ChatGPT burst onto the scene in 2023, and in just over half a year, it has become proficient in understanding and generating text, comprehending and creating beautiful images, and interpreting and writing code, bringing endless possibilities to various industries. From initial popularity in tech circles to now attracting interest from an increasing number of people around us, it seems we are not far from the AI era.

However, some people are skeptical about ChatGPT’s popularity, believing it’s just a temporary trend that won’t significantly impact our lives. So, how popular is ChatGPT really? How deep is its penetration? This article will discuss ChatGPT’s penetration from three aspects: search trends, application scenarios, and user characteristics.

ChatGPT's Penetration

Many people feel that when ChatGPT first came out, its popularity was high, with everyone discussing AI replacing humans and general artificial intelligence. But as time passed, they found it wasn’t as intelligent as imagined, so its attention is gradually declining. To verify this view, we can look at ChatGPT’s search trends through Google Trends.

Google’s search trends measure the number of times keywords are searched, which is a relative number between 0 and 100. In the selected region and time range, the moment with the highest search trend is assigned 100 points, representing the highest search volume for that keyword in this region and time period. If at a certain moment in the same region and time period, the search volume is half of the highest point, then the search trend at that moment is 50 points. If the search volume for a keyword is too low for Google Trends to gather enough data, then the search trend for that keyword is 0 points.

Google's global search trends for ChatGPT remain high

Globally, ChatGPT’s search index remains high, with search numbers peaking from March to May 2023, then dropping slightly, and starting to climb again in August, currently still at a high level.

For China, it reached its peak in February, then started to decline after March and April, and has now stabilized at about 1/5 of the previous search volume. There’s another indicator here, search trends displayed by region, with China at 100, meaning that in China, ChatGPT appears most frequently among all Google search keywords. In fact, the region can be further subdivided, such as into provinces and cities, with Qinghai Province having the highest regional search trend.

Google's ChatGPT search trends by province in China

Google’s data doesn’t accurately reflect the domestic situation in China, after all, due to special environmental reasons, not everyone can use Google search. To more accurately reflect the domestic situation, we can look at Baidu index or ByteDance’s index.

Looking at the indices from these two companies, in terms of total search volume, ChatGPT’s search volume peaked in March and gradually declined thereafter, basically consistent with Google’s data.

Baidu's nationwide search keyword trends for ChatGPT

Looking at specific provinces, in terms of absolute search numbers from Baidu, Guangdong Province has the most searches, followed by Beijing, Jiangsu, Zhejiang, and Shanghai. In addition to web searches, ByteDance also has search data for Douyin and Toutiao. Taking Douyin as an example, besides giving the top five provinces in terms of search numbers: Guangdong, Jiangsu, Zhejiang, Henan, and Shandong, there’s also an interesting city-level breakdown, as shown in the following image:

ChatGPT search numbers by city tier on Douyin

We can see that although first-tier cities don’t have the highest proportion of search numbers, they have the highest Target Group Index (TGI). This might be because their total population is not as high as new first-tier, second-tier, and third-tier cities, so the total search numbers are not high, but the group’s interest in ChatGPT is relatively high.

Demand Mapping

Previously, we looked at search trends from the perspective of ChatGPT’s single keyword search numbers, proportions, and regional distribution, but for each specific search, it might be based on different needs. For example, people might want to know:

  • How to register for ChatGPT?
  • How good is ChatGPT’s chat ability?
  • Why can ChatGPT generate images?
  • How is the voice chat experience with ChatGPT?

These questions are based on different needs and also reflect people’s interest in ChatGPT’s specific capabilities. To better analyze the focus on ChatGPT, we usually analyze through related queries. Google trends has related queries, Douyin has association analysis, and Baidu has demand mapping, all basically aimed at analyzing what kind of demands are based on.

Let’s first look at the search demand analysis within the global scope. The data provided by Google Trends here is relatively simple, with only related topics and related queries. Related topics are what users who search for ChatGPT also search for. Related queries are similar, showing what keywords users who search for ChatGPT also search for. Both of these data have two metric sorting methods, and Google only provides the top-ranking content.

