Audit Your X Followers Before You Optimize for More

Before you spend another quarter chasing X followers, measure the ones you have. In the B2B accounts we manage, a follower base sitting near the platform-wide 19.42% fake-or-spam rate does more than dent a ratio. It causes posts to fail X's first-ring distribution test, so the For You feed never picks them up no matter how good the writing is.

Your Twitter Follower Quality Audit Changes What the Algorithm Sees

Run the audit before the growth push, not after. X does not grant a post out-of-network distribution on merit alone. It shows the post to a slice of your own followers first, measures what they do with it, and only then decides whether the For You feed gets it. Roughly half of what a person sees in For You comes from accounts they do not follow, and that half is the portion your follower base controls. A seed audience padded with bots and dormant accounts fails that test quietly. There is no penalty notice and no warning, just a post that tops out inside your own timeline.

The second mechanism sits upstream of the first. TweepCred is a PageRank-style author credibility score that X recalculates daily across the follow graph, and the quality of the accounts pointing at you feeds it. When a follower base is dominated by low-quality accounts, that score drops, and a lower score caps how many of your posts even enter the candidate pool that ranking operates on. That gate closes before any engagement signal is measured. You can write the best post of your year and have it never reach the stage where the first-ring test would have judged it.

November 2025 removed the last escape hatch. The Following tab used to run as a strictly chronological timeline, which meant people who deliberately followed you saw your posts whether or not the ranker liked them. Both tabs now rank by Grok's predicted engagement score. Follower quality gates reach in the feed built for strangers and in the feed built for your own audience, and the second one used to be the safe harbor for accounts with a weak algorithmic profile.

The part practitioners consistently underestimate is how long a bad follower base keeps charging you after you start fixing it. TweepCred and the heavy ranker's predicted-engagement scores are computed on rolling historical data. In before-and-after data from managed accounts, the suppression from a bot-contaminated base persists for 3-6 weeks after cleanup begins, even when current engagement quality has already improved. The algorithm's memory of your low-signal past outlasts the past itself.

That lag produces the most common failure pattern we see, and it is worth naming: the content treadmill. Reach goes flat, so the account posts more. More posts against the same contaminated seed audience produce more failed first-ring tests, and every failed test is another data point telling the ranker that this author's posts do not earn engagement. Volume makes the problem worse rather than better. The input that needs changing is who gets the first look, not how many looks you ask them for.

None of this shows up in the metrics most B2B teams review. Follower count goes up. Impressions go down. The dashboard offers no causal link between the two, so the conclusion people reach is that the content stopped landing. Sometimes it did. Often the content is fine and the seed audience stopped responding, which is a different problem with a different fix.

What 19% Fake Followers Does to a B2B Account's Organic Reach

The best public measurement of how much of X is not real comes from a joint SparkToro and Followerwonk analysis of 44,058 randomly sampled public accounts that had been active in the previous 90 days. Using a 17-signal machine learning model trained on 85,000 labeled accounts, they classified 19.42% as fake or spam. That is roughly 4x X's own official estimate of under 5% of monthly daily active users. The authors were explicit that their figure likely undercounts the more sophisticated modern bots, since the signals that catch a lazy bot do not catch a well-built one. For context on how long this has been true, the standard academic baseline, from the Varol et al. Indiana University and USC work published in 2017 using Botometer scoring, put automated accounts at 9-15% of active accounts, and that paper is still cited as the reference point in current bot-detection literature.

Scale is the reason it does not get better. X suspended 800 million accounts in 2024 for manipulation and spam violations, a figure disclosed by X government affairs executive Wifredo Fernández to British MPs in March 2026. The platform has roughly 300 million monthly active users, so the enforcement number is nearly 3x the entire active user base. Read that as a supply statistic rather than a success statistic. Bots are being created faster than they are being removed, which means follower base contamination is a continuous operational condition, not a one-time cleanup target.

The peer-reviewed picture agrees. A USC Viterbi study published in PLoS ONE in February 2025, covering January 2022 through June 2023, found a statistically significant increase in bot scores among randomly sampled X accounts after the acquisition, at p=0.01, with no reduction in coordinated inauthentic activity on any of three separate detection metrics. Whatever changed at the enforcement level over that window did not reduce what the researchers could measure from the outside.

Now the part that matters for your engagement rate math. Bot accounts generate 128.79% more tweets than expected while receiving 50% fewer likes than expected. They are loud and they are ignored. When one follows you, it adds to the denominator of every engagement rate calculation anyone runs on your account, contributes no engagement to the numerator, and occupies a slot in the seed audience that X samples when deciding whether your post deserves out-of-network reach. On the formula and on the distribution test, a bot follower is not neutral. It is negative.

The operational detail we can add from running automated accounts: contamination is not evenly distributed over time, it arrives in waves. Follower quality degrades fastest in the 48-72 hours after a post goes modestly viral or gets picked up by a hashtag with heavy bot activity. An account that audits clean on Tuesday can be 8-12% contaminated by Friday after an outreach campaign, and that new contamination quietly suppresses the following week's reach before anyone connects the two events.

This is why the intuition that a good post fixes a weak account gets things backwards. The best-performing post of the month is also the event most likely to degrade your seed audience for the month that follows. Teams celebrate the outlier, then spend three weeks wondering why nothing since has traveled. If you want to see the mechanism spelled out in X's own words, the platform manipulation and authenticity policy describes what X considers inauthentic engagement and account behavior, and the Transparency Center publishes the enforcement volumes behind the suspension numbers.

Bots vs. Dormant Real Users: Why the Distinction Changes Your Cleanup Strategy

Most follower audit content treats fake, inactive, and low-quality as one category. X does not. A bot that its spam graph has flagged and a real person who stopped logging in two years ago are both dead weight in your follower count, but they hit your account through different mechanics, and that difference should drive the order in which you remove them.

A dormant real user is a neutral zero. Real creation date, real bio, some history of posting like a person, and no current activity. That account never sees your post in the first-ring test, so it never responds, so it dilutes the response rate on your seed audience. It does not carry a negative weight in the graph. A flagged bot is a different object. It sits inside a cluster that X's graph scoring already deweights, and its presence in your follower base is a signal about you, not just a missing signal from it. TweepCred is computed over the follow graph, which means the reputation of the accounts pointing at you propagates into your own score.

The behavioral tells separate cleanly once you look at cadence rather than appearance. Bots post at volumes no human sustains, or they post nothing at all for months and then produce a burst. They follow accounts in the thousands while collecting followers in the dozens. Their creation dates cluster, because they were provisioned in batches. Dormant real accounts look nothing like this. Normal creation date, a bio someone wrote, a handful of posts that read like a person had a bad week at work in 2019, and then silence. Missing profile photos correlate with bots but do not confirm them, and using the photo alone as your filter will remove real prospects.

For a B2B audit, we work in a fixed priority order and the reason is speed of recovery, not tidiness. Accounts that X's own systems have already flagged as spam go first, because those carry the graph penalty. Zero-activity accounts with no photo and no bio go second, since they are almost certainly synthetic and cost you nothing to lose. Dormant real accounts with some past engagement history go last, and for many B2B accounts they should not go at all. A real person at a target company who has not logged in for six months is worth more as a dormant follower than as a removed one, because dormancy is reversible and removal is not.

The distinction also changes what a clean audit result looks like. An account can be near zero on bots and still fail the first-ring test, because a follower base of real people who no longer open the app produces exactly the same silence as a follower base of bots. When we see that pattern, aggressive bot removal will not fix it. The fix is on the engagement side: replies weigh far more than likes in the ranker, and a small set of real followers who reply is worth more than a large set who never open the app.

One more thing the wave pattern implies here. Contamination that arrives in the 48-72 hours after an outlier post is almost entirely bot contamination, not dormancy. Dormancy accumulates slowly over years. Bots arrive in a burst attached to a specific event. If your contamination percentage jumps between two audits, you are looking at the first category, and the cleanup is fast. If it drifts upward over quarters, you are looking at the second, and no amount of removal will substitute for changing who follows you next.

Does a Twitter Follower Audit Raise B2B Engagement, or Just the Ratio?

Start with the honest part. Removing fake or inactive followers is mathematically guaranteed to raise your stated engagement rate, because the denominator shrinks while the numerator of real interactions stays where it was. That is arithmetic, not improvement. Every specific before-and-after percentage you find in practitioner content on this topic originates from a company selling follower-audit tools, and no peer-reviewed controlled study isolates the cleanup effect from everything else changing on an account at the same time. We build tooling in this space and we still cannot point you at one, which is why you will not find an invented lift number in this guide.

The question worth asking is whether cleanup improves reach, not whether it improves the ratio. There the mechanism is documented, because X open-sourced the ranker. The heavy-ranker neural network weights a reply from a follower at 27x the value of a like, and an author replying to a reply on their own post at 150x. A like is weighted at 0.5. That spread is the whole argument for follower quality in one line: a smaller audience that replies generates dramatically more algorithmic signal than a larger audience that scrolls past, and the largest single weight in the system is on a behavior that only happens when a real human says something you can answer.

Benchmarks give you the diagnostic threshold. X's median engagement rate across all accounts was 0.15% in 2024, fell to 0.12% in 2025, and sat flat at approximately 0.10% in Q4 2025, the lowest of any major social platform in an analysis covering 70 million posts. Those medians include every dormant and abandoned account on the platform, so they are a floor rather than a target. For B2B tech accounts in the 5K-50K follower range, the working expectation is 1-3%. Below 1% is the line where we stop assuming it is a content problem. At that level there is usually both a content factor and an audience factor, and cleaning the audience is the cheaper of the two to test.

The reach improvement, when it comes, does not arrive on a straight line and it does not arrive on day one. In managed accounts, absolute impressions typically dip in the first 5-10 days after a bulk removal. The ratio looks better immediately, since the denominator moved, but the raw numbers fall, because the first-ring test is now seeded from a smaller absolute pool and the ranker has not yet repriced the quality of that pool. Practitioners who pull the dashboard at day seven see the ratio up and impressions down, cannot reconcile the two, and conclude the cleanup backfired.

It is a transient trough, not the floor. The recovery tracks the recalculation cycles described earlier: the first-ring test responds within days once the remaining audience starts clearing it, while TweepCred takes 3-6 weeks because it is rebuilt on rolling history. If you are going to run a cleanup, decide in advance that you will not evaluate it before the trough has had time to close, and write that date down. Half the failed cleanups we see failed because someone reversed course in week two.

There is a version of this work with no downside worth mentioning. If your goal is a defensible engagement rate for a partnership conversation or a board slide, the ratio improvement is immediate and real. If your goal is distribution, the ratio is a leading indicator of nothing on its own, and the number to watch is impressions per post from accounts that do not follow you, measured after the trough.

How to Run a B2B Twitter Follower Quality Audit With the Tools That Still Work

The tooling answer changed in April 2023 and a lot of published guidance never caught up. SparkToro's free Fake Followers Audit, the tool everyone recommended for years, was shut down when X eliminated the free API tier. It no longer exists. Guides published for 2026 still list it as a free option, which is a useful test you can apply while reading: if a page recommends a tool that has been gone for three years, its other tool recommendations were not verified either. Tool recommendation accuracy is the cheapest credibility check available on this topic.

What remains works on paid API access, including SparkToro's paid platform. Expect three constraints regardless of which vendor you pick. You are paying, the audit runs inside X's API rate limits so large accounts take a long time to scan end to end, and the vendor is scoring your followers with a proprietary model you cannot inspect. Treat the output as a signal, not a verdict. Compare two vendors on the same account and the percentages will disagree, because the underlying classification problem is genuinely hard. Even the most rigorous public study on this, the SparkToro and Followerwonk analysis with its 17-signal model trained on 85,000 labeled accounts, said its own numbers likely undercount sophisticated bots.

The manual audit costs nothing and it is more informative than people expect. Take a sample off your followers tab, large enough to be worth counting and small enough that you can review each account by hand in one sitting. Flag no profile photo, no bio, a handful of posts total, a recent creation date, and a following count in the thousands against a follower count in the dozens. Three or more of those on the same account is your threshold. If the share of your sample hitting that threshold approaches the platform-wide 19.42%, assume the contamination runs through the full base rather than through the slice you happened to look at.

For B2B, algorithmic health is only half of what the audit should tell you. The other half is prospecting value: what share of your followers work at companies in your ICP. X's advanced search and follower export paths can approximate this, though the API tiers constrain how much you can pull at once. This number is usually more sobering than the bot percentage. We have seen accounts with a defensible bot ratio and almost no ICP presence, which is a growth-channel problem that no cleanup will solve, and it is the number worth putting in front of whoever funds the channel.

If you are auditing someone else's account before a partnership or content deal, the incentive to get it right is financial. Fake followers in influencer campaigns cost brands an estimated $1.3 billion annually in spend against audiences that do not engage, per CNBC reporting on Sprout Social and Social Media Today data. Run the same manual sample on their followers, check their engagement rate against the 1-3% B2B range rather than the platform median, and look at whether their engagement is consistent across posts or concentrated in isolated spikes. Spiky distributions on an otherwise quiet account are worth a question before money moves.

One timing detail that distorts baselines: X purges bots in batches, and your audit date sits somewhere relative to the last one. In October 2025, X removed 1.7 million accounts in a single purge aimed at reply spam, announced by then head of product Nikita Bier. Anecdotal improvement in reply quality was reported, but no before and after engagement metrics were published. An audit run the week before a purge and one run the week after will produce different numbers on the same account with no action taken by you. Note the date of your audit alongside the result, or you will misread the next comparison.

Cleanup Velocity: Removing Followers Too Fast Triggers the Problem You Were Fixing

The number of followers you remove matters less than the rate at which you remove them. Removing 50 followers per day over 30 days and removing 1,500 in a single session end at the same follower count and produce meaningfully different algorithmic outcomes. The rapid-removal pattern is itself a spam signal. X's systems flag it, and a flagged account can have its distribution temporarily limited, which manufactures exactly the reach suppression the cleanup was supposed to cure.

The reason is not that X is protecting bots. It is that bulk graph manipulation looks the same from the enforcement side regardless of direction. A platform that suspended 800 million accounts in 2024 has detection calibrated for accounts that add or drop follow relationships in large synchronized batches, because that is what the manipulation networks do. A weekend batch job pattern-matches against the heuristics built to catch the bots you are removing. X's automation development rules are worth reading before you point any tool at this, since they set the boundaries on what automated account actions are permitted at all.

Our operational ceiling in managed accounts sits close to that fifty-a-day pace and not far above it. Accounts that push well past it during a cleanup show reach suppression that is measurably different from the normal post-cleanup trough: it starts later, it goes deeper, and it takes longer to clear. The two are easy to confuse from the dashboard, which is why we hold the pace even on accounts where the operator is impatient. The trough you planned for is survivable. The one you triggered is not on any schedule you control.

For an account carrying tens of thousands of followers worth removing, plan in months rather than in weekends. The gradual pace has a second benefit beyond avoiding flags: it lets X's daily PageRank recalculation register your quality improvement as a series of small updates rather than one large discontinuity. Incremental change reads as an account improving. A single step change reads as an anomaly, and anomalies get inspected.

There is an order-of-operations point here too. Removing followers while an outreach campaign is running means you are draining the pool and refilling it with whatever the campaign attracts, including the bot wave that shows up in the 48-72 hours after any reach spike. Pause acquisition during the removal window, or at minimum do not run both at their maximum settings in the same week. Otherwise the audit number at the end of the cleanup will be measuring two processes at once and you will not know which one produced it.

The Following Count Your Audit Is Probably Ignoring

Every follower audit tool measures the accounts pointing at you. The accounts you point at are the other half of the graph, and on any account where automation has run a follow-to-gain-follow strategy, they are the half doing the damage. TweepCred is computed with PageRank over the follow graph, and PageRank is directional. The reputation of the accounts you follow flows through your node too. An outbound list stuffed with low-quality accounts you followed to farm reciprocity drags on the same score the inbound cleanup was meant to lift.

The human version of this problem is worse than the algorithmic one, because it happens at the exact moment it costs the most. An enterprise buyer checks your X profile before a first meeting. They see an account following several times more accounts than follow it back. They do not know what TweepCred is and they do not need to. The ratio tells them the audience was acquired through reciprocal-follow tactics rather than earned, and it undercuts the authority signal the account existed to project. Nobody mentions noticing this on the call. They notice it anyway.

This is where a standard audit stops short. Removing incoming fake followers fixes the denominator in the engagement rate formula, which is the number a tool can measure and sell you a report on. It does nothing about the credibility signal a prospect checks in eight seconds. Both sides of the ratio need attention, and on accounts with automation history the outbound side is usually the faster win, since the accounts you are unfollowing were never engaging with you and their removal costs no first-ring signal at all.

The workable target is that your following count should stay comfortably below your follower count once an account is past the early stage where following outward is how you get discovered. Above rough parity, the outbound PageRank drag starts working against you at the same time as the inbound contamination, and the two compound. Apply the same velocity discipline you are using on the removal side, since mass unfollowing is a more familiar spam pattern to X's systems than mass follower removal is, and pattern-matches faster.

One caveat worth keeping. Following relevant accounts in your ICP is not a liability, it is how you find the conversations where replies happen, and replies are the 27x signal. The problem is not a high following count. The problem is a high following count made of accounts you cannot name, chosen by a tool, in a category you do not sell to.

Bot Contamination Rebuilds: the Audit Routine That Accounts for That

A single audit is a snapshot of a system that refills. Follower quality degrades fastest in the 48-72 hours after a post goes modestly viral or picks up a hashtag with heavy bot activity, and an account that audits clean on Tuesday can carry 8-12% contamination by Friday after an outreach campaign. The degradation is silent. It shows up as next week's reach being worse for no reason anyone in the room can name, and by then the connection to Tuesday's audit result has been lost.

The supply side supports treating this as maintenance rather than a project. X suspended 800 million accounts in 2024, nearly 3x its monthly active user base, and the USC Viterbi PLoS ONE study found no reduction in coordinated inauthentic activity across any of its three detection metrics over January 2022 through June 2023. Enforcement at that volume with that result means the creation rate exceeds the removal rate. Your account is downstream of that. Audit cadence matters as much as audit methodology, and most published guidance covers only the second one.

The routine we run on managed B2B accounts has four parts and no complexity. Full audit every quarter. Extra audit after any post that meaningfully exceeds the account's normal reach, run inside the 48-72 hour window rather than after it. Weekly engagement rate check as a cheap proxy, where a drop with no corresponding content change is a contamination indicator worth investigating rather than a content verdict. And a standing monthly removal cadence at safe velocity, which keeps you from ever facing a backlog large enough to tempt you into the weekend batch job.

The monthly cadence is the piece people skip and the piece that does the work. Removal at a low steady pace never triggers the bulk-manipulation heuristics, never produces a deep trough, and lets the daily PageRank recalculation absorb your improvement continuously. It also means the quarterly audit becomes a verification step rather than the start of a two-month remediation, which is a much easier thing to keep doing after the person who was excited about it moves to another project.

Track your audit dates against X's purge history, which is public and lumpy. When the platform removes a large batch of bots, as it did with the 1.7 million accounts pulled in October 2025, your engagement rate improves without you doing anything, because the denominator dropped while your real interactions did not. The improvement is partial, the supply refills within weeks, and any baseline you set during that window is skewed. Note the date next to the number, compare like for like, and resist the temptation to attribute a purge's arithmetic to your own work.

Frequently asked questions

How do I check if my Twitter/X followers are real or fake without paying for a tool?

Spot-check 100 accounts from your followers tab manually. Red flags: no profile photo, no bio, under 10 tweets, account created in the past 90 days, following 5,000+ accounts with under 50 followers in return. If more than 20% of your sample matches three or more of these signals, you likely have broader contamination. SparkToro's free audit tool closed in April 2023, so manual sampling is the only no-cost method that still works.

What percentage of Twitter/X followers are bots in 2025-2026?

A SparkToro and Followerwonk joint analysis of 44,058 randomly sampled active accounts found 19.42% classified as fake or spam using a 17-signal ML model. That is roughly 4x X's own official estimate of under 5% of monthly active users. A USC Viterbi peer-reviewed study (PLoS ONE, February 2025) found bot scores increased post-acquisition with no reduction in coordinated inauthentic activity, suggesting the problem has not improved since early 2023.

How does follower quality directly affect your Twitter engagement rate and algorithmic reach?

Follower quality affects reach through two mechanisms. First, X runs a first-ring engagement test on the author's existing followers before granting For You distribution; a bot-heavy follower base fails this test and posts stay small. Second, follower quality feeds into TweepCred, a PageRank-style score recalculated daily; a lower score caps how many posts enter the candidate pool. Since November 2025, both feed tabs rank by predicted engagement, so follower quality gates reach in both.

Should I remove fake or inactive followers from my X account, and will it improve performance?

Removing them is worth doing, but the timing of when you measure matters. Removal is mathematically guaranteed to raise your stated engagement rate (smaller denominator, same real interactions). The algorithmic reach improvement is real but delayed: expect a dip in absolute impressions for 5-10 days before recovery, because the first-ring test is temporarily seeded by a smaller pool. TweepCred recovery takes 3-6 weeks because the score is built on rolling historical data.

What signals identify a Twitter follower as a bot versus a dormant real account?

Bots typically have no profile photo, post at non-human cadences (50+ per day or months of silence), follow thousands of accounts while having under 50 followers in return, and have creation dates clustered around a known bot wave. Dormant real accounts have a normal creation date, some past tweet history, and a bio. The distinction matters because X's graph scoring actively deweights flagged bots while dormant real accounts are neutral zeros; removal priority should reflect that difference.

How do B2B brands audit a Twitter account before an influencer partnership or content deal?

Check four things: (1) engagement rate against the 1-3% benchmark for B2B tech accounts in the 5K-50K range, (2) the following-to-follower ratio (above 150% is a reciprocal-follow signal), (3) a manual sample of 100 followers for bot signals, (4) whether recent post performance shows consistent engagement or isolated spikes, which can indicate bot-boosted posts. Tools like Modash or Audiense can automate the follower quality scan if the account is large.

What is a healthy fake-follower percentage for a B2B Twitter account, and when does it become a problem?

Given that the SparkToro study found 19.42% fake accounts across the platform broadly, some baseline contamination is unavoidable. For B2B accounts, above 15% fake or inactive followers starts to measurably suppress the first-ring engagement test and pull TweepCred down. The more diagnostic signal is engagement rate: a B2B tech account with 5K-50K followers below 1% engagement rate has a follower quality problem contributing alongside any content factors.

How fast can I safely remove fake followers on X without triggering account restrictions?

In managed accounts, the safe ceiling is 50-75 removals per day. Removing 1,500 followers in a single session pattern-matches against the same bulk-manipulation heuristics X uses to detect spam accounts, and can temporarily limit account distribution. For accounts with tens of thousands of followers to clean, plan a 60-90 day window. The gradual approach also lets X's PageRank recalculation register quality improvements incrementally rather than processing one anomalous reset.

Do fake followers hurt your Twitter reach even after you improve your posting consistency?

Yes, and the suppression persists longer than most practitioners expect. TweepCred and the heavy-ranker's predicted-engagement scores are built on rolling historical data. In managed accounts, the algorithmic memory of a low-signal follower base continues suppressing distribution for 3-6 weeks after cleanup begins, even when current content and engagement quality have improved. The first-ring test responds faster than TweepCred, so early-session distribution gains may appear before broader reach recovery.

What Twitter/X follower audit tools still work after the 2023 API shutdown?

SparkToro's free Fake Followers Audit closed in April 2023 when X eliminated the free API tier. Tools that still operate on paid API access include SparkToro's paid platform, Audiense, and Modash. All require paid subscriptions and work within X API rate limits, meaning large-account audits can take 24-48 hours to complete. If a guide you are reading still lists the SparkToro free tool as available, treat the rest of its tool recommendations carefully.

Sources and further reading

• SparkToro and Followerwonk joint analysis finding 19.42% of active accounts are fake or spam
• X Transparency Center platform manipulation enforcement data
• X's platform manipulation and authenticity policy