Costs
Part of Fraud and brand safety: set a suitability standard early
Influencer fraud screening mistakes: a score is not brand safety structure
A look at common mistakes in influencer fraud and brand safety screening, from trusting a score and misreading flags to paying on metrics partners can buy.
Screening goes wrong in two directions and both are expensive. Miss a manufactured audience and you have paid for nothing. Reject a good partner on a bad signal and you have lost them, told them why, and told everyone they talk to.
This page is about the errors in the screening process itself. What to look for in the first place is in fraud and brand safety.
What to take away
- A screening decision is a test with two error types, and almost every process is tuned as though only one of them existed.
- Rare things are rarewhen a test with a modest error rate flags a rare problem, most flags are wrong, and that is arithmetic rather than pessimism.
- Detection is always behind. Payment structure and contract terms remove more exposure than any check.
Mistakes about the evidence
Accepting a screenshot. An image of an analytics screen is an image. The check is a live view of the account, shared during a call, navigated by the person who owns it.
Reading a single number. A follower count, an interaction rate or a vendor score in isolation tells you very little. The signal is inconsistency between numbers, which requires more than one.
Trusting a third-party score with an undisclosed method. A number you cannot interrogate is a number you cannot defend when you decline someone on the strength of it.
Skipping the comments. Reading fifty comments takes five minutes and catches more than most tools, because purchased comments do not respond to anything specific in the post.
Confusing a clean audience with a suitable one. A real audience that has no reason to care about your product is not a fraud problem, it is a fit problem, and it belongs with influencer discovery.
Mistakes about probability
This is where most screening processes quietly fail.
Why most flags are false alarms
- Serious cases are uncommon in the pool
- Check is good but not perfect
- Many more clean candidates than serious ones
- Clean candidates produce more false alarms
- Genuine cases produce fewer true flags
- So a flag is not a verdict
Suppose serious cases are uncommon in your candidate pool, and your check is good but not perfect. Most candidates it flags will be false alarms. There are many more clean candidates, so they produce more false alarms than genuine cases produce true ones.
This does not argue against checking. It argues against treating a flag as a verdict. The reasoning is the base rate fallacy.
Two consequences follow.
Use a flag to start a conversation, not to end one. The conversation is cheap and it resolves most false alarms in one call.
Do not tell a candidate they failed a fraud check. Tell them what you saw and ask about it. The distinction matters because you will be wrong sometimes, and because you may be wrong about someone whose reputation you have just affected.
Mistakes about incentives
Paying on a metric that can be bought. If the fee depends on interaction, you have created a reason to purchase interaction, including for partners who otherwise would not. The general problem is a metric under pressure ceasing to measure anything, which is Goodhart's law.
Setting targets that only a purchase can hit. A partner who agreed to an ambitious number and is short of it near a deadline is being pushed toward exactly the market you are screening for.
Rewarding the internal team on campaign metrics. The people running the campaign then have the same incentive not to look too closely. Where you can, separate whoever screens from whoever reports.
Treating detection as the whole answer. Manufactured engagement is cheaper to buy than to detect, and the sellers adapt faster than the tools. Structure is what actually reduces exposure: pay in stages, tie payment to delivery and publication rather than to results a partner can manufacture, and keep the term short enough that a mistake is recoverable. The clause work is in contracts and disclosure.
Mistakes about brand safety
| Mistake | Why it happens | The fix |
|---|---|---|
| No written standard until a case forces one | Standards feel abstract until they are urgent | Write it before you need it |
| A standard written as a keyword list | Easy to operate, easy to fail | Describe the judgment, then give examples |
| Judging a partner on their audience's behavior | The comment section is not the creator | Assess what the creator does, and what they do about the rest |
| Reacting fast and publicly | Pressure to be seen to act | Establish the facts first; speed is worth less than accuracy |
| Demanding removal for taste | The takedown clause exists | Reserve removal for defined causes, and expect attention if you do not |
| Applying the standard to partners and not to your own channels | Nobody wrote it down as a general rule | One standard, applied to everyone |
The third row is the one that produces most of the unfairness. Every account with reach attracts comments its owner did not write and does not agree with. What is assessable is what the creator publishes and how they handle their own space.
Brand safety mistakes and fixes
Mistake
- No written standard
- Write it before you need it
- Keyword-list standard
- Describe judgment, give examples
- Judging audience behavior
- Assess what creator does
- Reacting fast and publicly
- Establish facts first
- Demanding removal for taste
- Reserve removal for defined causes
- Standard not applied internally
- One standard for everyone
The fix
- No written standard
- Keyword-list standard
- Judging audience behavior
- Reacting fast and publicly
- Demanding removal for taste
- Standard not applied internally
The mistake that undoes the rest
Running the screen after the deal is agreed. At that point the campaign has a date, someone has committed internally, and a flag creates a problem everybody wants to make go away. Screen before you negotiate, when declining costs nothing but a polite message.
And screen against a written standard rather than a mood, so that the decision is the same on a busy week as on a quiet one. What the standard is protecting is set in strategy and objectives, and how a weak result should be interpreted afterward is in measurement and ROI.
Bottom line
Screen before you negotiate on evidence you can interrogate rather than a score you cannot. Treat a flag as the start of a conversation, because when the problem is rare most flags are wrong.
Never pay on a metric a partner can purchase. Write the brand safety standard down before a specific case makes writing it look like a verdict.
Common questions
How much time should screening take?
Enough for a live look at the account, fifty comments and the growth chart. That is a short call, and it catches most of what matters.
What do you say to someone you are declining on these grounds?
As little as possible about your conclusion and as much as you like about your process being a fit assessment. You may be wrong, and a specific accusation is a serious thing.
Are detection tools worth paying for?
At volume, for triage, yes. As the basis for a decision, no.
Should past purchased engagement disqualify someone permanently?
That is a judgment. Accounts change hands and habits change. What matters is whether the audience is real now and whether the person can talk honestly about the history.
Who should own screening?
Someone whose performance is not measured by the campaign's results. Where that is not possible, at least write the standard down so the decision is not made under pressure.
