advertisements, such as restaurant advertisements.

Orientation of similar users: refers to the use of some people who have been identified to find similar users for orientation.

Keyword orientation: summarize keywords according to machine or manual analysis, and match them through similarity.

(5) Timely decision-making: the advertising trading system has high requirements for response speed, data throughput and accuracy. When the user’s request comes, it is necessary to convey information to the DSP of all parties, complete multiple jump communications within the specified time, and must complete the decision within 100ms. If it is later than 100ms, it is likely to be misjudged that the advertiser cannot display or give up the bidding.

2. Retrieval / sorting module

The most suitable and profitable advertisements are displayed to users, including analysis of user tags and context, inverted index of candidate advertisements, search in sorting, online calculation of ECPM sorting and other related functions;

The most suitable and profitable advertisements are displayed to users, including analysis of user tags and context, inverted index of candidate advertisements, search in sorting, online calculation of ECPM sorting and other related functions;

(1) Analysis context:

According to the content of the page, the algorithm determines how to identify the title / related keywords of the page.

Ignore messy HTML content and filter only relevant and important keywords.

Without losing the relevance of the recommended keywords, it expands with the increase of the number of web addresses.

(2) Inverted index:

(3) ECPM calculation:

The expected cost per mille (ECPM) is the most critical indicator for quantitative evaluation of earnings. It refers to the advertising revenue that can be obtained from thousands of actual or estimated advertising displays.

ECPM can be decomposed into the product of click through rate and click through value


a. Three variables, u and C, respectively represent advertisement, user and environment

r: Return

μ: Click through rate, which describes the behavior in the media

v: Click value, that is, the revenue that a single click brings to an advertised product, describes the behavior that occurs on the advertiser’s website

The product of μ and V quantitatively represents the expected CPM value of one or several displays.

3. Feature calculation module

Complete the optimization of advertising prediction model: including CTR (click through rate) optimization, CVR (conversion rate) optimization, CTR model and features that need to be calculated online;

The goal of the prediction is to find the most suitable advertisement for the specific audience users under the given media environment, and to maximize the revenue for the media. The simplest CTR prediction model, according to the historical log, calculates the three-dimensional CTR cross reference relationship. In the prediction process, when a user accesses a specific URL, if there is a CTR in the query dictionary, the ad with the highest CTR will be returned. If there is no ad, the ad will be returned randomly to accumulate subsequent data.

Existing problems: Based on the statistical data, the effect of the old ads is OK, but there is no prediction ability for the cold start ads.

In fact, we do the click rate prediction model online, and the algorithm used is logical regression. The following methods that may be considered in the future are:

Machine learning method: feature + model

Collaborative filtering method: recommendation system

The ranking model is based on the prediction results, and there are several kinds of advertising ranking models as follows:

Point wise approach: turn it into a classification problem or regression model to deal with

Pair wise approach: compare two ads whose priority is high and do not classify

List wise approach: learning sorting model for the whole advertising candidate set

4. Billing statistics module

Manage advertiser’s budget and consumption: including advertiser’s budget management, traffic revenue management, real-time billing, smooth consumption, global revenue and other functions;

5. Data analysis module

Store and manage all data generated by the trading platform, including big data storage, log system, data cleaning, data analysis, data highway, anti cheating, offline model optimization and other functions.

The above is just a brief introduction to the core modules of adexchange from the perspective of core functions. According to different businesses and companies, there will be differences in each platform. However, as long as the core business processes are around, the general functional division is basically the same mode. With the increase of platform volume, the architecture of the platform will change accordingly. This chapter leads you to understand the internal of adexchange Principle, later will share the internal details of online advertising with you from more angles.

2、 Four top advertising trading platforms in the world

1. Google Ad Exchange

Currently, the most popular exchange is Google ad exchange, which provides publishers with access to all advertisers in Adsense as well as other senior / major brand advertisers. They have terms to prioritize transactions, filter / block, guarantee anonymity and Google reputation, which is absolutely the preferred transaction for most digital ad traders.

2. OpenX

For most advertisers, OpenX is another popular option, offering incredible year analyst, which attracts many buyers and sellers to the exchange. The exchange ensures the full autonomy of stakeholders in the whole advertising platform. In addition, their customer support is first-class and responsive.

3. AppNexus

By visiting more than 692427 websites, appnexus offers a variety of options for people who want to advertise online. It also has a dedicated gateway, which can access the main advertising trading platform and advertising resource aggregators.

4. SmartyAds

Smartyads has more than 250 DSPs, caters to global user groups, and has a thriving community of buyers and sellers in an active open market. The market connects advanced publishers, mobile applications and direct advertisers, thus providing high-quality services.

#Related reading#

Commercial product manager | computing advertising (1): the origin and development of online advertising

Commercial product manager | calculation advertising (2): transaction mode and billing mode of online advertising


Author: Fantasy Volta; official account Author: Fantasy Volta, commercial and computing advertising product manager

This article was originally published by @ Fantasia vodka and is copyrighted by the original author. No reprint without permission

The picture is from unsplash, based on cc0 protocol

Tagged in: