Commercial advertising strategy
Author: Morphick
Edited by Hoh
Source: Moffick
Production platform: DataFun
Note: Reprint please contact the original author for authorization.
Guide:With the rapid development of the Internet, the market size of online advertising has risen, becoming one of the mainstream real estate models of the Internet, but also gave birth to a number of advertising giants, such as Google, Facebook, the domestic byte beat, Ali, Baidu, Tencent, advertising in the entire revenue occupies an important position.This article will introduce advertising dynamics, target conversion bids, data management platform (DMP), advertising auction kernel, mechanism design, optimal mechanism and so on.
▌Ads are dynamic
People vary widely, with different basic attributes: gender, age, income/consumption level, education level, scene status: geographic location (city level, province or city), context (channel, page content), terminal state (network environment, terminal type, operating system, operator), different historical behavior: different operations (reading, clicking, collecting, attention, downloading, purchasing, active and retaining) for different types of content. We know that advertising is to put the right content at the right time in the right way to the right people, for different types of users static display of a single idea and merchandise, one-piece, conversion effect is not ideal can be imagined, for which the dynamic advertising, specifically programd creative and dynamic commodity advertising (DPA), we describe:
Programd Creatives/Dynamic Creatives: Different users see different creative materials, and the same user presents different ideas at different stages (first/non-first time seeing ads, not converting users/converted users) to better match the user's scenario.
Creative production: Instead of making the entire idea directly from content to form in the past, set the content, title (m) and form (n) separately (decoupling), freely combine (m*n) by the system, expand the number and possibilities of ideas, and present them randomly (DPA can call the library footage, not satisfied with manual editing and replacement).
Exposure Strategy: Dynamically adjust your exposure strategy based on your creative conversion data (click-through rate, conversion rate): Increase the number of exposures to high-conversion footage, reduce the exposure of low-conversion footage, and improve the overall creative transformation.
DPA: Dynamic Merchandise Ads - Docking Library to show different items based on the different dynamics of visitors.
Scope of application: multi-category goods /services (thousands of people, personalized recommendations).
Commodity library docking (way): There are two ways to dock a commodity library, one isManual upload of commodity packets, the number of applicable goods is small, low frequency of change, one is the API interface, the number of applicable goods, high frequency of change.
Exposure principle: People who visit the media (website/APP) have personality tags, merchandise in the library has personality tags, match labels, and different users display corresponding product ads, including creative-floor pages-conversions.
Purpose/Effect: Show the right content to the right people and improve the matching and conversion performance of your ads.
Points to note:
Label mapping: the media end of the user's label system and the product side of the label system to do mapping, the two labels to be right on, commodity advertising can do the appropriate;
Data closed loop: advertisers need to pass back the user's subsequent conversion data, from creative exposure to click, arrival and final conversion orders need to establish a data model to dynamically adjust strategies to improve conversion;
Product types: Advertisers have enough types of goods/services to be shown for different users, and if there are fewer categories and the effect is relatively poor, it is mainly adapted to users through procedural ideas.
Label adjustment: The initial label may not be accurate, and the label for the product will be dynamically adjusted as the conversion data changes.
User ID: Users of different sources/end types need to be identified and uniquely identified, the PC side is primarily cookie Mapping, the mobile side is primarily using device number, Android (IMEI) - iOS (IDFA).
▌ your target conversion bid
Ads are billed in a variety of ways: CPM, CPC ( Click. )、CPA ( Behavior. )、CPS ( Into. )、CPD ( Time. ), the most common of these is CPM ( Mostly used in brand advertising ) and CPC ( More for performance ads ), the media after the presentation and click to charge advertising fees, how to basically no longer consider the scope of subsequent conversions, based on this ad delivery also formed a phenomenon: in the creative exposure and click stage performance is good, but in the follow-up performance is very poor, low conversion rate, which is very helpless for the effect of advertisers, because conversion is the main purpose, so they hope that the chain extends back, can ensure conversion effect, in this demand driven by the development of targeted conversion bids ( OCPM/C )。 Let's have a chat.
OCPM/OCPC：Target conversion bids, advertisers set specific conversion goals and estimated prices, and the DSP system is optimized according to the conversion data model.
Advertiser Goals:In the case of controlled costs ( By conversion target bid, billing is converted to CPM/C )for better conversion results.
Key categories of conversions:APP download class/H5 lead class.
App download class's main conversion objectives: installation-activation-registration-pay-active-retention;
H5 Lead Conversion Goals: Form Submission-Appointment-Online Consulting-Telephone Communication-Registration-Purchase.
