Introduction to Google Trends (Part Three)
July 25th, 2008 at 11:03 am | by Michael Giuffrida |This is my final introductory post for Google Trends. Previous posts dealt with the scaling and normalizing of data (Part One) and comparing terms across regions (Part Two).
Today I’m just going to talk about exporting data. This is a new feature in Google Trends and is only available if you sign in with your free Google account.
To export data, expand the “Export this data as a CSV file” node at the bottom of the page, and you’ll see two links: one for fixed scaling and one for relative scaling. (CSV files are lists of “comma-separated values” that load easily into spreadsheet applications like Microsoft Excel.) We’ll start with relative scaling, since that’s what we’re used to.
First, you get a list of search terms with their average Search Volume Indices. Remember, these are all scaled such that the first term has an SVI of 1.00. Here’s what we see for our facebook, myspace comparison search across all years and regions:
facebook facebook (std error) myspace myspace (std error)
1.00 0% 2.25 1%
Then we get the list of data points, the same ones you see in the actual graph. Here’s a sample of some consecutive data points:
Week facebook facebook (std error) myspace myspace (std error)
Apr 6 2008 4.15 2% 4.05 2%
Apr 13 2008 4.1 2% 3.95 2%
Apr 20 2008 4.45 2% 4.05 2%
Apr 27 2008 4.65 2% 4 2%
May 4 2008 4.7 2% 3.9 2%
May 11 2008 4.9 2% 3.85 2%
Yep, you only get one data point per week, starting on Sundays. But I’ll explain how to get around this later. ;-)
There’s really nothing different here. The data points are scaled, relative to facebook having an average SVI of 1.00. Standard error is an estimate of the standard deviation of the error — Google only uses a small sample of all of its searches to calculate the Trends numbers, so they estimate a 2% standard deviation of error, which is small enough for our purposes. If you download the CSV yourself, you’ll notice that back in 2004, the standard error for facebook was listed as “>10%”. This went down to 10% in the fourth quarter of 2005, then down to 5% at the end of 2006 and down to 2% in April of 2007 as more and more people searched for facebook and the size of the sample increased. The standard error of myspace went down to 5% much earlier, in May 2005, which makes sense because it had more search volume at that time.
We can easily take these numbers and make a line graph that looks very similar to Google’s:
Google Trends Exported
The spreadsheet also comes with the Region, City and Language data, which works the same way as the data shown above.
Now, like I said, this is the same old relative scaling we’ve been dealing with all along. What’s fun is that, only when exporting data, Google Trends offers fixed scaling. Whereas relative scaling scales the data to the entire time frame selected — if you look at data from 2007, it’ll be scaled such that the average Search Volume Index across 2007 equals 1.00; if you look at data from October, 2007 it’ll be scaled such that the average SVI across October, 2007 equals 1.00 — fixed scaling scales all data so that the average SVI equals 1.00 at some early, fixed point in time, for every time frame. According to Google, the fixed point in time is “usually January 2004.” And this only works for keywords with enough historical data to even have data in that fixed time frame.
This is useful because it allows us to compare a keyword across different time frames. For instance, say we want to see how facebook was doing in January 2007 and compare that with its search traffic in December 2007. If we just export the relative data, both sets of data will be scaled so that the average SVI for their respective time frames equals 1.00 — in other words, your data sets will both average to the same number. That’s not very useful for comparing across time. Instead, if you export data with a fixed scale, the January 2007 data and the December 2007 data will both have an average SVI scaled by setting the data from January 2004 to 1.00. Thus, we’ll be making a valid comparison.
Of course, you could just export the entire time frame with a relative scale, and comparing January 2007 and December 2007 would be very easy. So why bother exporting the two months separately with a fixed scale?
Well, remember how I told you that the exported data only gave us one data point a week, every Sunday, but that I’d show you a way to get around that? This is it. Exporting data by month gives you one data point per day. So by exporting fixed-scaled data month by month, we could eventually compile a list of search volume data for every single day from January 2004, or whenever our data starts, to the present.
And there you have it. Google Trends now allows you to export search volume data, down to daily points, without having to estimate from the pixel height of the lines on the graph you get online. You’re free to apply any statistical methods you do (or don’t) want to use. This means we can have actual science behind our trend analyses, making our conclusions much more credible.
This was Part Three of my three-part Introduction to Google Trends. Part One. Part Two.
Tags: google, google trends, making comparisons, scaling, search volume, statistics, time, tool