CreatorScore▸ Key findings
Six numbers from the data
- 55% of the findings that survived review sit on posts more than a year old; 25% are more than three years old. Only 19% come from the last 90 days.
- The pattern holds for the serious ones: 49% of high and critical findings are over a year old.
- 59% of kept findings were in what the creator said out loud. Captions and written posts held 26%; on-screen text and visuals 9%.
- 44% of raw automated detections were dismissed on a second review of the evidence. In the categories brands care about most, the rate is higher: 74% of hate-speech detections and 96% of written-violence detections did not hold up.
- In 41% of the checks that found something, the findings were spread across more than one platform. In 29% they were on three or more.
- 58% of checks surfaced at least one high or critical finding after review. 10% came back with nothing to review.
▸ 01
How old the real findings are
Most vetting advice says to review a creator’s recent posts. We measured where the findings actually were. For the 6,017 kept findings whose post date is on record, the median finding was 444 days old when the check ran, and the oldest was 16 years old.
- 55%of kept findings on posts over a year old
- 25%on posts over three years old
- 19%from the last 90 days
Kept findings by age of the post
High and critical findings are no more recent than the rest: 49% of the 467 are over a year old and 18% are from the last 90 days. A review of the last thirty posts would have missed four in five.
▸ 02
Where in the content the findings live
A caption is what a creator writes under a video. A transcript is what they say in it. The two are not the same content, and the findings are mostly in the second.
- 59%of kept findings in spoken words
- 26%in captions and written posts
- 9%in on-screen text and visuals
Kept findings by source
Reading captions is not reading the content. A process that does not transcribe video misses most of what a full read finds.
▸ 03
How often the detector is wrong
Every detection is re-read by a second pass that sees the evidence behind it: the quote, the frame, the surrounding text. Of 10,788 raw detections, 4,728 did not meet the category’s definition on that second read and were dismissed.
- 44%of raw detections dismissed on review
- 75%of on-screen text and visual detections dismissed
- 28%of spoken-word detections dismissed
Dismissal rate by source
Dismissal rate by category
Categories with at least 100 raw detections. Highest rate first.
The categories that matter most are the ones a detector gets wrong most. A keyword list reads a slur quoted in a news clip, a lyric or a joke about oneself as hate speech; a classifier reads a workout caption as violence. Presence signals like alcohol and profanity are close to certain. A detection count is not a risk count, and a tool that shows one without the review step behind it is showing the wrong number.
▸ 04
One account is not the creator
Brands usually vet the account they are about to pay for. Of the checks that found something, 41% had findings on more than one platform and 29% on three or more.
- 41%of checks with findings on two or more platforms
- 29%on three or more
- 5.6years of history read in the median check
Kept findings by platform
Platform counts reflect where creators post most and how far back each platform’s history reaches, not that one platform is riskier than another.
▸ 05
What survives review
Of the 6,060 findings kept after the second pass, most are low severity: presence signals a brand wants to see and decide on, not dealbreakers. The serious tail is small and real.
- 58%of checks with at least one high or critical finding
- 8%of kept findings high or critical
- 10%of checks with nothing to review
Kept findings by severity
▸ Method
Methodology
- Sample: every CreatorScore Social Background Check completed between 24 June and 5 November 2026. A check reads the full public history of the accounts a brand submitted for one creator, on up to 11 platforms, in the original language. The median check read 528 posts spanning 5.6 years.
- Detections: 10,788 raw detections. A detection is one post, caption, transcript excerpt, on-screen text or frame that a detector (keyword list, classifier or vision model) matched to a brand-safety category.
- Second review: every detection is re-read by a second automated pass that sees the evidence behind it (the quote, the frame, the surrounding text) and keeps, downgrades or dismisses it against the category's definition. 4,728 were dismissed. Kept means kept or downgraded. Brand-side dismissals after delivery are not counted.
- Age: the check's completion date minus the post's publish date, for the 6,017 kept findings whose post date is on record (99% of kept findings).
- Source: where the matched text or image came from. Spoken words are video transcripts; captions and written posts cover captions on image and video posts and text-only posts on X, Threads, LinkedIn and Facebook; on-screen text and visuals cover text read from frames and what the frame shows; several at once means the same post matched in more than one of those.
- Severity, category and platform are the labels on the kept finding. Platform counts reflect where creators post most and how far back each platform's history reaches, not that one platform is riskier.
- No creator, brand or agency is identified. The sample is CreatorScore's own production data, not a random sample of all creators: brands chose which creators to check, which skews toward creators already in commercial conversations. The number of checks is available to press and researchers on request.
What a vetting process should cover, with these figures applied, is in What is influencer vetting. The tool that produced the data is described at /influencer-vetting. The broader scoring data is in State of Creator Trust 2026.
Cite this report
Free to quote and republish the charts with attribution and a link (CC BY 4.0).CreatorScore, "What Influencer Vetting Actually Finds" (November 2026), https://creatorscore.io/research/vetting-findings-2026