What is the algorithmic mechanism of news feed advertising?Information flow advertising formula

Advertising is a game that spends money to compete to buy volume, and information flow advertising is no exception. It still has to comply with the two core indicators of cost (CPA=cpm/ctr*cvr) and volume (conversion volume=exposure*ctr*cvr). It is also necessary to optimize the funnel factors at all levels.

What is the algorithmic mechanism of news feed advertising?

Information flow advertising is just because of the addition of intelligent algorithms, we need to study not only competitors and users, but also machine algorithms, because it is estimated first, and this estimate involves whether it is accurate, high or low And the problem of machine reward (platform advertising return), which is related to whether you can cross the circle (0-1 exposure cold start and follow-up model maturity scale).

In addition, because it is a form of advertising to find people, the update of materials (targeting is only to delineate the coverage of the crowd, creativity is the key to attracting it) and thresholds (on the one hand, users are refreshing on the platform for no purpose, and the platform The content threshold is high, and the imitation of competing products will lead to the lack of attraction of the material) is higher than other advertising forms.

Information flow advertising formula

Therefore, in response to the appeal, our analysis of information flow advertisements found that we have to solve the following three core problems, which are explained in order: (the core goal is the auxiliary results of the first two)

1. Machine Algorithms: Cold Starts and Models

We know that the platform’s advertising revenue is to maximize the ECPM value (ECPM=cpa*Pctr*Pcvr*bid). Considering various frequency control factors such as user experience, the order is based on the ECPM value.In this formula, the only thing that can be determined is your bid cpa (the bid factor is adjusted according to whether factors such as cost and budget meet the needs of advertisers). The difficulty lies in Pctr and Pcvr, which is the estimated probability of exposure to conversion. The estimation is not fabricated out of thin air, it needs historical data reference, given a prior probability, after the real exposure, the data feedback is obtained and new parameters are added and then adjusted.And this historical data is the reference of past converted user characteristics, materials, accounts, industries, etc.After each estimation, the real feedback data is exposed and then the parameters are adjusted to determine whether to enter the next traffic pool.The lower estimate is less exposure, the higher the estimate is, the higher the cost is, and the estimate is consistent with the actual data. (The real data is high and the volume continues to increase, and the real data is low to reduce the impact factor of volume optimization).

(1) Cold start

An old account or plan will have historical data as a reference. For a new account and a new plan, how to estimate without data?Therefore, there is also a trial and error cost and trial and error time until the stability of the model is satisfied, (the number of models is stable, the more the number, the more accurate the model).We need to obtain real data before we can make a judgment. When the exposure fails, it may be that the system thinks that the ECPM value is indeed not high. We can optimize the factors that may be considered, but it is also possible that we think it is good, and the system thinks it is not good. have to use other methods.It takes 1-5000 impressions or more to get at least 10000 conversion.

In order to pass the cold start as quickly as possible until the model is stable, here is a formula,Cold start = high bid DMP crowd package industry package narrow first and then wide historical material budget time

High bid: higher than the industry average bid, such as 20% or higher, and then lower the high bid to see the real data feedback, it is also a reward for the machine, and bear the consequences of this high cost, but it needs to be combined with a small budget and adjusted after getting the data feedback .If there is data feedback from the high bid, it can be lowered again, and if there is still volume, the test is successful.

DMP crowd package: When there is no data to refer to in the advertisement, it is like a machine looking for a needle in a haystack and cannot test the probability one by one. In order to reduce the cost and time, the converted crowd (not the platform advertisement) is used to encrypt and upload the ID package, and let the system In this crowd model, expand the test.

Industry crowd package: If you don’t even have historical conversion data, you can use the industry crowd package. The predecessors have already helped you out of the model, and the more mature the industry is, the more accurate this crowd package is. Of course, it can also be found in Crossover operation is performed in DMP to obtain its own crowd model.

Narrow first, then wide: If the above two help systems are difficult to build models, such as some unpopular industries, it is recommended to use other conventional orientation methods to first narrow and then wide to test. (Advertising users are screened out of the precise targeting crowd, but the estimated exposure coverage also needs to be looked at).

