100% of statisticians would say this is a terrible method for predicting elections. However, in the case of 2016’s presidential election, analyzing the geographic search volume of a few telling keywords “predicted” the outcome more accurately than Nate Silver himself.
The 2016 US Presidential Election was a nail-biter, and many of us followed along with the famed statistician’s predictions in real time on FiveThirtyEight.com. Silver’s predictions, though more accurate than many, were still disrupted by the election results.
In an effort to better understand our country (and current political chaos), I dove into keyword research state-by-state searching for insights. Keywords can be powerful indicators of intent, thought, and behavior. What keyword searches might indicate a personal political opinion? Might there be a common denominator search among people with the same political beliefs?
It’s generally agreed that Fox News leans to the right and CNN leans to the left. And if we’ve learned anything this past year, it’s that the news you consume can have a strong impact on what you believe, in addition to the confirmation bias already present in seeking out particular sources of information.
My crazy idea: What if Republican states showed more “fox news” searches than “cnn”? What if those searches revealed a bias and an intent that exit polling seemed to obscure?
The limitations to this research were pretty obvious. Watching Fox News or CNN doesn’t necessarily correlate with voter behavior, but could it be a better indicator than the polls? My research says yes. I researched other media outlets as well, but the top two ideologically opposed news sources — in any of the 50 states — were consistently Fox News and CNN.
Using Google Keyword Planner (connected to a high-paying Adwords account to view the most accurate/non-bucketed data), I evaluated each state's search volume for “fox news” and “cnn.”
Eight states showed the exact same search volumes for both. Excluding those from my initial test, my results accurately predicted 42/42 of the 2016 presidential state outcomes including North Carolina and Wisconsin (which Silver mis-predicted). Interestingly, "cnn" even mirrored Hillary Clinton, similarly winning the popular vote (25,633,333 vs. 23,675,000 average monthly search volume for the United States).
In contrast, Nate Silver accurately predicted 45/50 states using a statistical methodology based on polling results.
This gets even more interesting:
The eight states showing the same average monthly search volume for both “cnn” and “fox news” are Arizona, Florida, Michigan, Nevada, New Mexico, Ohio, Pennsylvania, and Texas.
However, I was able to dive deeper via GrepWords API (a keyword research tool that actually powers Keyword Explorer's data), to discover that Arizona, Nevada, New Mexico, Pennsylvania, and Ohio each have slightly different “cnn” vs “fox news” search averages over the previous 12-month period. Those new search volume averages are:
“fox news” avg monthly search volume |
“cnn” avg monthly search volume |
KWR Prediction |
2016 Vote |
|
---|---|---|---|---|
Arizona |
566333 |
518583 |
Trump |
Trump |
Nevada |
213833 |
214583 |
Hillary |
Hillary |
New Mexico |
138833 |
142916 |
Hillary |
Hillary |
Ohio |
845833 |
781083 |
Trump |
Trump |
Pennsylvania |
1030500 |
1063583 |
Hillary |
Trump |
Four out of five isn’t bad! This brought my new prediction up to 46/47.
Silver and I each got Pennsylvania wrong. The GrepWords API shows the average monthly search volume for “cnn” was ~33,083 searches higher than “fox news” (to put that in perspective, that’s ~0.26% of the state’s population). This tight-knit keyword research theory is perfectly reflected in Trump’s 48.2% win against Clinton’s 47.5%.
Nate Silver and I have very different day jobs, and he wouldn’t make many of these hasty generalizations. Any prediction method can be right a couple times. However, it got me thinking about the power of keyword research: how it can reveal searcher intent, predict behavior, and sometimes even defy the logic of things like statistics.
It’s also easy to predict the past. What happens when we apply this model to today's Senate race?
Can we apply this theory to Alabama’s special election in the US Senate?
After completing the above research on a whim, I realized that we’re on the cusp of yet another hotly contested, extremely close election: the upcoming Alabama senate race, between controversy-laden Republican Roy Moore and Democratic challenger Doug Jones, fighting for a Senate seat that hasn’t been held by a Democrat since 1992.
I researched each Alabama county — 67 in total — for good measure. There are obviously a ton of variables at play. However, 52 out of the 67 counties (77.6%) 2016 presidential county votes are correctly “predicted” by my theory.
Even when giving the Democratic nominee more weight to the very low search volume counties (19 counties showed a search volume difference of less than 500), my numbers lean pretty far to the right (48/67 Republican counties):
It should be noted that my theory incorrectly guessed two of the five largest Alabama counties, Montgomery and Jefferson, which both voted Democrat in 2016.
Greene and Macon Counties should both vote Democrat; their very slight “cnn” over “fox news” search volume is confirmed by their previous presidential election results.
I realize state elections are not won by county, they’re won by popular vote, and the state of Alabama searches for “fox news” 204,000 more times a month than “cnn” (to put that in perspective, that’s around ~4.27% of Alabama’s population).
All things aside and regardless of outcome, this was an interesting exploration into how keyword research can offer us a glimpse into popular opinion, future behavior, and search intent. What do you think? Any other predictions we could make to test this theory? What other keywords or factors would you look at? Let us know in the comments.
Also, if you've enjoyed this post, check out Sam Wang's Google-Wide Association Studies! --Fascinating read.
This is super cool! Definitely shows the potential for more uses of keyword research for broader applications. And makes you wonder why Google can't use the reams of other data they don't publicly share to make even more accurate predictions (or can they?)
I think another interesting thing to look at would be to hone in depper on the counties/areas that flipped allegiances and voted for Trump, and find more specific predictors there. Or another way, see if in those counties shifted in searching for "cnn" vs "fox news" between 2012 and 2016 to match voting trends.
Interesting post Britney!
never thought of looking at the outcome/ prediction of an election like that.
Nevertheless reading your article I think it is not too surprising to find out that depending on peoples search results you can predict the outcome of the election. People search for things that they are interested in and if it is, let it be the two media sources you just mentioned, I think it is not too big of a surprise, especially if both cnn and fox are positioned so differently. People consume the media, believe the media - so the media they choose to listen to /follow through for sure has a huge impact on their political views.
I have to agree with you Joe. It would be really interesting to see how things shifted over the years and if we could see a trend in there! If there is a trend, I am sure that this might become a much bigger deal to predict elections in the future.
Thanks so much Abel! It was really fun to dive into.
Have you read Everybody Lies by Seth Stephens-Davidowitz (former data scientist at Google)? It looks at internet data (a lot of it Google Searches) and how search data and search trends say differently to what the polls say.
One of the really interesting parts it looks at is how searches change based on what is said in a presidential speech and how certain words can heavily influence Google searches.
Tom! Someone else just messaged me about Everybody Lies yesterday!!! I have to read that! -Thank you
Yes, great point Joe! We could also look at how things might (hopefully) start improving culturally and see those searches shift. Lots of interesting use cases here.
Wow! That was amazing. Could be interesting to see this approach with the Catalonia process in Spain and the upcoming elections.
Mario, that would be fascinating to find a similar keyword indicator!
Hello Mario! I'm also from Spain... let's definitely check this !!
Interesting post!! I'm from Spain, where the territorial elections in Catalonia will be held in 5 days!! Let's see if all this applies :)
Woaahhh This is mind blowing!
I love the scientific approach of this idea!
Fun stuff to read thanks!
That's very interesting. Estudying keyword research you can predict the next president, next fashion tendences or the business growth of any company.
In the politician subject I think in Spain (where I live) a lot of people changes his vote the last hours before vote and avoid a best prediction in the elections. Some companies that use this info normally fail lot in their predictions.
I'm introducing the Hungarian example. https://gerillamarketing.blog.hu/2018/01/07/a_duna_...
Hi Britney
They are different cultures, but here in Spain we often seek political information to do harm. We look for more information to not vote for someone than to vote for a candidate
Good point Lluis!!
Whaaaa?! I'm so confused. Why would you seek out information to not vote for someone?
Is there a resource or article you could refer me to? Sounds fascinating. Thanks for letting me know Lluis!
Great addition Lluis! I also experienced that with some of my family members who live in Europe! (but quite frankly I am not sure if this is just a phenomenon in Europe/or Spain).
If there are two candidates and you are not really convinced of either one, you will vote for the person you think might be less "harmful". So yes, you will look up all the information needed to decide, why you should not vote for them. But then again in this case, people would look up both candidates/ on different media channels. Nevertheless I think that depending on the media people listen to on a daily basis, they might be already influenced, depending on how the media channel is positioned.
Sure! I agree with you!
Very good post In Spain it happens that what the polls predict is not always what is really voted, but it is always very approximate. We will see what happens tomorrow in the Catalan autonomous elections. Long live Spain!
Very powerful
There's a lot of power in political statistics! I think this is an interesting study, but perhaps political affiliation through organizations could help in county level political keyword research like this ("membership fees for nra" vs. "aclu donation" type of keywords). Issue how to and why keywords ("what does national reciprocity mean" or "what is unemployment percentage") would also assist in this as long as they have generally the same volume. Campaigns look a lot at issues, legislation, and their opponent's positions for adword campaigns and the like, so it's probably a good indicator!
Gotta be careful about controls too - demographics are big in influencing political belief!
This is awesome! Keyword research tools clearly predicted the election far better than any poll during the campaign. What a great source to predict the next one. I wonder if this same theory applied to yesterdays Alabama election with Ray Moore and Doug Jones?
This is mindbendingly interesting. It’s rare to find an analysis that is so completely new.
Thank you so much Memli! We were a bit hesitent to publish something so different, but the feedback has been great.
Some great insights her; particularly the power behind GrepWords API.
Thats amazing.. makes me wonder, when Rand did a whiteboard friday he said that google adwords keyword planner tends to drop some keywords and keywords that are more of "to buy" driven. also would you say that keyword explorer by moz could have produce the same results?
The power of Internet searches is enormous. Good article and surprising.
Thank you.
Britney your posts always blow my mind great job on this. I was thinking that this would be interesting on some speculative investing topics you read for example Bitcoin or even some AI news that recently came out sometimes it is articles or main keywords people are searching for, and from positive or negative terms can affect how that company ect surges in the market.
Thanks for this analytical information of keyword madness