Data Deep Dive: Google Trends, Part III
We’ve dissected Google Trends, from understanding the underlying data to cleansing and enriching it. In Part III, we dive into the metrics and analytics. Plus, an analysis for Taylor Swift fans.
Welcome to the Data Score newsletter, composed by DataChorus LLC. The newsletter is your go-to source for insights into the world of data-driven decision-making. Whether you're an insight seeker, a unique data company, a software-as-a-service provider, or an investor, this newsletter is for you. I'm Jason DeRise, a seasoned expert in the field of data-driven insights. As one of the first 10 members of UBS Evidence Lab, I was at the forefront of pioneering new ways to generate actionable insights from alternative data. Before that, I successfully built a sell-side equity research franchise based on proprietary data and non-consensus insights. After moving on from UBS Evidence Lab, I’ve remained active in the intersection of data, technology, and financial insights. Through my extensive experience as a purchaser and creator of data, I have gained a unique perspective, which I am sharing through the newsletter.
Google Trends is widely used as an alternative data1 source, yet it’s often misapplied in arguments, leading to misleading conclusions. The data itself isn’t flawed; rather, it’s the misunderstanding of the data that leads to misinterpretation and poor application.
This article concludes our three-part series on Google Trends. In parts I and II, we discussed common questions addressed with the data, explained the underlying data that is extracted from Google Trends, and explained how to cleanse the data so it is usable. In this final part, we discuss enriching the data for analytics, the limitations of the data and some action items to begin using the data.
In the Dataset Deep Dive Series, the format covers the following topics:
Common questions addressed with the data (Covered in Part I)
Underlying Data (Covered in Part I)
Cleaning the Data (Covered in Part II)
Enriching the data (Part III)
Limitations to consider (Part III)
Action items to begin using the data (Part III)
https://trends.google.com/trends/
The use case for this deep dive on Google Trends was sparked by a well-known YouTuber in the music industry who recently used Google Trends to argue that music is not as good as it used to be and that music is becoming less popular. I thought this question of the popularity of music compared to other activities provided a great low-key topic to explore via Google Trends, using the best practices for getting quality results. It also helps contrast good and poor uses of Google Trends. Check out Part I for more context.
Sponsored Advertisement
World of DaaS in DC
Thursday, August 29th 2024, 6-8PM
Join DC's leading data executives, investors and self-proclaimed "data nerds" for an evening of drinks, networking and new ideas!
Some people you can expect to see:
Sara Fagen, Founder & CEO, Tunnl
Jason Hines - CEO & Co-founder, Gigasheet
Auren Hoffman - CEO, SafeGraph. Fmr CEO, LiveRamp (NYSE: RAMP)
Michael Southworth - CEO, Babel Street
Register: https://daasindc.rsvpify.com/?securityToken=UphNFxe8R6K2Jt18lJffjslyMo74zHrb
Enriching the data
In this section, we’ll explore how to transform cleansed data into actionable metrics that align closely with the specific use case.
In Part I, we discussed the investment questions that can be answered with Google Trends data in the context of what the underlying data represents.
The primary use is to address investment debates, where knowing top-of-mind awareness of a topic would be critical to assessing the potential for outcomes. Top-of-mind awareness refers to how prominently a topic or brand is in people's thoughts, which is the first step in the marketing funnel toward trial and repeat purchases. High top-of-mind awareness can eventually influence consumer behaviors that impact a company’s sales and profitability, reflected in the company’s revenues and profits. However, combining top-of-mind awareness with a dataset covering purchase or usage behavior allows for a deeper understanding of the conversion of awareness to purchase and usage.
In this section, we will discuss metric creation and analytics, including the types of analytics.
New product and media launches metrics
Assess potential success of a new product launch compared to prior product launches based on initial search activity
Assess if the latest media release is likely to be a blockbuster
Popularity growth, search rank, and market share
Assess if a disruptive innovation is gaining mindshare
Assess if new consumer behaviors are long-term trend or fad by monitoring awareness
Behavior shifts with economic impact
Assess the rising public consciousness of topics and themes
I’m going to continue to use the debate outlined by YouTuber Rick Beato as a data examples. The debate is whether music is less popular than other entertainment sources like video games and social media (though we showed that properly sourced and cleansed data would lead to a different conclusion).
