Bots don't optimise for size. They optimise for cover.
Bot rate on a social media account is the share of follower profiles that signal-based detection flags as non-human. Across 17,182 distinct X follower profiles we analysed between 2026-03-18 and 2026-05-24, that rate varied by 4x across account categories — and the variation is not what most outbound operators assume.
The cheapest persona for a fake X account to maintain is "aspiring solo founder following a prominent operator": generic startup-adjacent bio, no engagement-pattern red flags, blends in with the broader build-in-public scene. The hardest fake is "sales rep evaluating Apollo": no camouflage payoff, no engagement boost from other bots, the profile sticks out on inspection.
That asymmetry shows up cleanly in the numbers. If you are sourcing leads from competitor follower lists, the implication is concrete: the cleaning cost — both the noise you filter out and the LLM cycles you spend on profiles that go nowhere — depends on which kind of competitor you are mining.
Bot rate by category
| Category | Bot rate |
|---|---|
| Political / diplomatic | 61.1% |
| Solo / small-team software founders | 54.2% |
| Tech / marketing media | 49.2% |
| B2C SaaS | 45.1% |
| Community / events | 23.6% |
| B2B SaaS (Apollo, ZoomInfo, Lusha, Hunter, Leadfeeder) | 17.6% |
| Personal, non-public-figure | 11.9% |
The obvious explanation is size. The better explanation is cover.
Bot rate does increase with account follower count, which is what you would expect. High-profile accounts surface more often in news cycles, search results, trending lists, and platform follow recommendations, which makes them disproportionately visible to bots scanning for camouflage targets. Below 10K followers, bot rates in our sample cluster between 9% and 22%; above 1M, they cluster at 70% or higher.
But size does not explain the category gaps. A 23K-follower B2B SaaS account (ZoomInfo) showed 25.9% bot rate. A 66K-follower community/events account showed 23.6%. A 14K-follower solo founder account showed 18.1%, but the 100K+ founder accounts in the same category climbed to 58.5% and 85.9%.
The pattern is consistent: 17.6% bot rate on B2B SaaS accounts vs 54.2% bot rate on solo and small-team software founder accounts, despite the B2B SaaS group averaging around 8K followers per account and the founder group averaging around 211K. Smaller B2B SaaS accounts have meaningfully cleaner audiences than 10x-larger founder accounts. Size is part of the story. Category is the larger part.
How bots choose who to follow
A fake account following a prominent solo founder reads as "aspiring builder, paying attention to the indie hacker scene". The bio writes itself ("founder | building [project]", "cofounder, ex-[bigco]"), the follow choice signals nothing unusual, and the profile blends in with the actual fans, students, and aspiring-founder accounts already in that audience. The camouflage payoff is high. Bot networks also reinforce each other: following the same founder gives other bots a credible-looking "mutual connection" graph.
A fake account following a B2B sales tool — Apollo, ZoomInfo, Lusha — has none of that cover. Why would a throwaway account care about an outbound platform? The persona has to be a sales operator or marketer, which is harder to fake convincingly (specific industry knowledge, recognisable peer accounts, plausible posting history). The internal-engagement payoff is also low: bot armies do not follow each other from a Lusha account, so there is no graph effect.
Bots concentrate where the cover is good (personal-brand, political, mass-audience accounts) and stay out of where it is not (B2B SaaS audiences).
What this means for competitor-follower mining
If you are sourcing leads from a competitor or strong-signal account on X, three things follow:
1. Calibrate the cleaning overhead by competitor category. Mining a B2B SaaS competitor's follower list: expect around 17% noise. Mining a prominent solo founder: expect around 54%. That is a 3x difference in unit economics if you run this at volume.
2. Follower count alone is a poor proxy for audience quality. A 23K-follower B2B SaaS in our sample had a cleaner follower base than a 66K-follower community account. The middle of the size curve is where category dominates.
3. Cleaning cost is more than just filtering. Every bot that passes your filter and reaches the classification stage burns LLM tokens producing a verdict that converts to nothing. The economics of outbound from social signals are partly a function of how aggressive your bot screen is before the expensive stages run.
Why we filter bots before classification
This is why CTGO bot-filters at sync time, before the LLM classification stage runs at all. Cheap heuristic detection up front (signal-based: account age, posting history, bio entropy, follow patterns) catches the bulk of the noise. Only profiles that pass the screen reach the GPT pass that classifies them against a user's saved-search filters.
The economics: a tracked account at 54% bot rate would more than double the LLM spend per real lead surfaced if classification ran on every profile, since over half the GPT cycles would produce verdicts on profiles that convert to nothing. The category-by-category bot rate determines how much GPT spend an account avoids — and which categories deliver meaningfully better ROI per LLM token spent.
Methodology and caveats
The numbers above come from CTGO's classification pipeline observing 17,182 distinct X follower profiles between 2026-03-18 and 2026-05-24. Bot detection is signal-based heuristic and reports false positives and false negatives in both directions. Accuracy is consistent enough across accounts that the category-to-category comparison holds, but any specific number should be read with a ±5% tolerance.
Two things worth knowing before citing these numbers:
- This is a snapshot, not the truth. The higher-confidence findings (B2B SaaS at 17.6%, solo founder at 54.2%, tech/marketing media at 49.2%) rest on broader data within their category buckets. Other categories are illustrative only.
- The followers analysed were weighted toward more recent followers, where bots tend to concentrate, rather than a uniform sample of each account's full follower base.
Treat these as directional. The category gaps survive normal measurement noise, but precise numbers will shift as the sample grows. We refresh this dataset quarterly.