Each of these goals, as the path continues to extend, the number of users continues to decrease, and the higher the price of the more effective users later, the more the target is recommended to be set from front to back.
Prerequisites:Advertisers need to retest conversion data, and the platform tracks and closes the path data for users from creative exposure, clicks to landing pages to final conversions, to build data models.
How to do it:
Platform construction station：Advertisers use the tools provided by the platform to build a station, there is no need to deploy code separately, the platform's data has been opened, to achieve back-pass;
Advertisers own sites：The monitoring code for the platform is deployed at the site ( Web class JS/API, mobile SDK/API )。
The logic of the overall business process:Advertisers deploy monitoring codes - back-to-back conversion data - Platform data modeling - Advertisers identify optimization goals and prices - Platform fits forecast data - Dynamic optimization adjustments. There are several aspects that are important throughout the process:
Data modeling:Platform according to the back-pass user data, built in the various steps of the conversion model, the model construction needs actual data to support, and the early fluctuations may be relatively large, in the case of the delivery strategy unchanged, the more data accumulated, the more mature the model, the more accurate the prediction.
Optimization objectives:Conversion goals ( H5/APPç±» ) Bid ( The system recommends a bid ), in general, the DSP platform will be docked with all kinds of advertisers, data accumulation for various industries, according to the advertiser's delivery requirements combined with the platform's traffic situation, will give a suggested bid, the proposed bid within the threshold, too low not competitive, exposure will be affected, too high cost will increase.
Fitting forecasts:Estimate Reach people plus smart bids, estimating reach, exposures, clicks, and conversions based on advertisers' targeting criteria and bid platforms. Bids are launched in two phases:
The first stage, the machine learning phase, advertisers bid at the average conversion price (historical conversion price);
In the second phase, a fitted forecast model is established to calculate the conversion price calculated by the system (advertiser sets thresholds and the system recommends bid intervals).
Note:During modeling, the delivery strategy is not recommended for large changes, stability, advertising is generally sorted according to ECPM ( ECPM s 1000 s CTR s CVR s conversion bid ), modify the strategy needs to re-learn, mainly pay attention to the following points: targeted crowd settings ( The crowd is different, the labels are different, the prices are different )Bid. ( Directly affects the auction results ), the budget remains adequate ( The under-budget platform prevents overs run, and exposure is limited )Creative. ( Affects click-through rates ), landing page content ( Affects subsequent conversions )。
▌ Data Management Platform ( DMP )
With the development of the advertising industry, related technologies are also continuing to improve, especially since the popularity of procedural advertising, data-driven features in advertising is increasingly prominent, we have said before, the essence of advertising is to put the right content at the right time in the right way to the right people, here is the key point is: how to determine the people who are advertising is the right person. In order to effectively solve this problem, DMP has been developed ( Data management platform ), the goal is to segment the population through the label image to meet the advertiser's precise targeting, to achieve effective advertising, let's expand to say DMP.
User:From the above, we can understand that DMP's primary target audience is advertisers and agents.
Use claims:Advertisers use the DMP platform to create, analyze, and expand their use to deliver people with precision.
Classification:Advertising ecology ( RTB in particular ) There are many participants, according to different roles can be divided into advertisers self-built, DSP platform construction, independent third-party DMP, the three objectives are slightly different, advertisers self-built to improve their own advertising closed loop, the challenge is to have certain requirements for technology and scale, otherwise not worth the loss; Cash out by collaborating to divide or sell targeted groups of people.
DMP core features:
Create segments of the population, there are several main ways:
Using DMP's mature labeling system to create targeted crowd packs;
Advertisers upload their own user base data;
Build crowd packs for intended users based on ad serving performance.
Analyze user portraits:The proportion of each component structure and TGI of the population was analyzed from the dimensions of gender, age, province, city level, model, interest, etc ( Target group index )。
Seed user base expansion ( look-alike )：Analysis of the characteristic behavior of the extraction seed group, matching similar extended users within the platform range ( Advertisers choose to expand the number of users )。
Remarketing:Extract the converted user base for targeted promotion, in which the creative, landing page of the display content needs to be distinguished from the new users.
Crowd operations:New target groups are obtained by combining, excluding, and cross-operation between two and more groups.
Analyze delivery results:And according to this adjustment and optimization of the crowd package.
Here's a quick look at the main processes for creating, scaling, and computing populations:
Create a process: Determine the name of the crowd - Determine the grouping - Select various types of tags (attributes and interest behaviors) / Upload crowd packs / Ad serving users.
Extension process: Determine the type of extended device - Determine the traffic platform - Determine the number of extensions - Process the seed population - Determine the grouping - Determine the name.