Link to the video where Rick Beato tried to use Google Trends to make his point that music was becoming less popular: https://youtu.be/TU96wCDHGKM?si=F1e8zXQw_yxhe_cQ&t=185
I’m continuing with that debate because it’s a low-risk, non-investment debate to demonstrate how to effectively use Google Trends for data analysis. I wanted to pick a topic relevant for readers of the Data Score without turning the article into a full investment thesis (this isn’t investment research). I also want to avoid writing about anything that would be useful in my day job. So this topic was a low-risk way to show how to think about Google Trends data collection while also showing some ways it gets misinterpreted.
Analytic Type and Metric creation
The creation of metrics depends on the type of analysis intended.
New product and media launches metrics
Assess potential success of a new product launch compared to prior product launches based on initial search activity
Assess if the latest media release is likely to be a blockbuster
Product launches with high expectations—whether due to significant investment or company communications—often spark critical investment debates. However, the actual results of the launch will not be known until after the product launches. Google Trends can provide an early indication of the potential success of the launch. This approach could apply to the next Apple product launch compared to prior launches (such as the iPhone 16 versus prior versions). It could apply to major movie releases compared to prior movie releases (e.g Will “Moana 2” be as successful as prior November Disney/Pixar releases?)
Can Google Trends predict the popularity of Taylor Swift album releases? We look at search activity at the release of the album and in the following weeks. Think about this as an example where you could apply the same logic to investment debates around new product launches, new movies and TV series, or events where there are historic comparisons with ground-truth data to bench market against and have an early view of the next big launch before the purchase data is available.
Let’s use Taylor Swift’s album releases as an example to illustrate how search activity can reflect the relative popularity of her albums. The goal is to show the analytic framework, metrics and visualizations with this use case, and less so about showing a highly correlated analysis.
I am using “swift” as the search term to capture all searches related to the singer-songwriter. I got comfortable that I didn’t need to create a term like “swift -code” or try to build up the term from “taylor -ann -lord -elizabth -james… etc” I sanity checked the data by comparing to the prebuilt search term for Taylor Swift, and its going to be accurate enough to just use "swift.”
I pulled multiple versions of “swift” with random numbers added to the search string to extract more than one sample of the search term, and then averaged it into a single time series. I used data for the last 5 years to get weekly data points from Google Trends. In addition, I put together a list of Taylor Swift album release dates, marked in grey bars in the data below. You’ll see search activity for Taylor Swift spikes on the album release weeks, with a couple exceptions, including when her Eras tour was announced and tickets went on sale. Peak Taylor Swift search activity was during this year’s Superbowl.
Using the album release weeks as the anchor for the analysis, we can compare all search activity in the weeks leading up to the Taylor Swift album release date and the weeks after. Effectively, the higher the search activity around the album release, the more likely the album is going to be a commercial success. The chart below shows each week of search activity, where T0 is the album release week. Note that albums have typically been released on Fridays, but Google Trends’ week definition is from Sunday to Saturday. We also look at the following weeks to understand the popularity before search activity returns to prior levels.
To ensure comparability, we need to adjust for the upward trend in Taylor Swift’s search activity since Q4 2022. I used the average search 5 weeks prior to the album release as the baseline and subtracted the baseline from each week’s search data value. This rebases the values to the pre-trend at a value of 0 for the prior 5 weeks before the release, allowing the post-trend search values to be comparable for each release.
Now the sanity check. I used data I found on Wikipedia (for better or worse) that showed US sales for Taylor’s albums. I ranked the table below based on the Google Trends search data for the week of the release and compared it to US sales.
This comparison looks really good to me, as the search is indicative of the success of the album. For her new album releases in 2019, 2020, and 2022, the rank order of search activity around the release week matched the actual US sales. However, the most recent album has the highest search, but US sales figures are ranked second.
Why would the most recent album have the highest search activity but not the highest sales?
Maybe it wasn’t a good album? I did a very scientific study and asked my wife and oldest daughter, who are “Swifties,” and they said it was a very good album.
I’m pretty sure the data on US sales is not comparable to the older albums because less time has passed since the album dropped. The Wikipedia page references data for US sales for the most recent album as of July 2024, while the other albums have been out for much longer and the data referenced is as of January 2024.
A better ground-truth data point to compare would be to find the sales for the corresponding weeks of the search analysis.
I am not a "Swifty,” so I don’t have the success of the albums memorized, and I decided it was not worth spending the time to find the corresponding time period sales figures considering this use case of Google Trends is solely for the purpose of showing how to do the Google Trends part of the analysis (sorry to the Data Score readers who are Taylor Swift fans for the lack of rigor on the ground truth data point!)
Nevertheless, I feel pretty confident that it is just a matter of time before this latest album goes beyond the other recent albums in US sales.
