Bot Farms and Fake Engagement Guide

A bot farm is an operation that runs large numbers of fake or hijacked accounts to manufacture followers, likes, views, reviews and comments, and the same infrastructure is used for commercial fraud and for state-backed influence campaigns. Fake engagement distorts ranking systems, misleads customers and advertisers, and gives scams and disinformation borrowed credibility. In 2024 the US Justice Department and partners in Canada and the Netherlands disrupted an AI-enhanced bot farm that ran fictitious personas on X, and both the US Federal Trade Commission and the UK now treat buying fake reviews or fake social media influence as unlawful in defined circumstances. This guide explains how bot farms work at a level defenders need, the signals that expose them, the layered controls platforms use, what brands and marketing teams should check before paying for reach, and the rules that apply. It is for trust and safety, fraud and security teams at platforms and marketplaces, and for brand, marketing and procurement teams who buy influence.

Key takeaways

Bot farms are an identity problem. They depend on creating accounts at scale or taking over real ones, so the strongest controls sit at sign-up, verification and account recovery, not only in content moderation.
Clusters give them away. Single fake accounts can look human; groups that share creation patterns, infrastructure, behaviour and narratives rarely do.
AI made personas cheap, not invisible. The Meliorator advisory describes AI-generated profile photos and biographies, yet still lists detectable signals such as uniform user agents and formulaic profiles.
Buying fake engagement now carries legal risk. FTC rule 16 CFR 465.8 covers buying and selling fake indicators of social media influence for commercial purposes, and UK DMCC Act Schedule 20 paragraph 13 bans fake and concealed incentivised reviews from 6 April 2025.
Measure prevalence, not takedowns. A falling count of removed accounts can mean you are winning or that you have stopped looking; track the share of engagement that comes from inauthentic accounts.

What bot farms are and how they operate

The term covers several operating models that are often combined. The OWASP Automated Threats to Web Applications project gives defenders a shared vocabulary: OAT-019 Account Creation (bulk creation of accounts for later misuse, including fake social media accounts), OAT-016 Skewing (repeated clicks, requests or submissions intended to alter a metric, with examples including boosting likes, poll, rating and review fraud, and synthetic social media activity) and OAT-017 Spamming (adding malicious or questionable content, including review spam, reply bots and social media bots).

Model	How it works	What it is used for	Typical tell
Automated fake account network	Software creates and runs personas through web or API sessions	Follower and like inflation, amplification of narratives	Shared creation patterns, uniform client fingerprints, synchronised activity
Device or phone farm	Real handsets or emulators run apps to look like individual mobile users	App installs, views, reviews, engagement that must pass device checks	Many accounts per device lineage, emulator artefacts, identical app inventories
Click farm with human workers	Low-paid workers perform likes, follows and reviews by hand	Engagement that must defeat behavioural bot detection	Narrow working hours, geographic concentration, task-like bursts
Hijacked real accounts	Accounts taken over through credential stuffing or infostealer logs are rented out	Credible-looking engagement with account history	Sudden change in behaviour, language or location of an established account
Engagement pods and exchanges	Real users trade likes, follows or "sub4sub" arrangements	Gaming ranking algorithms	Reciprocal engagement graphs, engagement unrelated to content quality
State-backed persona operations	Software-managed fictitious personas with tailored backstories	Influence and disinformation	Narrative coordination across clusters, persona archetypes

MITRE ATT&CK describes the resource-development side of these operations: T1585.001 Establish Accounts: Social Media Accounts, T1586.001 Compromise Accounts: Social Media Accounts, and botnet acquisition under T1583.005 and T1584.005. Mapping to these helps security teams share findings with threat intelligence colleagues. See the Botnets Guide for the infrastructure side.

Case study: the Meliorator bot farm

On 9 July 2024 the US Justice Department announced the disruption of a bot farm that it said was developed by affiliates of Russia's state-run RT and operated with the involvement of the FSB. The department searched 968 X accounts, seized two US-registered domains used to create email accounts for the personas, and X voluntarily suspended the remaining bot accounts identified in court documents. A joint advisory (AA24-191A) from the FBI, the US Cyber National Mission Force, Dutch intelligence and police services and the Canadian Centre for Cyber Security described the software, known as Meliorator, in defensive detail:

Two components: an administration panel (Brigadir) that managed personas and their actions, and a back end (Taras) that generated personas and automated their activity.
Persona archetypes: fully detailed profiles, minimal profiles that mainly liked or reposted, and profiles built from scraped public data to look established, with AI-generated profile images and biographies.
Building credibility: personas followed genuine users whose interests matched their stated politics and accumulated followers and activity over time, then amplified narratives from both bot and real accounts.
Recommended platform mitigations: validate that accounts are created and run by real people using processes similar to Know Your Customer practices, strengthen authentication and verification, flag users with known-suspicious user agent strings, and make accounts secure by default, including MFA.

The lesson for defenders is that generative AI lowered the cost of credible-looking personas but did not remove the operational fingerprints of running them at scale. See the Disinformation and Brand Impersonation Guide for the brand side of influence operations.

Signals that expose bot farms

No single signal is reliable on its own: real users share networks, new users have thin profiles, and enthusiastic fans post in bursts. Combine signals across layers and act on clusters.

Layer	Signals	Notes
Registration	Velocity of sign-ups from one source, disposable or newly registered email domains, phone numbers from virtual or recycled ranges, form-fill timing, sequential or templated usernames	The Meliorator operators registered their own email domains, so domain age and ownership matter as much as known disposable lists
Device and client	Uniform or rare user agents, emulator or automation framework artefacts, many accounts per device identifier, headless browser traits	The advisory specifically recommends flagging suspicious user agent strings
Network	Data centre, proxy and residential proxy ranges, many accounts behind one address, impossible travel	Residential proxies make IP reputation weaker; use it as one input
Profile	AI-generated or reused photos, formulaic biographies, creation dates clustered in time, profile completeness that matches an archetype	Generated faces are improving; treat image analysis as supporting evidence
Behaviour	Likes and reposts with little original content, synchronised actions across accounts, activity at fixed intervals, engagement on content immediately after publication	Timing correlation across accounts is one of the strongest cluster signals
Graph	Dense follow rings, reciprocal engagement, follower growth in steps, followers concentrated in unrelated geographies	Graph analysis catches human click farms that behaviour models miss
Content and narrative	Near-duplicate text, the same links or hashtags pushed by a cluster, reviews posted in bursts with similar phrasing	Language models make paraphrasing easy, so look at shared targets and timing, not only shared text

Layered controls for platforms and marketplaces

Controls work best as a pipeline: make account creation expensive, make automation detectable, and make the engagement worthless when it gets through.

Sign-up friction proportional to risk: invisible bot detection for most users, step-up challenges for risky sessions, and phone or identity verification where abuse is concentrated. CAPTCHAs alone are a speed bump: OWASP lists CAPTCHA Defeat as its own automated threat (OAT-009). See the CAPTCHA and bot challenges guide and the Fake Account Creation and Sign-up Fraud Guide.
Bot management at the edge: client fingerprinting, behavioural analysis and rate limits on follow, like, review and message endpoints, including the API. See the Bot Management Guide.
Protect real accounts from takeover: hijacked accounts are the most credible fake engagement. MFA by default, credential stuffing defences and session protections reduce supply. See the Account Takeover Prevention Guide.
Trust tiers for new accounts: limit the weight, reach or rate of actions from accounts until they build history, so freshly created accounts cannot move rankings.
Discount, then remove: remove suspected inauthentic engagement from counts and ranking signals quickly, and remove accounts after review. YouTube's fake engagement policy, for example, states that artificial traffic is not counted toward views and spam subscribers do not count toward totals.
Cluster takedowns: act on the whole network at once; piecemeal removal teaches operators which signals you use.
Policy that names the behaviour: prohibit buying, selling and promoting artificial engagement, including through third-party services, and apply consequences to the beneficiary, not only the bots. YouTube's policy counts hired promoters' actions against the channel.
Information sharing: share indicators with peers and authorities, and vet indicators before blocking: the Meliorator advisory warns that some listed infrastructure belongs to legitimate services.

Choosing controls by size

Organisation	Priorities	Reasonable approach
Small community, forum or review site	Spam and fake reviews	Managed bot protection, email and phone verification for posting, review moderation queue, rate limits
Marketplace or app with ratings	Review fraud, seller boosting	Verified-purchase reviews, seller-linked account detection, graph analysis of reviewers, legal compliance programme
Large social platform	Engagement manipulation, influence operations	Dedicated integrity team, cluster detection models, trust tiers, threat intelligence and public reporting

For brands, agencies and buyers of influence

Most organisations never run a platform, but they do buy reach: influencer campaigns, social advertising, app install campaigns and review programmes. Fake engagement wastes that money and can now create legal exposure for the buyer.

Audit before paying: look at follower growth history, the ratio of engagement to followers, the geography and age of followers, and comment quality. Sudden step increases in followers and generic comments are warning signs.
Write it into contracts: require that creators and agencies do not buy followers, views, likes or reviews, give you the right to audit and claw back fees, and require disclosure of paid relationships.
Pay for outcomes you can verify: conversions and verified actions are harder to fake than impressions and likes.
Never buy reviews or incentivise them without disclosure: under the FTC rule and the UK DMCC Act, the business that commissions or benefits from fake reviews is exposed, not just the seller.
Watch your own accounts: a sudden wave of followers to your brand account may be a bot farm building cover or a competitor trying to discredit you; document it and report it to the platform.

Creator and brand accounts are also prime targets for takeover and resale. See the Security for Content Creators and Influencer Accounts Guide.

Rules that apply

Jurisdiction	Instrument	What it covers	Status
United States	FTC Trade Regulation Rule on the Use of Consumer Reviews and Testimonials, 16 CFR Part 465	Fake or false reviews and testimonials, including AI-generated ones (465.2); buying reviews (465.4); review suppression (465.7); selling, distributing, buying or obtaining fake indicators of social media influence, such as followers, views, likes, shares and comments, where the person knew or should have known they were fake and they could be used to misrepresent influence for a commercial purpose (465.8)	Announced 14 August 2024; published at 89 FR 68077 on 22 August 2024; the FTC can seek civil penalties against knowing violators
United Kingdom	Digital Markets, Competition and Consumers Act 2024, Schedule 20 paragraph 13	Submitting or commissioning fake reviews or concealed incentivised reviews; publishing reviews in a misleading way; a duty on traders who publish reviews to take reasonable and proportionate steps to prevent and remove them; offering services that facilitate this	In force from 6 April 2025
Platforms	Platform rules, for example YouTube's fake engagement policy	Artificial inflation of views, likes, comments and subscribers, promotion of third-party engagement services	Warnings, strikes and termination for repeat violations

This table summarises the instruments as published on the official sources listed below; check the current text before relying on it. See the Online Fraud Regulation Map for wider fraud regulation.

Where identity fits

Every account in a bot farm is a non-human actor presenting as a human. The questions identity teams ask about non-human identities apply directly: who created this account, what proves a person controls it, and what is it allowed to do? Three identity controls matter most.

Proofing at the right moments: stronger verification when an account starts to carry weight (monetisation, verified badges, advertising, high-volume messaging) rather than for every sign-up.
Separating declared automation from impersonation: legitimate bots and AI agents should be able to identify themselves, for example through declared automated accounts and signed agent traffic, so that undeclared automation stands out. See the Bot vs AI Agent Traffic Management Guide.
Recovery and takeover controls: account recovery is where bot farms acquire aged, credible accounts. Treat recovery flows with the same rigour as sign-in.

Detection and measurement

Metric	What it shows	How to use it
Prevalence: share of engagement (likes, follows, reviews) from accounts later found inauthentic	Real impact on users and rankings	Estimate with regular sampled human review, not only model scores
Time to action on detected clusters	How long fake engagement influences rankings	Measure from first detection to discounting and removal
Proportion of fake accounts caught at sign-up versus after activity	Where controls work	Shift detection earlier over time
Recidivism: removed operators returning with new accounts	Whether takedowns stick	Link new clusters to old ones by infrastructure and behaviour
False positive rate from appeals	Cost to genuine users	Track overturned decisions by signal type
For buyers: share of campaign spend with creators failing audit	Money exposed to fake reach	Report per agency and per campaign

How to implement

Baseline: sample accounts and engagement on your most valuable surfaces (reviews, follows, trending) and estimate prevalence by human review.
Instrument: log registration, device, network and action events with consistent identifiers so clusters can be built.
Deploy detection in monitor mode: score accounts and clusters without enforcement and compare against human review to tune thresholds.
Enforce in stages: first discount suspect engagement from counts and rankings (easily reversible), then restrict actions, then remove accounts after review. Keep an appeals route and the ability to restore accounts and counts.
Close supply: raise friction at sign-up and recovery where clusters originate, and strengthen account takeover defences.
Review quarterly: operators adapt, so re-test signals, update policies and share indicators.

Common failure patterns

Counting removals as success: reported takedowns rise or fall with effort, not with the problem. Measure prevalence.
Relying on CAPTCHAs and IP reputation: both are defeated cheaply with solving services and residential proxies.
Removing accounts but leaving the engagement: follower counts, ratings and rankings keep the benefit after the bots are gone.
Ignoring hijacked accounts: controls tuned for new fake accounts miss aged, compromised accounts rented to farms.
Buyers trusting dashboards: marketing teams pay on impressions and follower counts without auditing where they came from.
Acting on unvetted indicators: blocking shared infrastructure from an advisory without checking it can cut off legitimate users.

Framework mapping

Framework	Reference	How it applies
OWASP Automated Threats	OAT-019 Account Creation; OAT-016 Skewing; OAT-017 Spamming; OAT-009 CAPTCHA Defeat	Vocabulary for bulk accounts, metric manipulation, spam and challenge bypass
MITRE ATT&CK v20	T1585.001 and T1586.001 Social Media Accounts; T1583.005 and T1584.005 Botnet	Resource development for influence and fraud operations
Regulation	16 CFR 465.2 and 465.8 (US); DMCC Act 2024 Schedule 20 paragraph 13 (UK)	Fake reviews and fake social media influence
Government advisory	AA24-191A Russian State-Sponsored Use of Meliorator Software	Indicators and platform mitigations for AI-enabled persona farms

Practitioner checklist

Map your surfaces where engagement is counted or ranked (reviews, follows, likes, trending) and estimate inauthentic prevalence by sampled review.
Log registration, device, network and action events with identifiers that allow cluster analysis.
Apply risk-based sign-up friction and verification, and give new accounts limited weight until they build history.
Rate-limit engagement endpoints, including APIs, and detect automation frameworks and emulators.
Strengthen MFA and account recovery to cut the supply of hijacked accounts.
Discount suspect engagement from counts and rankings before removal, with an appeals route.
Act on whole clusters and track recidivism.
For marketing: audit creators before paying, write no-fake-engagement clauses into contracts and never commission undisclosed or fake reviews.
Check obligations under 16 CFR Part 465 and the UK DMCC Act with legal counsel.

Bot Farms and Fake Engagement Guide: How to Detect, Disrupt and Avoid Paying for Them