In September 2015, I stood on a big stage for Google in Berlin and showed the advantages of the new features of Google Trends in addition to voice search. It is a useful tool, but also offers a lot of potential for misunderstandings, which should be cleared up here. Search queries are enclosed in <> parentheses.
- 1. Misconception: Not all search queries are taken into account
- 2. Misconception: There are no absolute numbers
- 3. Misconception: A search query is different on Google Trends
- 4. Misconception: A rising line does not mean that searches were made more often
- 5. Misconception: Without a benchmark, Google Trends is worthless
- Bonus Misconception and Effects of Misinterpretations
- How can Google Trends be used sensibly?
- Summary
This text was updated on May 27, 2020, because Google had revised the help for Google Trends.
1. Misconception: The basis of Google Trends data
Google Trends is not based on all search queries entered on Google, but on a representative sample:
Google Trends data reflects the search queries that users make to Google every day. However, they may also include irregular search activity, such as automated searches with the aim of distorting our results. […] Although we have mechanisms in place to detect and filter such activity, these searches are sometimes stored in Google Trends for security reasons: If we were to filter them out in principle, the creators of such queries would know that we are tracking them. This in turn would make it more difficult to filter out such activities from other Google search products, where the most accurate representation of the real data is crucial. For this reason, users of Google Trends data should understand that it is not an accurate reflection of search activity.
Google Trends filters out certain types of search queries, for example […] Searches from a few people: Trends only analyzes data for popular search terms. Terms with low search volume therefore appear as 0
There is no definition of when a term is popular. And just by the way: Hasn’t anyone noticed that hormone-controlled terms never appear in the most popular search terms published in newspapers based on Google Trends? The first insight: We are talking about trends here, no more and no less.
2. Misconception: Lines and search volumes – There are no absolute numbers
This is the biggest and the most fatal misunderstanding. If one curve is above the other, this does not mean that one term was searched for more often than the other. The lines do not reflect absolute numbers. In this example, we are looking for acne and neurodermatitis (I explain below why I write this in quotation marks in the mask) for the country of Germany in 2015. Acne and neurodermatitis alternate in search interest, but acne seems to have a higher search interest more often:
And then I look at the data in Google AdWords Keyword Planner, for the same time period, for the same country:
What is interesting here is not the graph (I only took the screenshot so that you can see that I am searching in the same country for the same period), but the two lines below where we see the average search volume per month. Atopic dermatitis is far above acne, 74,000 to 18,100.
Averages can lead us in the wrong direction, so let’s also look at the data plotted for each month:
We see a similarity to the Google Trends graphic, namely that searches for neurodermatitis go down from May or middle of the year and up again from September. And, this will be important later, neurodermatitis gets around 100 on Google Trends, but acne never reaches this point. Otherwise, the curves of the AdWords data do not touch once as in Google Trends. They are far apart. So the second insight: We can’t claim from Google Trends data that one term is searched for more often than another (although Google Trends is often misused for this). Google Trends doesn’t offer absolute numbers. Sorry.
3. Misconception: What is a search query?
If you type in Google or in the Google AdWords Keyword Planner, you will only search for this term. If you enter trends on Google, it will automatically search for other terms, even if you have not selected a topic, but only this search term (see the [Help][4]). For example, you can restrict something by putting a term in quotation marks (“acne cream”), but this only restricts that it is not searched for, but it could be included. It is not said which search terms are included. An “exact match” does not exist, see again [the help][5].
Let’s take a look at the differences:
In this example, we compare the search terms and . If we add quotes, then the curves will look a little different:
Not a huge change, but there is a difference that we will remember again for later: If we enter the terms without quotation marks, then we get about 100, with quotation marks we get about 100.
It is surprising, because in the case of a one-word term, where no synonyms are searched, there can be no different order for the words in the search query. We cannot explain this phenomenon.
Finally, let’s take a look at what happens when we select the automatically identified topic “disease” (note: Google automatically makes out what the medical term for atopic dermatitis is):
Here, the trend data of terms that fit into the group of the disease are aggregated. The “search interest” in neurodermatitis comes close to the topic of acne in February 2016, but acne as a topic seems to have a greater search interest than neurodermatitis. Again, we don’t know which terms are grouped together. So it could be that the different data comes from the fact that Google Trends includes additional terms for both terms, but for the term acne terms are used for terms whose search interest is different and therefore changes the result. But that doesn’t sound very plausible. Third finding: Google Trends data is not comparable to AdWords data because the input is interpreted and enriched differently and we at Google Trends don’t know what it is with.
