Twitter is all about “live,” but its platform can’t all be live because recent doesn’t always equal relevant. Twitter proved this in February when it introduced a digest-style mini-feed in its main timeline that shows a sample of tweets someone might find most relevant even if they’re not the most recent. The company has quietly done a similar thing with its search results pages.
In September Twitter started sorting the top-most tweets on its default search results pages based on how relevant they are to the search query, rather than ordering them in reverse-chronological order as it had previously done. The relevance-based order only applies to the first handful or so of tweets atop the search results with the rest sorted in reverse-chronological order. If people want to simply see the most recent tweets related to their search query, they can still find the reverse-chronological firehose under the “latest” tab on the search results page.
Even though the newish top tweets section favors relevance over recency, recency remains a strong signal, according to a Twitter spokesperson.
For example, yesterday I searched “damn daniel video” on Twitter to see if the top result would be the most relevant one regardless of recency: the original clip that caught fire on Twitter back in February. It wasn’t. Instead the results were a smattering of tweets from this month, with the top 5 being out of chronological order and sorted based on how relevant Twitter thought they were to my query and without the original video being among them.
But the results were an improvement from May when I tried the same search query and the top result was a much-liked, days-old tweet that simply included the words “damn daniel” and “video” and the runner-up result linked to a “Damn Daniel Videos” parody profile that had never posted a tweet. At least yesterday’s top results actually included videos.
The relevance-based top search results section resulted from a series of tweaks Twitter has made to its search-ranking system in hopes of doing a better job at identifying the tweets people are looking for. According to Twitter, these changes have led to people interacting more with its search results and, more importantly for a social network that has been struggling with audience growth and turnover, sending out more tweets and spending more time on Twitter.
In a blog post detailing its search ranking adjustments, Twitter’s senior software engineer for search quality, Lisa Huang, outlined three things Twitter had to figure out before swapping in the relevance-ordered results: 1) the right mix between tweets and profile links in search results, 2) whether it’s better to fit one super-relevant tweet on a screen or a few that aren’t as individually relevant and 3) how to put tweets on equal footing when ranking them by engagement.
Tweets or profiles?
When Twitter had tried out displaying tweet results in a horizontal carousel at the top of the page and place a similar carousel showing relevant Twitter profiles somewhere below it, people were less likely to check out the profile results. Then Twitter switched things up. It splintered the tweets carousel into individual results for each tweet, and people were more likely to check out the tweet results without as much of a negative impact on the profile results. Twitter also worked out a way to algorithmically determine how high the profiles carousel should appear in the list of results based on whether the search query was likely meant to find profiles or tweets.
More less relevant tweets or less more relevant tweets?
Twitter’s also looking at how it can squeeze a wider variety of tweets into search results, so that flashier tweets don’t obscure the rest.
People seem to really respond to tweets that include photos or videos. They interact with them more and those tweets rank really high, according to Twitter. But the photos and videos make those tweets take up a lot of space on the screen. They crowd out other tweets that a person might be interested in, if not more interested in, so that a person would have to swipe or scroll down to see them. “Assuming that people have some subconscious limit to their scroll span, we end up showing them fewer results,” wrote Huang.
So Twitter is working on ways to spice up its search results, such as by now considering how much screen space a tweet would take up in order to gauge the total relevance score of all the tweets displayed on the screen. Twitter refers to this as the “relevance density.”
For example, let’s say a tweet with a large, vertical video takes up an entire search results screen and has a relevance score of 0.7. Then there’s a set of three tweets — two text-only ones and one with a smaller, horizontal video — that together take up a screen, have lower individual relevance scores but combine for an aggregate relevance score of 0.9. In the past Twitter would have prioritized the vertical video tweet because its relevance score topped each of the other three tweets on a tweet-versus-tweet basis. But now Twitter is teaming up those other three tweets, like the Power Rangers combining into the Megazord. As a result their aggregate relevance score would beat the vertical video tweet’s relevance score, and the three tweets would show up higher in the search results.
But Twitter still has to work to do when mixing up its search results. While a mix of more tweets with different or no attachments diversifies the results, Twitter also needs to make sure those tweets aren’t redundant. If you type a search query to see the latest touchdown pass in a live NFL game, you probably don’t want all the top results to be some variation of “Touchdown!” or all to be videos of the same play. You’d probably be more interested if one was a video of the touchdown, another had a photo of the catch and another mentioned the score of the game.
Making the primary signal an unbiased signal
The main signal Twitter’s search algorithm uses when ranking a tweet’s relevance is how likely the person behind the search query will engage with the tweet, according to the Twitter spokesperson. But estimating how likely someone is to engage with a tweet isn’t as straightforward as referencing the number of times other searchers have engaged with the tweet.
If a tweet is placed higher in the search results, then of course it’s going to get more engagements, if only because people are more likely to see it. So Twitter devised a way to limit the effect of what it calls “positional bias” by looking at the historical data on a tweet’s position in search results and the number of engagements it received. In short, the higher a tweet’s position in the search results, the lower the weight of the corresponding engagements.
But even after accounting for positional bias, Twitter is still fine-tuning the signal-to-noise ratio. It’s been experimenting with ways to filter out noise like people accidentally hitting the like button while scrolling through results or instances when some people engaging with the tweet and some not. The latter noise isn’t actually noise but a signal that a tweet may resonate with certain audience segments but not others. And Twitter is trying to home in on that signal so that it can better personalize its search results.