How fake follower detection actually works, why there is no reliable ground truth to measure it against, and the innocent explanations that trip every signal.
- There is no reliable label for a fake follower, so any accuracy claim is measured against someone else's estimate, not truth.
- Every detector leans on 1 main signal, engagement lower than the follower count suggests, and that signal has several innocent causes.
- Detectors only assess followers whose profiles are public and readable, which is not a random slice.
- With no shared definition, 2 tools can give different percentages for the same creator and neither is wrong.
- Validate against platform purges, the closest thing to ground truth, and treat the score as a flag for human review, not a filter.
Fake follower checks are standard in influencer tooling, and almost nobody explains what is under them. The short version is that they are a set of reasonable heuristics, applied to a sample, producing an estimate that cannot be checked against anything solid.
That is not a criticism. The heuristics are sensible and the output is useful. The problem is that it gets presented as a measurement, and buyers treat it as one.
Here is how the detection works, where each signal breaks, and what you can actually trust.
Why is there no ground truth?
Because nobody outside the platform knows which accounts are fake, and the platforms do not publish it.
It is also not 1 thing. 4 different populations get called fake and they behave differently:
- Bots. Automated, no human behind them. The easiest to spot.
- Click farm accounts. Real people, paid to follow. They look human because they are.
- Dormant real accounts. A person who signed up in 2019 and stopped using the app. Real, just gone.
- Mismatched real accounts. Active humans with no interest in this creator, often picked up through a giveaway.
Only the first is clearly fake. The other 3 lower engagement without anyone buying anything, and most detectors cannot tell them apart. So when a tool says 18% fake, read it as 18% of sampled followers that did not look engaged, which is a different and much weaker claim.
What signals do detectors use?
| Signal | What it looks for | Strength |
|---|---|---|
| Engagement versus followers | Likes and comments far below what the follower count implies | The main signal in most tools, and the most confounded |
| Follower growth pattern | Sudden spikes with no matching content event | Good when you have stored history. Useless on a first look |
| Follower profile quality | Followers with no photo, no posts, default usernames | Reasonable, and only works on readable profiles |
| Follower to following ratio | Accounts following thousands and followed by few | Decent for spotting farm accounts |
| Comment quality | Generic or emoji-only replies, repeated across posts | Useful. Also catches legitimate fan behaviour |
| Geography versus language | Audience country that does not match the content language | Weak on its own. Plenty of honest reasons |
| Account age versus size | A young account with a very large following | Weak. Viral growth is real |
Most products combine several of these into 1 score. That is sensible modelling and it makes the result hard to argue with, which is the problem we flagged in the screening risk taxonomy: a blended number cannot be taken apart when someone questions it.
2 of these signals also need data you may not have. Growth pattern needs stored history, because public sources return a current snapshot rather than a time series, as we covered in the freshness benchmark. And engagement rate has no standard denominator, so your number will differ from a competitor's for reasons unrelated to fakes. The 4 defensible choices are set out in how to build an influencer discovery product.
If you want measured audience data rather than an inference from the outside, the field list per platform is public. See the coverage list
Where does each signal fail?
This is the part worth printing out, because every one of these has an innocent explanation that looks identical to fraud.
| Signal | Innocent cause that mimics it | Who it hurts |
|---|---|---|
| Low engagement | Algorithm changes cutting reach across a whole platform | Everyone, at once, for reasons unrelated to followers |
| Low engagement | A creator posting less often than they used to | Anyone changing cadence |
| Low engagement | Large accounts naturally get lower rates than small ones | Macro creators, systematically |
| Growth spike | A post went viral, or a press mention landed | The creators doing best |
| Poor follower profiles | Followers with private accounts cannot be assessed at all | Creators with younger audiences |
| Geography mismatch | A bilingual creator, or a diaspora audience | Non-English creators, systematically |
| Generic comments | Fan accounts and younger audiences comment in emoji | Creators with teen audiences |
Look at the right-hand column. The failures are not random, they cluster on specific creator types. A detector that penalises non-English audiences and teen-heavy accounts is not just noisy, it is biased in a direction that will show up in who gets booked.
