Thread by @andst7

🧵This spam acc that compulsively shares and puts likes on porn content is considered an mDAU (monetizable daily active user) by Twitter

But it is not alone. We estimate that false, spam or automated accs could represent 12-14% of current mDAU

It's a long thread, let's start 👇🏻

/2 Twitter doesn't share info about mDAU. Nor its metrics and methodology.

Twitter defines mDAU as accounts that are logged in and can see ads. Usually, researchers have different ways to analyze accs, but no one seems to help estimate mDAU except one…

3/ We ran sponsored campaigns.

As the objective of our sponsored campaigns, we picked engagement. Twitter charges when people we target engage with our content. Impressions that don’t generate an engagement are free. We decided to run five campaigns for less than 24 hours.

4/ Targeting and campaign settings.

As demographic preferences, we selected: 🇺🇸, 🇮🇳, 🇵🇰, 🇮🇩, 🇧🇷 (countries with the largest population, excluding 🇨🇳).

As targeting features, we decided on 5 KWs related to 5 different topics: from e-commerce to phone accessories and beauty care.

5/ This step is essential.

We created a "hidden tweet," which means that the tweet is only visible as sponsored from accounts that are logged in and eligible to see ads (mDAU). The text of the tweet contained only 5 KWs. The content of the tweet didn't have any sense.

6/ We know that fully-automated or hybrid spam/fake accounts navigate Twitter and engage with tweets containing specific keywords or linked to particular accounts or hashtags.

We were right 🧞‍♂️

7/ At the end of the sponsored campaign, we had over 500 likes and a couple of retweets.

We downloaded the complete list of users that interacted with our sponsored campaigns with a scraper. We started analyzing them 🧐🔎🕵🏼

8/ We analyzed the dataset. We compared and found similar results even with different methods. We even performed a manual review of the whole dataset.

Botometer Pro API and Botometer Web;
GD;

9/ GD is the software we use to analyze Tiktok/IG, adapted for this analysis;
Botometer is a tool focused on Twitter. Many individuals and Universities developed it arxiv.org/abs/2006.06867

Furthermore, we performed manual reviews relying on our multi-year experience.

10/ Additionally, we set a similar campaign.

We got another 300 interactions and a similar percentage of spam, fake and automated accounts that interacted, but we didn't have time to perform a manual review.

So, we won't include the results. Anyway…

11/ We paid to get interactions from bogus accounts 😒😡

Some of these accs with signs of automation liked more than half-million tweets in a couple of yrs and produced just a few tweets. While some compulsively share porn content, others only RT but never make a single tweet

12/ Important notes and considerations 📝

Both humans 🧒🏿👦🏻👩🏻 and software 🤖 can spam;
We should consider a more extensive dataset in the future;
Our tweet got the most engagement from non-US countries 🌎;
The result should be repeated and validated over time ✅;

13/ Notes/Considera.[2]

The few accounts that RTed our sponsored tweet may have created some noise;
We invite researchers to make similar tests;
We tried to be conservative / our figures might be inaccurate considering the tech challenges.

Sharing is knowledge! ⚡️

14/ Conclusion

Twitter is wrong. Elon Musk @elonmusk is right. Now additional research can assess a more accurate number, but we're sure that 5% of Twitter is far from reality.

15/ Acknowledgments

I want to thank Kirill, Pasha, and Nicola who worked in the past days non-stop. I want to thank hundreds of people that supported me these days.

Again: If you liked the thread, please donate to a charity 🌱❤️

Ps: Now I'll be off for a while. Ad maiora 🙋‍♂️