The operation flow: determines the operation logic (combined, poor, intersected) - determines the packet - determines the grouping - determines the name.
Through the above description we have a general understanding of DMP, understand its importance, user needs and main functions, interested students suggest practical use, strengthen cognition and understanding, DMP through the precise segmentation of users, for advertisers to follow up to achieve thousands of people, dynamic commodity advertising and frequency control and other operations to provide a solid foundation, with the development of time, advertising digital-driven features will become more important and prominent, DMP in the future will be more dazzling.
▌ the auction kernel of the ad
With the rapid development of the Internet, the market size of online advertising has risen, becoming one of the mainstream real estate models of the Internet, but also gave birth to a number of advertising giants, such as Google, Facebook, the domestic byte beat, Ali, Baidu, Tencent, advertising in the entire revenue occupies an important position.
Online advertising operates in a complex way, especially in real-time bidding ( RTB ) Stage, not only rely on The Internet technology and big data, but also involve advertisers, agents, DSPs, ADX, SSPs, DMP, media, users and many other participants, this is the modern side of online advertising, but when we extract these specific forms, we will find that the kernel of online advertising can be traced back to a long-standing category of behavior - auction. Speaking of auction everyone's mind will auction scene: the host shouted 1 million once, 1 million twice, 1 million three times, the transaction, this is a form of public auction, auction has multiple forms and mechanisms, let's start a chat.
From the auction of items, the auction in ancient times has been used to sell a variety of items: jewelry, paintings, antiques, online advertising to sell things in a different form than their predecessors: exposure, bidders ( This can be specifically for advertisers ) The purpose of participating in the auction is to gain exposure and present their products to the audience.
Formally speaking, the common form of auction mainly includes four categories, the British auction is the auctioneer to bid a lower price, at least two bidders to participate in the bid, the price all the way up until the remaining bidder is interested; It's a public price-cutting auction process in which the auctioneer offers a high price and the price drops all the way down until an auctioneer is interested; Bidders submit bids in a sealed form, with the highest bidder winning the auction, the difference being that the winner of the sealed auction pays his own offer with the first price sealed auction, and the winner of the second price sealed auction pays the second highest bid.
From the above description, we can deduced that the main mechanism of the auction consists of two aspects: allocation rules and payment rules, allocation rules determine how auction items are distributed, payment rules determine how bidders pay. About the distribution can be first-come, first-0, random selection, high price, the last, all look at the mood of the auctioneer, in the auction theory the highest price is a standard auction; ( FP )pay the second highest price ( SP )to pay for losses caused to others as a result of their participation ( VCG )。
The development of online advertising ( Contract Ads - Spot Ads - Live Auction Ads )The rules are evolving, and the initial contract ads are first-come, first-7 ( Pre-agreed, scheduled delivery, pay the agreed price ), to the bid ads adjusted for the high price, the meaning of the high price also after some development changes, initially according to the display bid, at this time the highest price is the highest bidder, and then Google will click rate into the auction formula, developed according to the ECPM auction ranking, at this time the price refers to the highest ECPM ( ECPM=1000*pCTR*bid ), wherebid is a bid for clicks, pCTR refers to the estimated click-through rate of the ad, if according to the current development of the hot OCPX model ( Convert bids by target )The calculation of ECPM should increase the conversion rate prediction on the basis of pCTR, which is more demanding for big data and algorithms, if the platform is designed according to the optimal bidding model, the bid participating in ECPM calculation is a virtual valuation after fitting the advertiser's valuation distribution function, the calculation will be more complex and of course the benefits will be more considerable; The pros and cons, in a simple sense, the broad price in a particular case the yield is higher, but the volatility is large, the second price is more stable than the first price ( Incomplete information about Nash equilibrium )VCG, for its part, is incentive-compatible and encourages telling the truth, but at a higher computing cost. We'll talk about the mechanism design of the auction, the dynamic floor price, and the squeeze factor in the following article.
Online advertising developed to the real-time bidding stage, each exposure is the result of bidding, that is, every ad display we see has gone through the request-inquiry-bid-bid-win-release process, the daily exposure of advertising in billions of dollars, for response speed ( From request to final presentation, you need to control it in hundreds of milliseconds )High contigies require very high performance, not only that, but also the use of big data for click-through rate, conversion rate forecasting, while achieving billing, anti-cheating, etc., more complex than traditional auctions, the technical content is also higher.
Online advertising pry up and down, forming an industrial chain, involving tens of billions of resources, it adhering to the core of the auction, in the development of modern technology in the wave of continuous evolution, and finally grew into a generation of auction master.