During this 5-year period, Taylor re-released her songs to gain ownership of her masters, in addition to her already retained rights as the songwriter. Surprisingly, Google search trends didn’t align as expected with these re-releases. I think I would need to do some thinking about the relationship between the popularity of the original release and the audience size that would care about making sure they purchased the versions where Taylor Swift owns the masters rights, such that Google Search activity would be unrelated. That said, I wanted to show these data points for full transparency.
Think about this as an example where you could apply the same logic to investment debates around new product launches, new movies and TV series, or events where there are historic comparisons with ground-truth data to bench market against and have an early view of the next big launch before the purchase data is available.
Consider combining search data with inventory data and reviews data
Keep in mind that search activity may signal high awareness and anticipation of the new launch, but execution of the launch could lead to differences downstream in the funnel. For example, a new product launch where high search activity compared to prior comparable product launches exists could be disrupted by supply chain constraints not allowing for demand to be met. Conversely, search may be relatively low, but the initial revenues could be higher than expected. It’s important to understand if the distribution channel has received and paid for product, but the offtake by the consumer is not there, the impact on the financials can happen at a later date. This is where combining data sets could reveal a more complete picture. For example, web-mined inventory data and web-mined review data could be combined with search data for a more complete picture.
Popularity growth, search rank, and market share
Assess if a disruptive innovation is gaining mindshare
Assess if new consumer behaviors are long-term trend or fad by monitoring awareness
Even without an event catalyst, Google Trends data can be a powerful tool for monitoring top-of-mind awareness in the early stages of the marketing funnel. This works best when there is something new happening in the industry, which means consumers are looking for information before making a purchase. Share of mind is the first step before trial purchases, eventually leading to regular use and finally brand and product loyalty. When something is newly disruptive in an industry, the rest of the funnel is unknown; however, awareness can be measured first.
Using the social media platforms discussed in Part II, we can convert the search activity into a year-over-year percentage change to compare the growth of new entrants, such as searches for TikTok, when growth was off the chart (growing well above the upper range shown below, capped at 200%).
In addition to looking at growth rates, which can be a function of a small base, we can convert the search activity to a share of the competitive set.
Please note: the analysis below is meant to be an example. This isn’t accurate because there's been a huge number of social media apps over the last 20 years that are not included in the analysis. This isn’t meant to be for investment research, and instead of spending days creating the full comp set, it’s best I just leave it to the analysis below as an example.
Changes in rank order typically carry more weight in investors minds (and quant models) than marginal changes in metrics like search activity. Within the compset, we can convert the search scores into a rank, where 1 is best. I like to show these charts on a logrithmic scale2 because a move from Rank #2 to Rank #1 is much more meaningful than, for example, a move from Rank #50 to Rank #49. Visually showing the logrithmic scale shows that it is increasingly hard to move up to #1 overall.
Again, the chart below is indicative of the types of analysis. This isn’t a completely accurate analysis because there's been lots of change in the social media landscape over the last 20 years with new platforms coming and going, and that's not included in the analysis. This isn’t meant to be for investment research, and instead of spending days creating the full comp set, it’s best I just leave it to the analysis below as an example.
There are a number caveats to consider with this type of analysis, which will be discussed below. But it’s worth calling out here
Threads is the number 1 ranked social media app on the iOS app store; I couldn’t separate threads from Instagram in search activity. There isn’t a prebuilt search term that I could find either. So it’s missing explicitly.
TikTok is the number 5 ranked app across all apps but is shown as 5th in search activity. The generation that uses TikTok doesn’t use Google to find TikTok.
The search for the social media platform formerly known as Twitter is not always an indicator of interest in using the platform, as the management change has led to lots of interest in what is going on at Twitter. So awareness is high, but maybe not for comparable reasons to the other searches. Note: Because it’s not possible to back out all reasons for searching for "X,” besides the new name for Twitter, I used the pre-built search term to extract the data for X.
The market share and rank analysis are very much dependent on having a full competitive set within the analysis, which, if done properly, could take days to do for this sector. This is just meant to be an indicative analysis and not investment research-grade analytics. This isn’t investment research.
Again, the above analysis is only meant to provide an example and not a properly designed analysis because of the incomplete competitive set (this isn’t investment research). For real applications of the approach, think about scenarios where an industry is seeing new entrants or a new product is disrupting the market. The category doesn’t need to be a digital category to work. I’ve seen successful applications in multiple consumer packaged goods categories where changes in top-of-mind awareness were early signals that consumer preferences for specific products were changing (for the better or worse of the new brand). Clearly, people aren’t going to search for an alcoholic beverage, plant-based protein product, or e-cigarette product every time they purchase the product. However, these examples did work with Google Search leading the way because the newness of the product or category led to the search activity in total, which was representative of overall top-of-mind awareness.