  • Top. The most popular topics, with scores calculated on a relative basis: the most frequently searched topic gets 100 points, a topic searched half as often gets 50 points, and so on.
  • Rising. Related topics with the biggest increase in search frequency since the last time period. Topics marked as “Breakout” might be new topics with almost no previous search numbers, so their search frequency increase can’t be evaluated.

Let’s look at the related topics and related queries globally over the past 30 days, sorted by the rising metric. The results are as follows:

Google's rising related searches data for the last 30 days

We can see that the DALL-E topic is popular, as well as ChatGPT vision in related queries. However, the top rising search term parafrase is a bit strange; looking into it, it’s mostly searched in Indonesia and doesn’t have much connection with ChatGPT. Looking at the top results, related topics are artificial intelligence, OpenAI, etc., and related queries are also mostly normal, like ChatGPT login, ai ChatGPT, and so on.

Douyin Association Analysis

Douyin’s association analysis currently supports selecting a one-week period and then analyzing search-associated words and content-associated words, somewhat similar to Google’s related queries and related topics. The following image shows the search-associated word analysis for the period from October 16, 2023, to October 22, 2023:

Douyin's search-associated word analysis for the last week

We can see that ChatGPT’s related searches on Douyin are surrounded by a series of related keywords. These keywords are represented by dots, with their relationship to ChatGPT represented by their distance from the center - the closer, the more closely related. Larger circles indicate higher search index and more people searching. Red dots indicate rising search index, while blue dots indicate declining search index. You can also hover your mouse over a related keyword to view specific search content.

The top related queries here include “how to download”, “Android phone”, “PC version”, “write thesis”, “female version”, “dialogue”, “paid”, “chatbot”, etc. There are also some strange ones, like “horror” in the above image. At first, I didn’t understand why it would be associated with ChatGPT. After hovering the mouse, I found out it was searching for ChatGPT horror dialogue, which was quite unexpected. These basically reflect the main needs of Douyin users when searching for ChatGPT.

In addition to search-associated words, there are also search-associated contents, which are mostly artificial intelligence, AI, and the like. We won’t display them here.

Baidu Demand Mapping

Douyin’s search association analysis mainly focuses on searches within the Douyin APP, and many people might search after seeing related videos, so the keywords include dialogue, female version, writing thesis, and so on. For web searches, the results might be different. Here we refer to the demand mapping in Baidu Index, with data from the last week shown in the following image:

Baidu's search demand mapping for the last week

You can also choose the time range based on the time progress bar below. Currently, Baidu supports viewing on a weekly basis. From the above image, we can see that the associated words for web searches here are not consistent with Google and Douyin. On Baidu, the search content mainly focuses on the following:

  • Free ChatGPT Chinese version
  • ChatGPT Chinese version webpage
  • AI writing
  • ChatGPT trial

I tried selecting other time periods, including when ChatGPT first came out in March, and the relatively cool period in June and July. The top search contents in Baidu’s demand mapping basically revolve around how to use ChatGPT and other related content. It’s all OpenAI’s fault for setting so many restrictions and not letting us use ChatGPT.

User Characteristics

We’ve already looked at overall search trends and demand mapping. Now let’s see which user groups are more interested in ChatGPT. We’ll mainly analyze from aspects such as age, gender, and interests. Google doesn’t publish user demographic data for searches, so we can’t see the demographic distribution of keywords on Google. For domestic users, both Douyin and Baidu have this data, which we can analyze through Douyin’s user portrait and Baidu Index’s user characteristics.

Douyin User Portrait

Douyin’s user portrait is shown in the following image. Looking at the TGI (Target Group Index), the 18 to 23 age group is the highest, and then the index decreases with age, showing that young people are more interested in ChatGPT. In terms of search proportion, the 31 to 40 age group accounts for the largest share, with the 18 to 30 age group having a similar proportion. The proportion of those over 51 is quite small, indicating that ChatGPT doesn’t have much attention among the elderly. Looking at gender, there’s a significant difference between men and women, with males having noticeably higher search proportions and TGI index than females, suggesting that ChatGPT is more attractive to males.

Douyin's user portrait for ChatGPT

Douyin also provides the interest distribution of ChatGPT-related users. In terms of search proportion, the top five are: fashion, food, travel, culture, and sports, and these groups also have relatively high TGI indexes. It’s possible that in the interest distribution of the population, these labels originally had a large base of people, and they are also more receptive to new things, so they show more interest in ChatGPT.

