People are asking us if we shadow ban. We do not. But let’s start with, “what is shadow banning?”
The best definition we found is this: deliberately making someone’s content undiscoverable to everyone except the person who posted it, unbeknownst to the original poster.
We do not shadow ban. You are always able to see the tweets from accounts you follow (although you may have to do more work to find them, like go directly to their profile). And we certainly don’t shadow ban based on political viewpoints or ideology.
We do rank tweets and search results. We do this because Twitter is most useful when it’s immediately relevant. These ranking models take many signals into consideration to best organize tweets for timely relevance. We must also address bad-faith actors who intend to manipulate or detract from healthy conversation.
As a specific example, if a search result has 30,000 tweets, here’s what we take into consideration when ranking:
- Tweets from people you’re interested in should be ranked highly
- Tweets that are popular are likely to be interesting and should be higher ranked
- Tweets from bad-faith actors who intend to manipulate or divide the conversation should be ranked lower
This last bullet is the basis of our work around serving healthy public conversation. Here are some of the signals we use to determine bad-faith actors:
- Specific account properties that indicate authenticity (e.g. whether you have a confirmed email address, how recently your account was created, whether you uploaded a profile image, etc)
- What actions you take on Twitter (e.g. who you follow, who you retweet, etc)
- How other accounts interact with you (e.g. who mutes you, who follows you, who retweets you, who blocks you, etc)
We know this approach is working because we see fewer abuse reports and spam reports.
What Happened This Week
Yesterday, we identified an issue where some accounts weren’t auto-suggested in search even when people were searching for their specific name. To be clear, this only impacted our search auto-suggestions. The accounts, their tweets and surrounding conversation about those accounts were showing up in search results. As of yesterday afternoon, this issue was resolved.
There were a number of follow up questions relating to our thread yesterday that we wanted to address:
- “How many people were impacted by the search auto-suggest issue?” Hundreds of thousands of accounts were impacted by this issue. This impact was not limited to a certain political affiliation or geography. And, to be clear, these accounts were only impacted within search auto-suggestions-- they still appeared in search results. This issue has now been resolved.
- “It looks like this only affected Republican politicians. Were Democratic politicians also impacted?” Yes, some Democratic politicians were not properly showing up within search auto-suggestions as result of this issue. As mentioned above, the issue was broad-ranging and not limited to political accounts or specific geographies. And most accounts affected had nothing to do with politics at all.
- “OK, so there was a search auto-suggest issue. But what caused these Republican representatives to be impacted?” For the most part, we believe the issue had more to do with how other people were interacting with these representatives’ accounts than the accounts themselves (see bullet #3 above). There are communities that try to boost each other’s presence on the platform through coordinated engagement. We believe these types of actors engaged with the representatives’ accounts-- the impact of this coordinated behavior, in combination with our implementation of search auto-suggestions, caused the representatives’ accounts to not show up in auto-suggestions. In addition to fixing search yesterday, we’re continuing to improve our system so it can better detect these situations and correct for them.
We’re focused on making these systems better and smarter over time and sharing our work and progress with all of you. We think it’s critical to promoting healthy public conversation on Twitter and earning trust.