▌ mechanism design
In the last section, we came up withThe kernel of the ad is the auction。 Auction as a way of selling has a long history, there are many forms of expression, the real impact of the auction results is the auction using the mechanism: not only can control the final commodity ( Exposure. )Belonging can also determine how much the buyer of the goods will pay ( Billing price ), which can be said to be critical to the advertising platform, let's have a chat.
An auction is an activity in which the seller allocates items to the buyer for a paid amount through a mechanism with two overall objectives:
The seller gets the desired return
Impact on society ( whether the auction is effective )
When auctions can be allocated to the highest-valued buyers anyway, we say they are valid. In an auction, the buyer's estimate of the value of the item is called valuation ( Represented by v )The buyer's bid for the auction item is called the offer ( Represented by b )。 For ease of analysis we make the following assumptions: there is an indivisible individual auction, the number of potential buyers is N, buyer i valuation of the item isvi ( Suppose the buyer knows the value of the auction to himself and is not affected by other buyers, called private value ), the quote isbiThe seller does not know the buyer's true valuation, but knows the cumulative distribution function F of the buyer's valuationi(v), and continuous density functionsfi(v)，FiIndependent of each other, assuming that the range of V is 0, w,fi≥0。
The auction mechanism mainly includes ( Q，M )In two parts, Q represents the configuration rules that determine the probability that buyer i will get the itempi(b), M on behalf of the payment rules, determines the buyer i's expected payment ti(b), then the utility of buyer i ( Consumer surplus ) ui=vipi(b)-ti(b), at which point each mechanism defines an incomplete information game between buyers and sellers (not knowing the buyer's true valuation only knows the valuation distribution and knows its valuation).) )"If, for buyer i, the maximum return is achieved given that the other buyer's strategy remains unchanged, then he has no incentive to adjust his offer." ( Risk neutral ), the system as a whole is stable, and we say that at this point constitutes a Nash equilibrium of the mechanism. The direct mechanism requires each buyer to report its own valuation independently at the same time, and if each buyer 3d reports its own valuation to form a Nash equilibrium, then we call it true equilibrium, and according to the principle of display, any equilibrium result of any mechanism can be replicated by the true equilibrium of a direct mechanism, i.e. the two are equivalent, so we can focus on the direct mechanism.
The above says that in the direct mechanism each buyer 3 report their own valuation, if the formation of Nash equilibrium, called real equilibrium, then this mechanism is indicated that the mechanism is incentive compatible, that is, to tell the truth for the buyer is a weak advantage strategy, that is, for the buyer i, estimatedvi, the quote isxithe utility of the buyer i ( Consumer surplus ) ui(v)=vipi(v)-ti(v)≥vipi(x)-ti(x), in our previous connection mentioned in the single item second price sealed auction to meet the incentive compatibility (allocation rules: the highest quote, payment rules: two-price settlement), we can get the truth through simple reasoning ( bi=vi ) is a weak-dominant strategy.
The reasoning is as follows: assuming there are N buyers, buyer i is valuedvi, the quote isbi，pj=maxbj≠i, ifbi＞pj, the buyer gets the item, the buyer utilityui=vi-pjInstead, the item is not obtained, and the buyer is available at this timeui0, divided into two broad categories to analyze separately:bi＞viAnd.vi＞bi；
① bi＞vitime
If.bi＞vi≥pj, get the item, at this point the buyer i utilityvi-pj≥0, withbi=viThe time is the same;bi＞pj＞vi, get the item, at this point the buyer i utilityvi-pj< 0, the utility is lowerbi=viTime-utility ( The utility of the non-obtained item is 0 )Ifpj＞bi＞vi, do not obtain items, at this time the buyer utility is 0, utility andbi=viThe utility is the same ( The utility of the non-obtained item is 0 )。
② vi＞bitime
If.vi＞bi≥pj, get the item, at this point the buyer i utilityvi-pj≥0, withbi=viThe time is the same;vi＞pj＞bi, do not obtain the item, at this time the buyer i utility is 0, the utility is less thanbi=viTime-utility ( Get the item, buyer i utilityvi-pj＞0 )Ifpj＞vi＞bi, do not obtain items, at this time the buyer utility is 0, utility andbi=viThe utility is the same ( The utility of the non-obtained item is 0 )。
To sum up, tell the truth ( bi=vi ) is a weak-dominant strategy, and the reasoning is complete.