Why does the sample matter so much?
Because a detector can only assess followers whose profiles it can read, and private accounts are not readable.
That is the same selection problem we worked through in the audience demographics post. Sampling error gets small quickly, around 2 percentage points at 2,500 sampled followers. Selection bias does not shrink at all, no matter how many you sample.
It is also worth knowing that follower lists are not exposed by the major platform APIs at all, which is why detectors work from what is publicly visible rather than from a complete list. The field-level picture is in social data API coverage.
So the question to put to a vendor is not how large their sample is. It is what share of followers they could not assess, and what they do with that share. Dropping it quietly assumes the unreadable followers look like the readable ones, and on a creator with a young or privacy-conscious audience that assumption is doing a lot of work.
What can you actually validate against?
Platform purges. When a platform removes a batch of inauthentic accounts, followers disappear, and that is as close to a verdict as you will get.
# Validate a detector against what the platform actually removed.
# 1. Snapshot the follower list and your flags, with a date.
snapshot = {"taken": "2026-04-01",
"followers": [...],
"flagged": {...}} # account_id -> your score
# 2. Wait. Re-check after a platform sweep.
# 3. Compare.
removed = snapshot_followers - current_followers
precision = len(removed & flagged) / len(flagged)
recall = len(removed & flagged) / len(removed)
# Honest caveats, state them with the numbers:
# - accounts also vanish because people quit or go private
# - purges are infrequent and uneven across platforms
# - this measures agreement with the platform, not truth
#
# It is still the only external check available, and a vendor
# who has never run it has not tested their detector against
# anything but their own labels.Run that once and you learn more about a tool than any accuracy claim on its website. You will need to hold snapshots for a while, which means storing follower lists over time, and that is worth doing anyway for the growth signal.
How should you use the number?
As a prompt to look closer, not as a filter. If a creator runs several accounts, check them together rather than separately, which is what identity resolution is for.
- Set a review threshold, not a reject threshold. A high score means a human opens the profile, not that the creator is excluded.
- Check the comments yourself. 2 minutes of reading tells you more than the percentage did, and it is the same manual step that sits behind influencer vetting.
- Compare like with like. Engagement rates fall as accounts get bigger, so judge a creator against others of similar size.
- Ask the creator to connect. Phyllo's social data API returns real reach and real audience location instead of an inference, which settles the question outright. That split is in authenticated versus public social data, and the permission that unlocks it on Instagram is listed in the Instagram scope reference.
That last one is the honest resolution. Fake follower detection exists because you are guessing about an account from the outside. Once a creator connects, you can read what the platform actually recorded and the guess becomes unnecessary.
Summary
Fake follower detection is a reasonable set of heuristics producing an estimate nobody can verify. The signals are sensible, the main one is engagement lower than the follower count implies, and it has several ordinary explanations including platform reach changes, posting less often, and simply being a large account.
The sample is the other weak point. Detectors assess only followers whose profiles they can read, and the ones they cannot read are not a random group.
Use the score to decide who to look at. Read a few comment threads. And where the decision involves real money, ask the creator to connect so you can see measured numbers instead of inferred ones.
Want measured audience and engagement data instead of an estimate from the outside? Get a demo
How accurate is fake follower detection?
Nobody can say. There is no reliable list of fake accounts to check against, and platforms do not publish which they consider inauthentic, so any accuracy figure is agreement with another estimate.
Why do 2 tools give different fake follower percentages?
They sample different followers, weight signals differently and define fake differently. There is no shared standard, so a creator can be 7% on one tool and 22% on another with neither being wrong.
Is a low engagement rate proof of fake followers?
No. Engagement falls when a platform changes distribution, when a creator posts less often, and simply as an account grows. Low engagement is a reason to look, not a conclusion.
Can you detect fake followers on a private audience?
Only partially. Detectors assess only followers with public, readable profiles, so private ones usually drop out of the sample. For younger or privacy-conscious audiences, that can be a large share.
What is the most reliable check?
Asking the creator to connect, which returns the platform's own reach and audience figures. Short of that, comparing your flags against accounts the platform later removes is the only external check.