It’s worth noting that the highest TGI index is actually for the technology category, but their search proportion is not high. This might be because there aren’t many people in this group to begin with, and they’re already quite familiar with ChatGPT and using it proficiently, so they don’t search for the ChatGPT keyword anymore.

Douyin's user interest categories for ChatGPT

Baidu User Portrait

Baidu also provides a user portrait for searches, and from the results below, the age and gender distribution is basically consistent with Douyin’s. However, in addition to search proportion and TGI index, Baidu also provides overall network distribution, which shows the population distribution of various groups and can serve as a reference for analysis. For example, we can see that the overall network distribution of males and females is basically the same, but in terms of search proportion and TGI for ChatGPT, males are significantly higher than females.

The interest categories provided by Baidu are slightly different from Douyin’s. The top 10 interest performances are film and music, education and training, etc., while the highest TGI are software applications, home appliances and digital products, games, etc.

Baidu's user interest categories for ChatGPT

Embracing AI

Overall, ChatGPT’s search popularity surged from the beginning of the year, reached a peak, and then fell back somewhat, but it still maintains a relatively high level without a cliff-like drop. Whether globally or domestically, ChatGPT is gradually penetrating into various groups of people, with more and more people becoming interested in it. This indirectly proves that ChatGPT, as a new generation of artificial intelligence achievement, has broad application prospects and is absolutely worth trying.

For what can be done with ChatGPT, you can refer to my previous series of articles, such as:

For how to better use prompts to ask questions to ChatGPT, you can refer to my previous Chinese guide based on OpenAI’s official best practices, a total of 6 articles:

  1. Write clear instructions: GPT doesn’t have mind-reading abilities, so if you want a short output, you can directly tell it to be around 100 words. If you want an explanation that a child can understand, you can directly tell it to explain as if speaking to a 10-year-old child, trying to be as simple and easy to understand as possible. For detailed discussion, see the article ChatGPT Prompt Best Guide One: Write Clear Instructions.
  2. Provide reference text: GPT can talk more than the most talkative person, and its answers may be made up or completely off track. Just like some practice books can help students do better in exams, you can provide reference text to GPT to help it answer more accurately and reliably. For detailed discussion, see the article ChatGPT Prompt Best Guide Two: Provide Reference Text.
  3. Break complex tasks into simpler subtasks: For tasks that are too complex, the current GPT4 still struggles a bit and has a higher error rate. Therefore, you need to break down the task into granularities that GPT4 can handle, and then assemble the results of each step yourself. For detailed discussion, see the article ChatGPT Prompt Best Guide Three: Breaking Down Complex Tasks.
  4. Give GPT time to “think”: If asked to multiply 17 by 28, people wouldn’t immediately know the answer, but could take time to calculate it. Similarly, you need GPT to use a chain of reasoning to think step by step, rather than immediately giving an incorrect answer. For detailed discussion, see the article ChatGPT Prompt Best Guide Four: Give the Model Time to Think.
  5. Use external tools: A major reason humans have become the rulers of all things is that we use tools. Similarly, you can compensate for GPT4’s weaknesses by providing outputs from other tools to GPT4. For example, a code execution engine can help GPT4 do math and run code. For detailed discussion, see the article ChatGPT Prompt Best Guide Five: Leveraging External Tools.
  6. Systematically test changes: If you can systematically evaluate GPT4’s capabilities, you can gradually optimize and improve its abilities. In some cases, modifications to prompts may improve performance on some individual examples but lead to overall performance degradation on more representative examples. Therefore, to ensure that changes to prompts have a positive effect, it’s necessary to define a comprehensive test suite. OpenAI has open-sourced the evals evaluation tool. For specifics, see the article ChatGPT Prompt Best Guide Six: Systematic Benchmark Evaluation.

Finally, it’s worth reminding that ChatGPT is not yet artificial general intelligence. It can sometimes be foolish, have hallucinations, and make things up, so you should verify ChatGPT’s answers. You can look at the article Real Examples Showing How ChatGPT Can Make Things Up! to understand ChatGPT’s embarrassing moments.