Consider combining search data with app usage, clickstream and transaction data
This is another example where combining Google Trends data with other data sets really enhances the understanding of what is happening and narrows the range of possibilities for future outcomes. Search data, app usage data, transaction data, and clickstream3 data can be combined to more deeply understand the underlying trends.
Behavior shifts with economic impact
Assess the rising public consciousness of topics and themes
This use case requires more creativity to think through, but the metrics could be the same as the above. I don’t have an easy example from the "Rick Beato debate that “music is becoming less popular.” Instead, I will note some past use cases of Google Search that worked as verified by ground truth data, including tracking material changes in unemployment claims and tracking COVID outbreaks by geography.
These types of analyses need both creativity in the creation of search terms and ground truth to ensure that the search activities are indicative of the thoughts of the population. The search terms used need to reflect what people actually search for when seeking information on the topic of interest from an analytical point of view.
For example, searches for the single word “debt” may seem like a way to track the population's focus on their debt. However, I think people would search for many topics related to their personal finances. Also, a simple search for “debt” would include searches related to “national debt” or policies like “student loan debt forgiveness.
Conversely, if we make the search terms precise, Google Trends will not return values for the full set of search behaviors related to the analysis. For example, if one wanted to understand consumer awareness of brands focused on sustainability compared to traditionally manufactured products, pulling data for search terms like "sustainable fashion" vs. "fast fashion" or "eco-friendly packaging" vs. "plastic packaging" is unlikely to capture what is really on the minds of consumers via Google Trends because the specificity of the search terms is limiting the results compared to search terms that reflect the actual search activity of the population.
Creativity in crafting search terms and ensuring robust sanity checking of the data can allow for this type of macro analysis.
Combine Google Trends data with other data types
In the use cases and metrics above, there are multiple times where I noted that the combination of Google Trends with other data sources becomes even more powerful. In "A Different Approach to Revenue Estimates Leveraging Alternative Data,” published in March 2023, I discussed how various data sources could be combined to understand the marketing funnel, from awareness to brand and product loyalty.
Limitations to consider
Even the most valuable alternative datasets have limitations and caveats to consider when leveraging the tool. Google Trends can be a powerful tool in the tool belt, but it has its limitations. Over the three parts of this deep dive, the common theme has been that the tool itself is easy to use, but the many nuances about the tool and data mean there are many ways to use Google Trends incorrectly and come to incorrect conclusions about an investment debate.
Here are some of the biggest limitations and caveats to consider:
The way the data is provided by Google Trends limits the types of analyses that are possible.
Google Trends returns values relative to all search activity over the time period and geography selected. It’s not an absolute search.
Term creation is complicated. Be careful about the dual meanings of words and consider all synonyms when crafting search terms.
Comparing terms requires the queries to include both terms and if more than 5 terms and geographies are needed, a control term is needed to stitch together the data. When combining data from multiple Google Trends queries, it’s essential to use a “control term”—a stable search term used in all queries—to ensure that the results are comparable across different datasets. (see Part II for a detailed discussion.) All queries will return values between 0 and 100, where 100 is the maximum share of the search for the query. This is a frequent source of misunderstanding and misuse of the Google Trends data by people unaware of the inner workings of the tool.
Remember, the share of search is relative to the geography. A value of 50 in Canada versus. 40 in the US does not mean there are more absolute searches in Canada versus absolute searches in the US; it’s that the term’s share of total searches in Canada is higher than the term’s share of total searches in the US.
Avoid worldwide geography unless the topic of investigation has the same geographic relevance as Google searches total relevance (the global will underweight China and overweight the US versus the global population).
Be careful with Google’s pre-built terms; they are a black box and subject to change as well as errors.
Top-of-mind awareness is not a purchase.
There is likely a strong correlation in the beginning of the product or service’s life as the public’s consciousness of the new product or service grows, which may mean the search data is highly correlated with purchase data. But that’s a risky assumption to make without acknowledging the steps in the funnel between awareness and purchase. Over time, as a product matures, there could be lots of reasons why the population begins to search for a product, but that doesn’t mean that they will purchase the item after the search. iPhone’s are a great example, where in the early years of both the iPhone and Google Trends data, searches for iPhones and iPhone sales were correlated. But over time, the number of searches for iPhones included other reasons for searching, like “lost iPhone” and “broken iPhone.”. The relationship between search and sales begins to break down over time.