Historical material: The estimation will not only combine the characteristics of the user (the appeal model is), but also depends on the creativity and page. Here, the history can be combined with other accounts or advertising traffic materials and industry traffic materials in the past.Copy or learn from the core points in the running material. (Copywriting, pictures, scenes,person, props, music, duration, etc., break up a creative material, disassemble it, disassemble it, and reassemble it. )

Budget: This is the premise that affects the volume, and the minimum value will be taken in combination with the account, balance, plan, group, and advertisement. (Other details explained below)

Time: At present, each platform has different time for cold start, at least it is recommended to observe it for 2-7 days.

2. Calculation model of information flow advertisement placement

1. Quantity

The greater the number of conversions, the more sufficient the data dimension, and the more accurate the estimation can be.Now the platform has 0 numbers directly into the intelligent algorithm (also based on sufficient data in similar industries).The requirements of each platform are different, 6, 10, 20, 50 or even more, that is, the budget of an advertisement must be sufficient to achieve model stability.But it also depends on the cost of this transformation in your own industry and your own budget capabilities.If the industry is a few yuan or tens of yuan, then even 50 conversions will be a thousand yuan budget, but in some industries, the average CPA can reach hundreds or thousands, you can set a minimum conversion data budget to prevent the cost from being too high. .

2. (Crowd conversion material)

In terms of disassembling the model, it can be understood that different demographics see different conversion methods for different materials, and even the level of bids will affect the model (the traffic pool for testing is different).The more in-depth conversion methods (such as direct purchase, or even the purchase of different customer unit prices, such as 1 yuan and 9 yuan, 49 yuan products.) The more difficult it is, of course, this also depends on the industry. (If there are industries such as educational forms and popular purchases, it is recommended to use the same method to learn from reference data).

2. Material update

We will draw on our own or industry historical data as the estimated prior value so that the model can be found smoothly.But after going through the model, it is bound to face the decline of the material.Moreover, as mentioned above, the core of the information flow advertisement is the material, and the orientation only defines a covered group, allowing the system to find these characteristics, but in the end, whether the user acts or not depends on the material.This involves the amount of material, the frequency of new releases, the selling point, the form of expression, and the source of inspiration. (detailed below)

3. Core objectives: cost and volume

The optimization of the above two problems still needs to return to our final core goals: cost (CPA=cpm/ctr*cvr) and volume (conversion volume=exposure*ctr*cvr), which needs to be disassembled like SEM advertisements. It is to solve the influence factors of exposure, cpm, ctr, and cvr and the optimization actions that can be performed.

(1) Exposure

1. External factors: daily activities of the platform, duration, user tonality, competing products (quantity, schedule, bid), holidays, frequency control (such as large pictures, the number of similar advertisements, etc.)

2. Internal factors: orientation, ecpm value (cpa*Pctr*Pcvr*bid), budget, time period, multiple accounts, advertising space, material type (whether all categories), billing mode, running volume mode, etc.

(2) ctr

Advertising space, material, style, time period, crowd, etc. (It still depends on the external changes of the market and user environment)

(3) cvr

Crowd, page (content conversion entry), creative page relevance, etc.

(4) cpm value

Own bidding, industry competition, platform-based bidding

0. The 1~XNUMX entry rule of the information flow advertising algorithm model

Here we will further refine or supplement, what steps need to be done in the 0-1 process of an information flow advertisement?

A good advertisement is to impress the right people (targeting, crowd model) in the right way (products, materials, selling points) at the right time and in the right scenario (platform, advertising space), and at the same time, it needs to be scaled at a low cost.

company's product:

Only when the product has the monopoly differentiation advantage, the product has the transformation advantage, otherwise it depends on the channel competition.Most of them analyze the advantages of their company's products in the case of full market competition, avoid the strength of competing products, and can hit the pain points of users, so that they can be reflected in the follow-up materials.After understanding the advantages of the company's products, you can find the selling point of materials that can be externalized.

(1) Company: Including the time of establishment, background, nature, scale, honor, service, cases and other dimensions to analyze, whether there is an externalized selling point.

(2) Product: Extract points that can be externalized from user concerns such as price, function, emotion, and scene.

