How the Fake Follower Audit works
What the audit looks at for each follower of an X (Twitter) account, how it turns that into a score and a verdict, and how accurate it was when we checked it by hand.
What it reads
The audit works on the same public profile fields as a twtData followers export: account creation date, profile photo, post and like counts, followers and following counts, username, display name, bio and verification type. Only public profile data is used: no private accounts' posts and no direct messages.
The signals
Each follower collects points for these signals. Each one is shown in the report as a reason you can check on the profile yourself.
- New account: created less than 30 days ago (3 points) or less than 90 days ago (2 points). Bought followers are usually new.
- No profile photo (2 points).
- Never posted (2 points), or fewer than 5 posts (1 point).
- Follows far more accounts than follow it back: following 200+ with at least 20 times more following than followers (2 points), or following 1,000+ at 10 times or more (1.5 points). No followers at all scores 1 point.
- Auto-generated username: letters followed by six or more digits, or eight or more digits at the end (1.5 points).
- No bio (1 point).
- Calls itself a bot in the bio or name ("bot", "automated", "auto-posts"; 3 points).
- Spam-style bio: follow-back bait, "send me crypto", adult-account wording (4 points), or the stock fragments of a random-profile generator (2.5 points).
Some things lower the score: a year or more of real posting and liking (100+ posts and 100+ likes) takes off 1.5 points, and a paid blue checkmark takes off 1 point. Spam networks do buy checkmarks, so it lowers the score but never clears an account. Accounts verified as a business or government, and accounts older than five years with 500+ posts, are treated as real without scoring.
The score and the verdict
The signal points are scaled to 0-100 (8 points or more is 100). That makes up 65% of the final score. The other 35% comes from our machine-learning profile model. On its own that model is weak: it scored ten self-declared bots as human in a test on 2 October 2026, which is why it is only one part of the score. When the model is unavailable, the score is the signal part alone.
- Likely fake: score 70 or more.
- Suspicious: 40 to 69.
- Real: under 40.
The report's overall risk level is our own scale: low when under 5% of followers are likely fake and under 20% are flagged in total; high from 15% likely fake or 40% flagged.
Two things a single-profile checker can't see
When followers arrived. X lists an account's followers newest first, and the export keeps that order. The audit groups flagged followers by when they followed and marks a burst when the newest 1,000 followers are flagged at twice the rate of older ones or more, and at least 10 points higher, with 200+ scored accounts on each side. Bought followers tend to arrive together.
Followers X already removed. Followers in the list whose profiles can no longer be read are accounts X has suspended or that were deleted. The report counts them separately.
How accurate it is
We first checked the audit against which followers X suspended in the weeks after an export. Re-checking a 4,750-follower sample from five exports two to four weeks later found no suspensions at all, too short a window to use as labels.
So on 3 October 2026 we reviewed 150 followers by hand, 50 from each verdict, judging each from the profile and its recent posts without seeing the verdict:
| Audit verdict | What the review found |
|---|---|
| Likely fake (50) | 24% fake or bot, 72% inactive or empty, 4% real and active |
| Real (50) | 90% real and active |
| Suspicious (50) | A mixed group: 42% real and active |
The misses led to two signals (spam-style bios and the 10x follow ratio) and the credit for a year of real posting. Re-scored on the same 150 after that, every "likely fake" account was fake or inactive. That figure is optimistic, because the rules were tuned on those accounts. On the 4,750-follower sample the current rules give 84.5% real, 12.4% suspicious and 3.1% likely fake. The model alone would have called 41% of the same sample bots.
Limits
- A flag is a reason to look closer, not proof. Many "likely fake" accounts are abandoned or empty rather than bought, and the report says which reasons fired.
- The review is small (150 accounts) and was done by us. It is a check, not an independent benchmark.
- "Suspicious" is a mixed group by design, and it is reported separately.
- The audit judges followers from public profile fields. It cannot see who paid for followers or how an account is run.
Cite this page
twtData (2026). How the Fake Follower Audit works: method and accuracy. https://twtdata.com/bot-detector/methodology/