Negative sentiment searches cannot be separated from positive sentiment searches. Searches driven by negative events or sentiments (e.g., scandals, controversies) are mixed in with positive searches, making it difficult to discern whether the search activity is due to favorable or unfavorable interest. A controversy related to a search term could cause searches to spike, but blindly assuming all searches for a topic are positive sentiment can lead to bad conclusions.
Search behavior changes over time. Search behavior does change over time as a reflection of the changes in how people find information and also changes as the information people need changes. For example, data from Spotify and other streaming services could be a better indication of the popularity of music, with searches happening in the music apps instead of on Google. Likewise, applications where there is no web browser interface would mean searches are happening in the app store and not on Google. Furthermore, some topics can become uninteresting to the population, but they are still happening without people searching Google for information.
Google Trends is better used for business-to-consumer (B2C) use cases, but it can also be used for business-to-business (B2B) use cases. However, be careful not to combine B2C and B2B searches in the same query because the volume of consumer-focused searches will materially outweigh the B2B use cases and cause the data return to have values less than 2, which means the data will behave accordingly.
The share of category search activity is a powerful analysis for top-of-mind awareness shifts, but its important to be careful to cover relevant universes of search terms. As new products and services enter the market, search behavior could change. If the analysis doesn’t pick up on that change by refactoring the series of search terms, it could lead to an incorrect conclusion.
Make sure to tag each run of the analysis with a point-in-time stamp4 because Google Trends can restate their model output. Also, as new peak values of search term activity are reached, the historic data will be restated, reflecting the new 100 peak on the chart.
Action items to begin using the data
Google Trends is a widely used but often misunderstood tool. This deep dive aims to bring clarity to its effective use, empowering you to leverage the data with confidence.
Access Google Trends: https://trends.google.com/trends?geo=US&hl=en-US
Don’t get lost down the rabbit hole checking out random search terms. You could type almost anything and get data back in seconds. You could lose hours if you aren’t focused on specific outcomes you are after.
If automating the collection of data, get appropriate compliance advice. Please note that Google will even throttle manual access if the page count is too high.
Spend the time to create well-crafted search terms, considering geography and language constraints of the tool. Always sanity check the output data with ground truth.
Extract more than 1 sample of the data for each query by appending random strings to each query and extracting each term multiple times.
Stich together multiple queries using a control term
Enhance the search index data to be aligned with the specific analysis goals and questions to be answered.
Regularly test the terms used in the analysis, looking out for changes in Google user behavior and changes in the target topic or industry in question in the analysis.
Increase the power of insights by combining the data with other sources, creating a funnel from top-of-mind awareness to loyal customer purchases or repeat human behavior
- Jason DeRise
Alternative Data: Alternative data refers to data that is not traditional or conventional in the context of the finance and investing industries. Traditional data often includes factors like share prices, a company's earnings, valuation ratios, and other widely available financial data. Alternative data can include anything from transaction data, social media data, web traffic data, web-mined data, satellite images, and more. This data is typically unstructured and requires more advanced data engineering and science skills to generate insights.
Logarithmic Scale: A logarithmic scale (or log scale) is a method used to display numerical data that spans a broad range of values, especially when there are significant differences between the magnitudes of the numbers involved. Unlike a linear scale where each unit of distance corresponds to the same increment, on a logarithmic scale each unit of length is a multiple of some base value raised to a power and corresponds to the multiplication of the previous value in the scale by the base value. A logarithmic scale is nonlinear, and as such, numbers with equal distance between them, such as 1, 2, 3, 4, 5 are not equally spaced. Equally spaced values on a logarithmic scale have exponents that increment uniformly. Examples of equally spaced values are [base = 10] 10, 100, 1000, 10000, and 100000 (i.e., 10^1, 10^2, 10^3, 10^4, 10^5) and [base = 2] 2, 4, 8, 16, and 32 (i.e., 2^1, 2^2, 2^3, 2^4, 2^5). Source: https://en.wikipedia.org/wiki/Logarithmic_scale
Clickstream data: Clickstream, or web traffic data, refers to the record of the web pages a user visits and the actions they take while navigating a website. Clickstream data can provide insights into user behavior, preferences, and interactions on a website or app.
Point in time-stamped history: This phrase refers to a dataset that provides the time data to show how data has been revised. So it includes not only the time period the data was related to but also the date when the entire data set was originally released or revised. This allows investors to use the data in back-testing models as if it were seen in real time before revisions.