However, the differences between AdWords and trends are probably still not explained. What could be other reasons?
4. Misconception: Anything that rises or falls is a trend in Google Trends
Now it’s getting a bit mathematical. Google Trends does not offer absolute numbers, all data is displayed on a scale from 0 to 100. And now we remember the two clues above again, when one of the two search query curves touched the 100. Touching the 100 has a lot of meaning, because everything else is calculated from this highest point of search interest!
But it gets even more complicated: First of all, search interest is the search volume for a term divided by the search volume of all terms. Since we don’t know the basis (i.e. how many searches there were in total on that day) and this basis changes every day, it is possible that the line of search interest for a term changes, even though this term is searched for the same number of times every day. The point at which the maximum of this search term/all search terms ratio is reached becomes the maximum, i.e. 100 in the Google Trends chart, and all other values, including those of comparison terms, are derived from it in a normalized manner. And only for the selected period. If this is changed, the maximum is also recalculated. This leads to differences that could lead you to choose exactly the period that suits what you would like to sell Examples:
Misinterpretation: Compared to the weather, interest in Trump has hardly increased in the last 12 months
Misinterpretation: In the last 30 days, interest in Trump has not increased as much as that in Wetter. In fact, however, the data from Google Trends has not yet been updated, the election day and the previous day are missing.
The chart for the last 7 days: Can we now claim that Trump was searched for more often than the weather?
These last graphs are very nice because they plausibly show that the Germans did not necessarily search less for the weather, but the search queries for the term Trump on 9.11 had a significantly higher proportion in the population of all searches compared to the 6 days before. Everything else is calculated from this one maximum. Over 30 days, however, the weather had a maximum (EDIT: Because the day after the election was not yet possible and it is not even two days later! Thanks to Jean-Luc for this hint!), and then the calculations are made from there. That’s why this data can be so different.
Edit: If you look one day after the election and only look at the last day (i.e. data that is about 10 minutes old), then the result looks like this:
Again, the weather is higher, although Trump comes close. Or? Well, I took this screenshot in the evening, and the peak of Trump searches probably took place more than 24 hours ago. It is very likely that I would have gotten a different image if I had made the same query this morning.
What do we take away from this misunderstanding?
- The period of observation is immensely important, and you should not believe any trends chart without looking at several time periods
- All observations start from the maximum and can then be seen relatively from this maximum, but at the same time depend on the total volume of all search queries, which we do not know.
- It may not be so obvious in the graphs, but the 7-day graph is calculated on an hourly basis, the 30-day graph on a daily basis, the 12-month graph on a weekly basis. The Google AdWords Keyword Planner delivers data on a monthly basis. This is another reason why the data are not comparable.**
- Learning: “The last 30 days” does not necessarily mean that the last 30 days are really in it
5. Misconception: Without a benchmark, Google Trends is worthless aka Everything that rises or falls is a trend in Google Trends, Part 2
Depends on how you define trend. After the Brexit vote, a journalist found out through Google Trends that the British had only googled what Brexit meant after the vote, many newspapers wrote about it, and only [after a data analyst had looked at what was really happening][13] did everyone row back. Yes, there was more searching for it, based on… see the 4th misconception But compared to a popular search query, there was only a twitch. Data must always be put into context to get a sense of what that really means. Although I’m not a football fan, I like to use the search query, but also to see how relevant a term really is.
The weather is always sought, but a certain seasonality can also be observed here, Bundesliga is primarily sought by a part of the population, and only seasonally, but interesting observations can also be made here.
We can see the Friday game (first small dent in the blue curve), the Saturday games (biggest dent in the blue curve) and then again smaller dents for the Sunday games. The weather here has the highest swings, especially on Monday and Tuesday (maybe because of the snow?). So was the weather searched for more often? No, not necessarily! Since this is calculated on an hourly basis, everything is calculated from the maximum on 8.11., 5 and 6 o’clock, at this time the search interest (search query/all search queries) was highest, and everything else is calculated relatively. So it may theoretically be that on Saturday afternoon more searches for Bundesliga were carried out than for Wetter on 8.11., but the total population of search queries was higher! What do we take with us? We always need a reference point, a benchmark, something to compare search interest to. And best of all, we also know something about this point of reference, such as in this case that it snowed and there were outliers because of it.
