Crypto X is one of the noisiest data environments on the public internet. Airdrop farmers run thousands of accounts, engagement pods trade likes in private Telegram groups, and follower packages are sold openly for a few dollars per thousand. For a project paying creators, that noise has a direct cost: part of every sponsored post is shown to accounts that will never sign a transaction.
The good news is that inflated accounts leave patterns in the data. No single signal below is conclusive, but together they separate organic audiences from manufactured ones with reasonable confidence. Below are the signals, how to collect them, a small scoring routine and the false positives to watch for.
What data you can realistically get
Before choosing signals, it helps to be honest about access. The official X API is paid, and the lower tiers have tight monthly read limits, so pulling the full follower list of a large account is usually out of reach. In practice you will work with a mix of three sources:
- Historical follower counts from public tracking services such as Social Blade, which record daily snapshots for many accounts.
- A sample of followers and repliers, fetched through the API or exported by hand from the account's follower and reply lists.
- Post-level metrics for the account's recent posts: views, likes, reposts, replies and quotes, plus timestamps.
A random sample of 500 to 1,000 followers is enough for most of the statistics below. Avoid taking only the newest followers, because purchased batches sit together in the list.
Signal 1: the follower growth curve
Organic growth on X looks like a noisy, mostly smooth line with occasional bumps when a post goes viral. Those bumps have a visible cause, such as a thread with unusually high views or a mention from a larger account.
Purchased growth looks different. Plot daily follower deltas and look for step functions: jumps of several thousand followers in one or two days with no matching spike in post views. Another pattern is the "sawtooth", where a large jump is followed by a slow decline over the next weeks as X removes some of the fake accounts. A simple check is to compute a rolling median of daily growth and flag any day where growth exceeds five times that median without a corresponding post above the account's usual view count.
Signal 2: follower to engagement ratios
Engagement rate naturally falls as accounts get bigger, so comparing a 20,000-follower account with a 500,000-follower account directly is misleading. A better approach is to compare an account against peers of similar size in the same niche, using median values rather than averages so a single viral post does not distort the picture.
Useful ratios include median views per post divided by follower count, replies divided by likes, and reposts divided by likes. Bought likes rarely come with bought replies and quotes, so an account with plenty of likes but a very low reply-to-like ratio deserves a closer look. The opposite case, where views are tiny relative to followers, suggests a large share of the audience is inactive or fake, because those accounts never load the timeline.
Signal 3: account age distribution of followers
This is one of the most reliable signals and one of the easiest to compute. For each sampled follower, take the account creation date and build a histogram by month. Real audiences have a broad spread with a long tail of older accounts. Purchased followers are usually created in bulk, so they cluster in a few narrow periods, often within the last year.
Combine the creation date with simple profile features to estimate the share of low-quality followers:
sample = random_sample(get_follower_ids(handle), n=1000)
profiles = lookup_users(sample) # batched, 100 ids per call
flagged = 0
for p in profiles:
score = 0
if p.post_count == 0:
score += 1
if p.has_default_avatar:
score += 1
if p.following > 20 * max(p.followers, 1):
score += 1
if days_since(p.created_at) < 120:
score += 1
if ends_with_long_digit_run(p.username):
score += 1
if score >= 3:
flagged += 1
age_hist = histogram(month(p.created_at) for p in profiles)
report(flagged / len(profiles), top_peaks(age_hist, k=3))
No threshold here is universal. A newer account in a fast-growing niche will naturally have younger followers. What matters is contrast with two or three trusted accounts in the same topic.
Signal 4: reply quality analysis
Replies carry the richest signal because they are the hardest form of engagement to fake convincingly. Collect the replies to the account's last 20 to 30 posts and run a few checks:
- Near-duplicate text. Split each reply into word shingles and compute Jaccard similarity between pairs. Clusters of replies with similarity above roughly 0.6 across different posts point to templates or automated tools.
- Lexical diversity. Count unique words divided by total words across all replies. Generic praise ("bullish", "great project", "to the moon") pulls the ratio down sharply.
- Topic relevance. Check whether replies contain any noun phrases from the original post. Bots rarely mention specifics.
- Repeat repliers. Count how many unique accounts produce the replies. If a small group of accounts appears under nearly every post, you may be looking at an engagement pod.
Language models make generated replies more fluent than they used to be, so treat text quality as one input. Repetition and the identity of the repliers usually reveal more than grammar.
Signal 5: timing patterns
Organic engagement follows the audience's day. If an account's followers are mostly in Europe, likes and replies should build during European waking hours and slow down at night. Automated engagement tends to arrive either in tight bursts, with dozens of likes landing in the first one or two minutes, or as an unnaturally even stream around the clock.
Where reply timestamps are available, compute inter-arrival times for the first hour. A healthy distribution is skewed but continuous. A cluster of replies arriving within seconds of each other, from accounts that also reply together under other creators, is a strong sign of coordination. Repliers whose own posting activity covers all 24 hours every day are another red flag.
A step-by-step method
Putting the signals together, a practical audit of one account takes an afternoon:
- Pull at least six months of daily follower counts and mark every step jump that has no matching viral post.
- Collect metrics for the last 30 posts and compute median views per follower, reply-to-like and repost-to-like ratios.
- Repeat those ratios for two or three trusted peers of similar size and compare.
- Sample 500 to 1,000 followers, run the profile scoring above and build the creation-date histogram.
- Export replies from recent posts and run duplicate, diversity and repeat-replier checks.
- Look at reply timing for a few posts published at different times of day.
- Write down a verdict per signal (clean, unclear, suspicious) rather than a single score, so the reasoning stays visible to whoever reviews it.
Technical checks answer whether an audience is real. They do not answer whether it is the right audience for a specific product, which also depends on niche, language, region and the creator's history with sponsored posts. For the wider process around these checks, including what to ask creators directly, there is a practical walkthrough on how to vet crypto influencers that covers the non-technical side.
False positives to keep in mind
Every heuristic above can misfire. A creator who went viral during a market event can show a genuine step jump. Audiences in some regions simply have more default avatars and fewer posts. Some real followers are lurkers who never post at all. And a creator may have been targeted by someone else buying followers for them, which happens in competitive niches.
That is why the method relies on several independent signals and on comparison with trusted peers. When three or more signals point the same way, the conclusion is usually solid. When only one does, ask the creator for their native analytics and look again.
Keeping the checks repeatable
If you audit creators regularly, it is worth turning the steps into a small pipeline: a script that pulls the data, stores raw snapshots, computes the ratios and histograms, and outputs a short report per account. Store the raw data, because follower lists change and replies get deleted. Rerunning the checks a month later is often the clearest way to see whether growth is real.