The reason for studying incentive compatibility is that it can guide advertisers to tell the truth and participate in bidding with their true estimates of exposure, and only advertisers who believe that value for money will continue to participate, creating a healthy bidding environment for the advertising platform, which is conducive to long-term development, such as Facebook's VCG mechanism to meet incentive compatibility (configuration rules: high price, payment rules: buyer i pays the amount equal to the loss to other buyers due to their participation) But obviously not everyone thinks so, the next section we show how to choose a mechanism to maximize the seller's return, although it is not incentive compatible, nor effective and fair, but may wish to have a large number of followers, this is probably the so-called choice of different bar.
▌ the optimal mechanism
Auctions have two objectives: seller returns and social utility. Different objectives determine the choice of different mechanisms, which lead to different probability functions and expected payments for bidders to obtain items, as well as the characteristics of the mechanism, such as incentive compatibility, personal rational budget balance, etc. As we just said, in a direct mechanism, a mechanism that meets the incentive compatibility feature encourages advertisers to tell the truth ( Quotes - Valuations ) is a weak possession strategy ( U≥U' ), will form a real balance, advertisers will continue to participate in bidding, conducive to the long-term development of the advertising ecology, but there is a phenomenon: it can not guarantee that each auction to maximize the auctioneer's income, now let's explore how to choose the mechanism to maximize each income, this mechanism is called the optimal mechanism.
Our assumptions remain the same: an indivisible individual auction with N potential buyers and Buyer i's valuation of the itemvi ( Suppose the buyer knows the value of the auction to himself and is not affected by other buyers, called private value )The seller does not know the buyer's true valuation, but knows the cumulative distribution function F of the buyer's valuationi(v), and the continuous density function fi(v)，FiIndependent of each other, assuming that the range of V is 0, w,fi≥0。
Now construct a function that values V, Y(v) s v-1-F(v)]/f(v), called the buyer's virtual valuation function, assuming that Y is an incremental function of v, which we call a general problem. The allocation rule is the high price, the price at this time refers to the buyer's virtual valuation price, that is, the auction item is assigned to the highest virtual valuation bidder, because we do not assume that the buyer is symmetrical, so different buyers have different virtual valuation functions, the slope of different functions may be different, the highest valuation of the buyer virtual valuation is not necessarily the highest, so the auction under the optimal mechanism is not a fair mechanism;-i"A is the highest virtual valuation other than buyer i, then buyer i's payment at this time."Y-1(A), whichY-1Is the buyer i virtual valuation function of the counter-function, we call this mechanism is not the optimal mechanism without reserved price.
Let's show it more intuitively with specific graphics, assuming that there are two buyers 1 and buyer 2, the virtual valuation function is shown below, when their valuation is v1
For advertising systems, Advertiser 1 pays a price that is not v0, because ads are auctioned according to ecpm, and we're now pushing back the actual price that the winning advertiser should pay, assuming a click-by-click billing ( cpc ), pCTR is the estimated click-through rate, then:
The competitive value of Buyer 1Ecpm1=1000*pCTR1*Y1
The bid value of Buyer 2Ecpm2=1000*pCTR2*Y2
Assume.Ecpm1＞Ecpm2, then Buyer 1 gets the exposure, Buyer 1 clicks to pay the actual price is:
Y-1[Ecpm2/(1000*pCTR1)]
Y-1Is the counter-function of Buyer 1 virtual valuation.
Although the actual billing calculation of advertising is one step more, but the essence has not changed, from figure 1 above, we can see that the valuation of buyer 2 is clearly higher, but the item was assigned to buyer 1, obviously not a fair bidding environment, nor is it a valid auction, which involves price discrimination. Look at people under the dishes, know that you are a rich master, you want the same things have to pay more than others, how do you know you are a rich? This is based on the system's understanding of your past, the collection of historical data, through the construction of simulations to come together your valuation distribution function, some big data is familiar meaning, intuitively feel very unfair, but in real life price discrimination everywhere, such as industrial electricity and residential electricity prices are different ( Three levels of discrimination )Taxis are charged on the steps of the kilometers and on the personal tax ladder ( Secondary discrimination), monopolists different buyers different prices (first-level discrimination) abound, but why the use of virtual valuation function when the seller's return is maximized, which involves an explanation of economics, limited by space, we will be interpreted in the following article.
The author describes:
Morphick, an advertising practitioner with a strong interest in advertising.
Link to original text:
https://www.zhihu.com/people/mophic/posts
If you like this article, please click on the upper right corner, share the article to the circle of friends
Welcome to DataFunTalkCalculate the adCommunication group, with peers zero distance communication. If you want to enter the group, please add the weChat (micro-signal) of the students who skip class:DataFunTalker ), reply:Calculate the adSkip class will automatically pull you into the group.
——END——
Go to "Discovery" - "Take a look" browse "Friends are watching"