Our follower-quality scoring

Why a score and not a yes/no

“Real follower” is not a binary on any social platform. Even an entirely human-operated account drops some signals (no bio, no posts) when it's genuinely new. A binary classifier optimised against real-world data ends up with either too many false positives (humans flagged as bots) or too many false negatives (well-disguised bots flagged as humans).

We score on a 0–100 scale across three weighted components and only report a follower as “low quality” if its score is below 40. The composite is more conservative than a binary check on any single signal, which is the right trade-off when the consequence of a false positive is rejecting a follower a customer has paid for.

The rubric

Each signal returns 0 (absent) or 1 (present). Component scores are signal averages on a 0–100 scale. Composite quality = weighted sum across the three components.

How followers are sampled

Auditing every delivered follower is neither feasible nor useful — at our throughput it would consume every minute of platform-API budget and still produce a noisier number than a properly-sized random sample. We sample instead.

• Sample frame: all followers delivered across the test orders described in delivery-speed methodology during the most recent calendar month.
• Sample size: 200 followers per platform per month, drawn uniformly at random — large enough to keep the 95% confidence interval on the headline percentage within ±3 points.
• Audit time: sampled followers are scored 7 days after delivery, so any short-lived bot-purge by the platform has already settled.

What we report

Three numbers per platform, refreshed monthly:

• Median composite score — the 50th- percentile follower in the sample.
• Pass rate — the percentage of sampled followers scoring ≥ 40, the threshold below which we treat a follower as low quality.
• Component breakdown — the average per- component score, so an unusually weak component (e.g. dropping engagement signals) is visible without having to dig into raw sample data.

What disqualifies a delivery batch

A delivered batch is rejected and re-run end-to-end if either of these triggers fires during the post-delivery audit window:

• Median composite score on that batch falls below 50.
• Pass rate (composite ≥ 40) on that batch falls below 80%.

The thresholds are intentionally above the platform-wide minimums we publish on /buy-* so a batch is replaced before the public number ever moves.

Out of scope

• Existing followers of the target account are not audited — only those delivered by Likes.io.
• Demographic / niche-fit scoring (whether a follower is on-topic for a creator) is a separate concern handled by the engagement-rate calculators and is not part of this rubric.
• Private accounts in the sample contribute fewer signals (no public posts to count). We exclude them from the audit pool rather than score them on partial data.