Brand safety

Influencer Fraud Detection: A Practical Guide

How to read inflated reach and fake engagement signals, why a single signal never decides on its own, and how to document a finding in a report.

TubeDetect editorial team5 min readPublished: 10/09/2026Updated: 10/09/2026

This topic is also covered in Turkish: Turkish version of this guide.

Fraud is a probability question, not an accusation

Audience checks are often reduced to one question: is this account fake? That question has no single answer, because an account can carry purchased followers without being worthless as a whole. A clean looking account can still show artificial engagement during a campaign period.

A more useful question is whether the results an account produces are consistent with the audience it claims. An inconsistency is enough to change the campaign conditions or to ask for more verification, and it keeps the decision based on observation rather than on suspicion.

The language of a finding matters too. Fake is a legal claim that needs proof, while inconsistency can be supported with observable data. Reports therefore state observations and dates instead of shares or percentages: what was seen, over which period and in which source.

Four signal groups

No single signal decides anything; signals gain meaning when read together. These four groups cover most manual reviews and each can be checked in minutes. Write the date next to every signal, because an undated finding cannot be verified later.

Use the account own history as the reference. Low like counts are not necessarily a risk if they match the account normal line. Risk appears when the account leaves its own pattern: followers rising while engagement falls, comments drifting away from the topic, or audience geography not matching the campaign market.

Growth curve: does the follower count jump independently of content performance?
Engagement distribution: do comments and likes arrive in regular blocks, or do they change with content type?
Comment quality: are comments about the topic, or do the same phrases repeat?
Audience fit: do audience geography and language match the target market?

Sources: TikTok: Community guidelines · Meta: Community standards

The limit of automated tools

Fraud detection tools produce a score, and a score used as the only decision tool misleads. Private accounts, newly growing creators and niche audiences can fall outside the model entirely, so an account can be rejected for a low score even though it fits the campaign.

A two step verification is more reliable: collect observable evidence first, then compare the tool output with it. Where the two disagree, the report states the disagreement instead of smoothing it over. Hiding a contradiction weakens the basis of the decision.

A second risk of using tool output alone is scale. When twenty candidates are ranked in one table, teams drift towards the highest score and set the campaign job aside. A high score indicates low risk, not suitability, and the two are not the same thing.

A worked example: two inconsistency findings

A fictional example: a snack brand reviews two candidates. The first account shows two sudden follower jumps in the last 90 days, and in both weeks content performance stayed below its usual level, while comments remain relevant and topic focused. The finding reads: follower growth jumped twice independently of content performance, with comment quality consistent with the topic.

The second account grows steadily, but like counts cluster around the same narrow range across posts and many comments are short phrases unrelated to the content. Both findings are recorded with dates and links to the reviewed content. At no point is the word fake used.

The decision follows the effect on the campaign: the first candidate is approved conditionally with a performance condition after the first piece and an audience geography check, while the second is planned with a format shift and a stricter reporting window. Neither candidate is dropped, and the risk becomes visible.

Suggested image alt text: Time series chart of follower growth and engagement for two candidates

Add a 90 day chart of follower growth and engagement for both candidates, with sudden jumps marked and dated. It should show the difference between a jump and a steady line.

Write the finding down

Documenting a concern is the cheapest way to act without risking the campaign. A finding written with an observation, a date and a source can be discussed with the creator and used to restructure an offer. A concern shared verbally can neither be verified nor discussed with the creator.

Low engagement is an opinion. In the last 90 days comments ranged between 3 and 8 on twelve posts, while follower growth jumped twice in the same period is a finding: arguable and verifiable. That difference decides the value of the whole document.

Sharing findings with the creator also strengthens the relationship. When a team shows the data and verifies it together, the creator reads the process as professional and transparency increases in the next collaboration.

Observation: what was seen, with which numbers?
Date: over which period and in which pieces of content?
Source: platform data, screenshot or creator statement?
Effect: how does this finding change the campaign conditions?

Conclusion: turn suspicion into a condition

Audience verification is not about passing a verdict, it is about making the risk of a decision visible. With dated signals, a comparison between tool output and evidence, and a condition attached to every finding, the campaign stays both safe and runnable.

To review candidates with one template, create a Brand Profile and add the account. TubeDetect reports growth, comment signals and commercial history findings with their dates and sources, and states the limits of the review. Credit packs are listed on the pricing page.

Frequently asked questions

Can a fake follower share be measured exactly?

No. Platforms do not share full audience data, so any number is an estimate. Reports therefore show observable inconsistencies and their date range instead of a share, which keeps every finding verifiable.

Is low engagement a sign of fraud?

Not on its own. Content type, account size and audience structure all shape engagement. What matters is whether engagement is consistent with the account history and the audience it claims to reach.

Should a campaign be cancelled when a finding appears?

Not every finding requires cancellation. Changing the conditions, asking for extra verification or narrowing the scope is often the more proportionate step, and cancellation stays reserved for findings that conflict with brand values.

How is comment quality assessed?

By looking at topical relevance, repeating patterns and the nature of the questions. Technical questions about usage indicate genuine interest, while single word repetitions raise the probability of artificial engagement.

What if the tool score contradicts the evidence?

The contradiction is written into the report and the decision becomes conditional. A score neither rejects nor approves a candidate on its own; both outputs are read together before the call is made.