Platform information:

(1) Data algorithm: including the daily activities of the platform, usage habits and duration, data dimensions, and orientation methods.

(2) User portraits: mainly to analyze the tonality of users of the platform, and which copy style and style they like.

User information: user portraits, user needs, user concerns, user consumption

(1) User portrait: natural attributes, device attributes, interest attributes, behavior attributes (search,E-commerce, social, APP, LBS)

(2) User needs: the underlying motivation and pain points of users to use your product/service

(3) User focus: that is, the reason why users choose you. (from product and endorsement)

(4) User consumption: consumption ability, consumption psychology, consumption concept

The above information can be used in index tools, keyword demand maps, industry reports, competitive product analysis, user survey interview feedback, community social comment platforms, advertising platform DMP portrait analysis, customer service sales interviews, CRM data analysis, etc.

Competitive product information: It mainly analyzes its material externalization selling points and company product information, and finds differentiated selling points that can avoid its advantages but satisfy target users.

Crowd segmentation: core, target, potential audience and how to target them

Core targeting: words (such as brands, competitors), dmp conversions, behaviors (follow, search, purchase, download, LBS itself or competitors)

Targeting: words (such as generic products), industry packages, primary core interests

Potential orientation: words (such as crowd, industry words), secondary and tertiary related interest packages

Creative page:

(1) Different groups of people can adopt different creative selling points, such as the core group’s main brand and activities, the target group’s main differentiated product selling point, and the potential group’s main focus on welfare discounts and creating interest desires, pain points and anxiety amplification, etc.

(2) Take education as an example: people (teachers, students, teaching assistants, parents, single/multi-person), machines (props), materials (textbooks, gift boxes, books, pens, notes, mind maps), methods (methods, Skills, knowledge points), and the related factors involved in the ring (classroom, family, community) are dismantled and combined.

(3) Forms of expression: graphic (three pictures, big picture, small picture, grid, angle), video (oral broadcast, plot, hand-painted, ppt...).

(4) Test sequence: as many as one, then from one to many. (Multiple selling point material forms test, find out the running volume material, and extend around the material).

(4) Page information: the same principle as the SEM page part (especially note that the header image and the outer layer are strongly related or even consistent, and the creative image is directly converted).

(5) Sources of ideas: creative inspiration tools for advertising platforms, manual reading, tripartite crawling tools, keyword demand maps, etc.

Bid budget:

1. Budget

（1）、1.5-2倍转化目标数量预算。（如单日100转化量，cpa为100，则可设置15000-20000）。

(2), it is best not to be less than 1.5 times the actual consumption budget. (If the actual consumption is 10000, it should not be lower than 15000).

(3) Accounts and ad groups can be set. There is little difference between plan settings and the final budget depends on the minimum value of the balance, account, plan, and group, and the actual available balance of the advertisement will be taken.

(4) New material advertisements are added every day for backup, and the budget for the number of stable models converted per day should be set aside for advertisements that are online at the same time. (For example, in industries with high CPA, set aside 6 CPA budgets for 1 advertisement). If the CPA is 100, the budget for a single advertisement should be at least 600. If the daily budget is 1200, it is recommended to launch 2-4 advertisements at the same time.Observe the data of the first 24 hours, promptly remove advertisements with bad data, and roll out new ones.

2. Bidding

(1) Bid by industry and search or acceptable CPA, and increase by 5% on the basis of the suggested bid.

(2) If it is not possible to start in a cold environment and there is still no data, increase the bid until there is data performance. (Exposure more than 3000-5000 and then observe and adjust)

(3) If there is still no data feedback, a combination of billing and running volume models, small budgets, and shallow conversion goals can be used to accumulate conversion data and view materials and crowds. (such as cpm, cpc fast running volume).

data analysis:

Vertical: Focus on cost (CPA=cpm/ctr*cvr) and volume (conversion volume=exposure*ctr*cvr) and sorting formula ECPM=cpa*Pctr*Pcvr*bid to analyze which link data is lower than the average of the market, and the core is the worst The problem lies in finding the influencing factors that can be optimized in this link.

Horizontal: Platform, account, business, plan, group, advertisement find out the core difference dimension that affects the target from the whole to the part, and optimize around this dimension.