Google Correlate

» see correlate.googlelabs.com as it lived, on the wayback machine

~ the last words ~

google trends asks what a word did. correlate asked which words did what you did, and that is a stranger question with a very much smaller audience.

the mechanics were simple to describe and unusual to use. you handed it a time series, a spreadsheet of numbers by week or by month, and it searched its own query volume for terms whose curve moved in the same shape as yours. unemployment claims or car sales or whatever you happened to have. it came back with a ranked list of searches and a correlation coefficient for each. there was also a mode where you drew the curve freehand with the mouse instead of uploading one. and it could work across american states rather than across time, matching a geographic pattern in your data against a geographic pattern in searching.

it was released on 25 may 2011 out of google labs, which google closed later that same year, and it came directly out of the flu trends work. in 2008 google had noticed that searching for flu symptoms tracked actual flu cases, and flu trends was built on that observation running in one direction. correlate was the generalization: rather than google picking a phenomenon and finding the queries, anyone could bring their own phenomenon.

i am a computer engineer and i have spent years on problems that involve finding structure in a signal, and correlate is still not a tool i ever had a use for. that is the honest distance here. it answered a question that belongs to economists and epidemiologists, people who have a national statistic in one hand and want a faster proxy for it. my work has cameras on one end of it. the deep data was american too, so even the shape of what it could answer was drawn around one country's statistical agencies.

~ the record ~

~ how it worked ~

the thing that made correlate feel slightly illicit was that it inverted the usual direction of search-data privacy. trends gives you aggregate history for a term you already suspect. correlate gave you terms you had not thought of, ranked by how well they tracked something in the physical world. hand it a series of housing starts and it would hand back queries about mortgage calculators, and hand it something with a strong seasonal shape and it would hand back everything else that peaks in december, which is where the failure mode lived.

correlation on seasonal data is close to worthless, because nearly everything with an annual cycle correlates with nearly everything else with an annual cycle. the tool told you that, in the sense that the coefficients were right there. it did not stop you believing it. flu trends itself overshot the 2012-13 flu season by roughly double, and time ran a piece in 2014 on what that said about big data generally. correlate inherited the same weakness and never pretended otherwise.

~ who actually used it ~

economists, mostly. hal varian, google's own chief economist, and seth stephens-davidowitz both wrote up how to get real work out of google's data series, and a hands-on primer of theirs from 2015 walks through the tool directly. papers came out of it on predicting unemployment and on correlating pedestrian counts with search behavior. it earned a small permanent place in the literature.

what it never earned was a place in the marketing stack. trends became a thing journalists cite and agencies screenshot. correlate required you to arrive holding your own dataset, which most people do not have, and to understand what a spurious correlation looks like, which most people do not. that is a tool with a real user base of a few thousand researchers worldwide, and a few thousand researchers is not a number google keeps a product alive for.

~ what got buried with it ~

google announced the shutdown for 15 december 2019 and cited low usage, and there was a brief round of complaint from seo forums who had been using it as a keyword discovery hack rather than as a statistics tool. no replacement was offered. trends is still there and does not do this. the wikipedia article for correlate was folded into a list of google products, which is the fate of a service too small to argue about.

correlate.googlelabs.com answers nothing now. the interesting residue is the datasets: correlate would let you download the matched series as a csv, and some of those csvs are still sitting in the supplementary materials of papers from that decade, frozen search-volume curves for queries nobody can regenerate any more.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »