At the top of his rally in the critical Wisconsin county of Waukesha on Wednesday night, Donald Trump accused the search giant Google of both impeding and bolstering his candidacy.
"A new post-debate poll, the Google poll, has us leading Hillary Clinton by two points nationwide," he said, "and that's despite the fact that Google search engine was suppressing the bad news about Hillary Clinton. How about that."
I pay attention to polls, and I wasn't exactly sure what he meant by "the Google poll." Google doesn't and hasn't polled over the course of the campaign, as far as I was aware. Nor was I familiar with anything recent alleging that Google was "suppressing" anything.
As it turns out, this is classic Trump: Running full steam ahead with any sketchy evidence that seems like it might be helpful to him. Here's each thing he was referring to, so you can judge for yourself.
The 'Google poll'
The "Google poll" appears to be a cumbersome reference to a Google Consumer Survey conducted by Independent Journal Review. The survey did show Trump with a 1.7-point lead. It also showed Clinton winning the debate by a 52-to-48 margin. (Independent Journal Review is a sort of right-leaning Buzzfeed.)
The Google Consumer Survey process uses one of two methods to collect data, either through a phone app or by asking questions of people visiting a website. Have you ever been presented with a survey before you could read a news story? That was probably a Google Consumer Survey. They're built to be inexpensive, easy to set up and quick to deliver results.
It's less clear how accurate they are. In 2012, Nate Silver ran an experiment to test various polling methodologies, and found that Google was fairly close to the final result in the presidential race. Other studies have found that the surveys might be a useful tool for fleshing out social science research.
Because of the methodology used in the survey, though, The Washington Post doesn't consider Google Consumer Surveys as a sound way of drawing inferences about the national population. It's a survey, not a scientific poll. Nor does Google really pitch the service as a poll substitute, incidentally. Perhaps for marketing reasons, they emphasize use of the surveys as a marketing tool more than a polling tool.
There isn't much data provided from IJR's survey beyond the main numbers, also known as the top-lines. We don't know who was included in the poll; we don't know how the participants were screened to ensure they'd watched the debate. If the poll was a pre-news-story survey that demanded responses before people could read an article, that may have prompted more random clicks than a more traditional poll. It's hard to tell.
Google argues that its results are more accurate than other Internet-based polling. One such pollster is SurveyMonkey, which has done regular polling for NBC News. It found that Clinton won the debate 52-to-21.
The Google 'suppression'
In June, a conspiracy theory spread around the Internet that Google was tamping down search results critical of Hillary Clinton. The idea was that searching for things like "hillary clinton cri" didn't automatically suggest "hillary clinton criminal" as an option — though "donald trump cri" would. That sort of thing.
The report, a video from a site called SourceFed, was quickly rebutted by Google and debunked by other sites and people. Google said that its algorithms did filter out disparaging phrases after finding that it "too often predicted offensive, hurtful or inappropriate queries about people."
This particular conspiracy theory re-emerged this month thanks to a lengthy exploration of the idea published at Sputnik News. The author, Robert Epstein, is a research psychologist who was once the editor in chief of Psychology Today and who has written repeatedly about what he calls the "Search Engine Manipulation Effect." In this report, he runs through a number of searches showing examples of negative Clinton suggestions being hidden and negative suggestions about Trump being shown. He finds some occasions in which negative Clinton suggestions get through, which he offers may be allowed through to hide the fact that others are suppressed. He also notes that "if you try to replicate the searches I will show you, you will likely get different results."
Google's algorithm powering its autofill is sensitive to a lot of context, including the user and the time period and the emergence of new concepts. Epstein argues that there's also a layer of politically motivated filtering. As Google noted in June, it seems a little odd to hide autocomplete results for "hillary clinton criminal" from autofill but then show results for "hillary clinton criminal." There's also some significant selection bias at play. Epstein thoroughly explores negative stories about Clinton, but doesn't tackle Trump to the same degree. One blogger at Medium in June noted that selecting some anti-Trump terms allowed a similar re-creation of results.
It's also worth noting in the current environment that Epstein published his study at Sputnik News because, he says, the site "agreed to publish it in unedited form in order to preserve the article's accuracy." Sputnik was founded and is supported by the Russian government, suggesting that they probably weren't terribly eager to edit Epstein's work anyway. How'd Trump see the story? Well, it was picked up by Breitbart.
Stepping back from politics, it's not clear what Google's motive for suppressing autofill responses would be. If you go to Google and search for "hillary clinton cri," what are you expecting to see? Epstein's broader campaign is about the negative effects of search engines, which certainly could lead to some confirmation bias. What good does it do Google? Why damage the company's reputation solely to make it sort of harder to find negative things about Hillary Clinton? It doesn't really make sense.
Unless you're looking to seize upon anything that portrays your opponent in a negative light, no matter how iffy. Or if you're looking for any good news you can find about yourself, despite any questions that might surround those numbers. So Google is good and bad in one sentence for Donald Trump, because Google, on this day, happened to offer a bit from each column.