The basic idea of ​​information flow data analysis, what data should be considered for Baidu information flow optimization?

Information flow advertising often encounters various account problems, but people do not know how to effectively eliminate and solve them. They can only perform simple operations, such as: controlling the planned budget, adjusting the planned bid, frequently creating new plans, directly pausing the plan, etc., and only a small number of people know why they do this and what impact such operations will bring.

This is why many advertising specialists have worked for many years but have not become excellent advertising optimizers and are merely backend advertising staff. This problem arises because we do not have data support or basic data analysis thinking before we operate.
Then the next question is, for advertisers, the backend operations are already very familiar to them. When we are operating, have we thought about the following questions in our minds?

• What problems are there with the account?

• What is the purpose of doing this?

• Why do you have to do this?

• What impact will it have after the operation?
I think some people have never thought about the above-mentioned phenomena, which is why I will talk about the basic ideas of data analysis of information flow next. The steps are: we raise effective questions → establish reasonable hypotheses → find supporting data → integrate and analyze data → verify the analysis results (using Bytedance as an example to explain).

1. Ask effective questions

After discovering plan anomalies, identify the issues according to different stages of the plan life cycle, which are generally divided into three time periods: learning period, stable period, and decline period.
1. The learning period is when a system is in the process of learning and exploring potential target users. During this period, there will be the following problems, such as no display volume or very little display volume for the newly created plan? The newly created plan has no clicks or has a low click-through rate? The new construction plan has not been converted? And so on...
2. The stable period is the process in which the system has collected a certain amount of converted people and can find the target population more accurately. During this period, there will be the following problems: the actual bid is too high, the click-through rate suddenly increases/decreases, the conversion rate is low, and the conversion cost is high. And so on...
3. The decline period is when a plan suddenly shows a decrease in display volume, a decrease in conversion rate, an increase in conversion cost, and a decrease in conversion volume. When this situation continues for several days, it means that the plan has begun to enter the decline period. During this period, there were intermittent instabilities such as smaller display volumes, higher conversion costs, and a sharp drop in spending power, etc.

2. Establish reasonable assumptions

What should I do if the new plan does not show any results during the study period?

The first step is to think about the influencing factors and list reasons 1, 2, 3, 4, etc. for example:

1. Non-delivery time

2. Limited by flow control

3. Low Bid

4. Narrow or wrong orientation

The second step is to determine the reasons for the output of these listed factors.

Non-delivery time: It may be that we only checked the evening delivery during the delivery period, and the delivery time has not arrived during the day, or we checked the delivery on a certain day, or the delivery has been completed, etc. Output reasons. For example, the targeting is narrow or wrong: that may be due to the wrong selection of filtering area, multiple redirections, checking a very narrow targeting, etc.
The third step is to investigate each one according to the reasons for output. Whether the promotion time has been set, whether the budget is too small and the bid is too high, causing the advertisement to trigger flow control and affect the display, whether the bid is too low compared to the industry average, and whether the targeted coverage population is too narrow.

3. Find supporting data

Find data that can support the conclusion as evidence to support the hypothesis. Common channels for obtaining data include:

1. Data reporting/plan diagnosis

For example, whether it is during non-delivery time, you can view the specific delivery time by editing the advertising plan in Ocean Engine or clicking on the details of the advertising plan.

You can also check whether the bid is too low by looking at the industry analysis or plan diagnosis in the account diagnosis in the Bytedance Engine tool.

These two pictures basically show whether the quality and bid of the plan are within the industry average.

2. Monitoring and tracking (application download API/deep conversion SDK and data/platform landing page monitoring (page insights))

3. Third-party platforms

4. Authoritative industry data

4. Integrate and analyze data

After acquiring the data, it is necessary to integrate and analyze the data. This can be done synchronously when acquiring the data. Sometimes tools are needed for further integration and analysis.

1. Advertising management platform (data reporting/crowd insights, planning diagnostic tools)

Integration and optimization of three types of reports of Bytedance, advertising reports: account/group/plan/creative reports, audience analysis: post-investment user portraits, advertising products: programmatic creativity/video advertising/material statistics, etc. A variety of visual charts are provided, including time trend (line chart), cumulative comparison (bar chart), and proportion analysis (fan chart).

2. Data Platform

3. Third-party data platform

5. Verify the analysis results

There are two possible outcomes:

The first type of data verification is consistent with the results. Of course, everyone is happy at this time because the direction is correct. The previous process took a lot of time and energy, and the efforts were not in vain in the end.

Don't be happy too soon. Of course, there is another situation. The data verification is inconsistent with the results. Oh my God, then we may need to start over to check other reasons that we have not checked yet. It is also possible that the data we used was calculated incorrectly. Optimization itself is a matter of constantly asking questions → assuming reasons → finding data → analyzing data → verifying results. So we have to go further and further on the road of optimization.

Finally, I want to say that every hypothesis, collection, and analysis is like a summary and improvement of oneself, and a precipitation and sublimation of a stage.