Website Data Analysis

In the past few months, I have been working in a Canadian company that advertises customers on Facebook. Let's briefly talk about their attitude towards data. This is a small Startup company with less than 20 people in total. Four of them (including me) are technology, and all the remaining CEOs are Account managers. Of course, the CEO is also doing Account Manager many times.

When I first arrived at the company, I felt that their code was very bad and their database design was also very bad. Later I learned that when I started up, I went to an Indian company for outsourcing. They were not satisfied with the outsourcing. So after the first phase of the project was completed, I took it all and made it myself. However, the sequelae are also left behind.

The company's data model is clear that as long as it can make money at a price lower than the CPA price given by advertisers, it will find a way to increase advertising coverage. However, everyone understands that increasing the coverage rate may lead to a decline in the conversion rate. However, if we accept this assumption, there is no chance to make money. It is precisely because they believe that apart from common sense, there are still some things beyond experience.

For example, keywords ...... Some keywords are useful to some people and useless to others. If you don't do data mining and want to combine advertising words or keywords, you won't be able to make any money if you're exhausted.

So ...... The company designed several basic core algorithms in the Code:
1. A stop-loss trigger that automatically stops any loss advertisement.
2. A cron that automatically publishes advertisements, and the program has been scanning. Once you find that some advertisements can make money, you can freely combine these advertising elements and then automatically publish them to the advertising system. In this way, we can find some more profitable advertising forms unexpectedly.
3. I have done A lot of advertisement update algorithms and developed an automated A/B testing strategy to update the advertisement price based on the price fluctuations of Facebook ads.

I am deeply touched by these algorithms. The so-called data analysis is not a product manager who runs to O & M. The database administrator or engineer said: What data do I want to track now? Could you help me out. Then, you can check whether the data is reasonable.

In this company, as long as a data model has an impact on revenue, it will be directly coded into the system and become the automatically executed code. Based on this data-oriented principle, the Code is subject to an infinite number of reconstruction operations, because no one knows where the next data mode will occur and the relationship between the data.

I think many companies in China are still analyzing data in the daily report, but they are still talking about the data only to verify the product manager's ideas. Is this action too slow?

Next, let's talk about it casually. It may not be of reference value:

1.

For most websites, if you want to use data-oriented, you must establish A system-level A/B testing mechanism. For interface reconstruction, a product manager and an engineer can use the system for at least one day.
3-4. System-level A/B testing should ensure that the System-level A/B testing can be quickly launched and the data can be seen immediately. Once the critical value is exceeded, the testing should be completed, the data should be retained, and A report should be generated (directly sent by email, instead
Ideally, go to the background and check again)

2. for companies that are engaged in social networking websites or have complex user data models, a matching system should be established between the interface presentation and user data. In this way, the product manager can design several presentation modes and drop them into the matching system. In a short time, it will be able to discover the differences in the data presented by users, and then systematically solidify this mechanism.

3. Create an internal tag system for different users through cookies or user login information to see if these tags are significantly different in system 2. If yes, it can be solidified to improve key indicators.

Therefore, my current feeling for data analysis is:
1. To improve a data indicator, it is useless to stare at it. You must find several more operable data indicators that affect the data and adjust them.
2. The possibility of data analysis should be sufficient. The basis for full analysis is to test the possibility of a sufficient number. If you want to test whether the color of the icon will change from green to red. Why don't we test the blue, purple, and yellow colors?
3. If the problem can be explained by small-scale data, there is no need to extend the test period or expand the test scope.
4. Make full use of computers to help you collect and analyze data, shorten the data analysis cycle, and reduce the cost of data analysis.
5. If necessary, ask the computer to help you find pattern, because the computer is not biased.