Great article about what weight to put into studies, and the old correlation != causation. Even when something is closely correlated, causation is not proved. I often think of correlation falling into 3 groups: 1/ https://twitter.com/CountXero/status/1283404428578414592
1. Random coincidence, the type of thing you find here: https://www.tylervigen.com/spurious-correlations
They align, but that's just because they happen too. Fun stuff, but increasing mozzarella consumption isn't going to increase the amount of qualified civil engineers. 2/
They align, but that's just because they happen too. Fun stuff, but increasing mozzarella consumption isn't going to increase the amount of qualified civil engineers. 2/
2. 'Reverse' correlations, where the trend is connected to the thing you are studying, but that's because it's affected by it, it doesn't control the thing you are studying. I used to work in aerospace labs, & a very sensible boss taught me a lesson that stuck with me here: 3/
Don't heat the aeroplane. We used to do stuff for fast jets, & one of the things that happens when something is sodding off through the atmosphere at plus 1 mach is they get hot from the friction. There was a clear chart of leading edge temps & speed with near +1 correlation. 4/
On that data alone, if you confused the symptom for the cause, to make a fast aeroplane, you just need to heat it up. The hotter the faster. (That's not a good way to make a fast plane faster by the way). These kind of correlations can be useful for working out things you need 5/
To do, In the plane example make sure the plane can stand the heat. In a web example, making sure the site loads just quickly with 1,000 users on at the same time as opposed to just 100 at the same time becomes important IF you started ranking in 1 for a heavily searched term 6/
3. Actual Causation!
That's what we want, but we can only determine this is what's happening if we change the variable highlighted & that has repeatable results. Some are easy, like does adding a noindex tag affect ranking position? Yes, it's an easily measurable thing with a 7/
That's what we want, but we can only determine this is what's happening if we change the variable highlighted & that has repeatable results. Some are easy, like does adding a noindex tag affect ranking position? Yes, it's an easily measurable thing with a 7/
strong binary outcome. To study it is simple. Ask a different question, "Do clicks from different locations affect google's understanding of the relevance of a page to that term / city?" & it's a whole lot muddier. You might never determine this a 3 not matter how big your 8/
study as there's way to many unknowable variables, as we don't have access the the same picture google has, we don't know if there's something else doing this and we're actually seeing a 1 or 2. 9/
Correlation studies are fun to read, a lot of work goes into some of them. But they are the starting point at best to the hard work of determining if the factors they highlight fall into 1, 2 or 3. Given the ever changing & opaque nature of the algos, by the time you're done 10/
It might well have changed. That's why I tend to lean towards working with the know, verified factors that the search engines tell us about, and then just work to make the site match the goals they are seeking. Sounds boring but gives better ROI IMHO. /fin.