6. Bonus Misconception and Effects of Misinterpretations
A big misconception, apart from how Google Trends works, is the assumption that users really only search for EVERYTHING on Google. For example, Google has lost some search queries to Amazon, and if you look at your own search behavior, you no longer go to Google for every search, but directly to where you know that you will immediately find what you are looking for. Be it the bookstore around the corner with an online ordering service, be it the pizza service or clothing store. In fact, some trends are not even reflected in Google tools because they take place on Pinterest and other platforms. So Google search queries cannot be used to infer the needs and wishes of all people.
What bad things can happen if you misinterpret the data? My experience shows that in the worst case, it is better to take wrong data than to have none at all. As a result, budget decisions may be made incorrectly. But it can get worse. If, for example, scientific articles use Google Trends, here for analyses in the financial market or here in health research, without understanding how the tool works. If scientists don’t even understand this or don’t take the time to check how it works, how can you ask the average user to do that? The journalism example has already been described in the previous section, and here too one would expect journalists to do a little more research. Mark Twain is said to have once said:
Never let the truth get in the way of a good story.
7. How can Google Trends be used sensibly?
If you need absolute numbers, you can only use the Google Ads Keyword Planner. Point. Alternatives such as keywordtool.io or tools with an integrated search index such as Sistrix only provide partially correct data.
So what does Google Trends have that Google Ads Keyword Planner doesn’t? Google Trends is incredibly up-to-date, data is in within a few minutes. You can understand the search interest of the past on an hourly basis. And for more than 10 years, unlike Keyword Planner, where it’s only the last 4 years.
Google Trends can therefore be used very well in combination with the Keyword Planner, for example, to understand when are the best times to control campaigns more.
Summary
The mechanism “I enter two terms in Google Trends and see which one is more popular” simply does not work that easily. This will be difficult to get out of people’s heads, because the interface is wonderfully intuitive and literally tempts you to this interpretation, and it is not completely wrong. It is difficult when decisions with serious consequences are to be made on the basis of this data, and additional data is necessary here. Keyword Planner data is no longer available to everyone, but it is not directly comparable anyway, because Google Trends data does not compare pure terms.
And yet Google Trends is more than just a gimmick. With all the above, you just have to invest a little more brainpower to build a real story out of it.
Comments (since February 2020 the comment function has been removed from my blog):
Jean-Luc Winkler says
- November 2016 at 18:18 Hey Tom, very exciting post! In paragraph 4, you look at the data “last 30 days” and “last 7 days”. The data of the display “last 7 days” includes 7×24 hours from today’s last full hour retroactively (e.g.: on 09.11. at 17:17 is the last data record from 09.11. at 17; 00h). This period includes 09.11., when the curve shows the strong maximum. The data of the presentation “last 30 days” include 29 daily rates, starting from the day before yesterday retroactively (e.g. considered on 09.11., the last data point is 06.11.). So the day on which the curve shows the maximum on 11/9 is only shown in the “last 7 days” plot, not in the “last 30 days” plot, right? So let’s wait until the day after tomorrow and see if there is a deflection in the curve in relation to the curve that exceeds the curve. Best regards, Jean-Luc
Tom says
- November 2016 at 20:44 Very good point, I’ll add that!
Olaf Kopp says
- November 2016 at 11:23 Hi Tom, thank you for this very enlightening post. I’ve come across the differences between Keyword Planner and Google Trends several times now. I find it very annoying that Google is very opaque here and, as with other things, not consistent when it comes to data bases and more…
Tom says
- November 2016 at 11:33 Hallo Olaf,
in fact, (almost) everything is explained in the help, but hardly anyone bothers to read it because it looks so easy
LG
Tom
A>utonomes Trading says 18. March 2017 at 22:13 enter a fixed start date and a variable end date in the trend and download it in the CSV and compare it. The data changes back and forth “percentage-wise” as the mood takes you. And if, for example, you only download 14 days, the sum from Sunday to Saturday, you should get the same percentage difference as with the weekly data. So the time of the data query is also decisive for the same data set. (historcial random statistic dates)
S>argon says 3. April 2018 at 18:51 Thank you! I was just amazed how search volume suddenly fell far below 100, after I had already almost finished for the keyword page… Now I know where all the people have gone!
4]: https://support.google.com/trends/answer/4359550?hl=de&ref_topic=4365530 [5]: https://support.google.com/trends/answer/4359582?hl=de&ref_topic=4365530
[13: https://medium.com/@dannypage/stop-using-google-trends-a5014dd32588#.fvs83xks6