celebrity · vetting
How to Vet a Celebrity Endorser, Run Live on Celsius
Everyone sells a three-layer celebrity vetting process. We ran ours against nine names for Celsius to see what it caught, and the automated screen was wrong about two of the five people it flagged.
The worst-scoring name on our screen, at 0.11% engagement against 1.1 million followers, turned out to be completely fine. The number was our own bug.
Every agency sells the same three-layer celebrity vetting process, and it always sounds airtight on a slide. Screen the follower base for fake accounts, audit the audience by hand, then scan for category conflicts.
So we ran it, properly, against one brand, and wrote down every call as we made it.
The brand is Celsius, the energy drink. The candidates are nine names, eight of them people who have never worked with Celsius, plus Noah Lyles, who appeared in a Celsius ad campaign in April 2024 according to Marketing Brew. Lyles is in the group as a benchmark, because it is worth knowing whether the person a brand already picked would survive its own screen.
Where the nine names came from
The eight candidates come out of our own celebrity map, which holds 351 announced celebrity and creator drinks deals since 2021, each one carrying a source link. We pulled the people whose profile fits an energy or functional drink, which in this category means athletes almost entirely.
The engagement numbers come from our Instagram records, where all nine accounts are on file with follower counts, post history and per-post likes and comments.
Nothing here is a survey or an estimate, every number below is a stored row we can point at.
Layer one, the automated screen
We set the thresholds before looking at anybody, so the screen could not be tuned to produce a tidy answer. Engagement below 1 percent gets flagged as an audience that has stopped responding, above 10 percent gets flagged as implausible at this size, and more than 200 likes per comment gets flagged because likes are the cheaper thing to buy.
Then we ran it on the stored numbers, which is exactly what a tool would see.
| Person | Followers | Stored engagement | Likes per comment | Screen says |
|---|---|---|---|---|
| Allyson Felix | 1,138,464 | 0.11% | 6 | Flag |
| Chris Bosh | 2,517,607 | 0.55% | 32 | Flag |
| Christen Press | 728,809 | 2.14% | 431 | Flag |
| DeAndre Hopkins | 1,673,329 | 19.35% | 36 | Flag |
| Jalen Green | 2,287,765 | no data | no data | Flag |
| Baron Davis | 347,141 | 1.31% | 51 | Pass |
| Emma Raducanu | 2,929,035 | 3.49% | 131 | Pass |
| Gabby Thomas | 1,197,106 | 1.61% | 125 | Pass |
| Noah Lyles | 1,605,798 | 2.00% | 173 | Pass |
Five flags out of nine. If we stopped here, which is what a lot of vetting actually is, we would have told a client that Allyson Felix has a dead audience and DeAndre Hopkins looks manipulated.
Both of those calls were wrong.
Layer two, opening the posts behind every flag
The manual layer is not a second opinion on the same number, it is going and looking at the posts the number was built from.
Allyson Felix, flagged at 0.11 percent, is fine. Of the 12 posts we hold for her, 9 hide their like count. Instagram reports a hidden like count as -1, and our pipeline was averaging that -1 in as though it were a number of likes. That is what produced 0.11 percent. It is a fault in our own maths, and it would have cost a client the best name on the list.
Her comments were visible the whole time, running 81, 117, 740, 975 and 5,395 across recent posts, which is a healthy spread for 1.1 million followers.
DeAndre Hopkins, flagged at 19.35 percent, is a sample-size illusion. His account keeps only 3 posts. An average built on 3 posts, two of them from 2019 and 2020 with 476,397 and 153,478 likes, is not an engagement rate, it is a leftover. The screen was right to stop on him and completely wrong about why, and the note in the file would have said something false.
Chris Bosh, flagged at 0.55 percent, is half right. His likes genuinely are low, 11,809 as a median against 2.5 million followers. His comments are the second strongest in the whole group. His audience talks, it just does not press like.
Christen Press is the one flag that survived. Median 16,288 likes against 38 comments. That gap is not an artifact and it did not move when we recomputed it.
Jalen Green cannot be scored. There is no engagement data on file for him at all, so the honest output is unknown, and unknown is a different thing from a pass or a fail.
Because likes can be hidden and means can be pulled by one viral post, the number worth comparing across all nine is the median comment rate.
| Person | Median likes | Median comments | Comment rate |
|---|---|---|---|
| Emma Raducanu | 102,620 | 638 | 0.0218% |
| Gabby Thomas | 24,594 | 207 | 0.0173% |
| Chris Bosh | 11,809 | 290 | 0.0115% |
| Allyson Felix | 6,535 | 128 | 0.0112% |
| Noah Lyles | 29,612 | 124 | 0.0077% |
| Baron Davis | 335 | 21 | 0.0060% |
| Christen Press | 16,288 | 38 | 0.0051% |
Noah Lyles, the one person here Celsius has already run a campaign with, has a lower comment rate than four candidates they have not used.
That ordering is the useful output of the whole exercise, and no automated pass produced it. This is the part brands hand to us, because the screen is cheap and the reading is where the hours go, and a number that nobody opened is how a good candidate gets thrown out.
Layer three, the conflict scan that leaves nobody standing
Now the category conflict scan, which is meant to check the person is not already promoting a competitor.
We ran all eight candidates against every deal in the map. All eight have a prior beverage deal. Applied as written, the rule no competing products returns zero usable names, which means the rule as written is useless.
It only works once you accept it is three separate questions wearing one coat.
Shelf substitute, meaning the same fridge. Emma Raducanu was announced as Evian's global brand ambassador in December 2021, and Olipop announced Gabby Thomas in August 2021. Both sit next to an energy drink at the till.
Positioning contradiction, meaning it points the audience the other way. Three candidates came back attached to Grey Goose, which for a brand sold on training and performance is a tone problem rather than a shelf problem.
Except it is not even that, and this is where the scan failed in the opposite direction from layer one. Our own record labels the Grey Goose deal an endorsement at high confidence. Opening the source shows an interview series called In The Moment, hosted by the broadcaster Cari Champion and made with SpringHill. Chris Bosh, DeAndre Hopkins and Christen Press sat for interviews. That is a content appearance, and calling it a liquor endorsement over-flagged three of our eight people.
Live exclusivity is the only one of the three that is an actual legal blocker, and none of the eight tripped it, because every deal on record predates 2024.
Allyson Felix is the cleanest of the group on this layer. Her prior deal is Pure Leaf iced tea, from a March 2022 announcement tied to a grant program with the SeekHer Foundation, which is a long way from an energy drink in both product and shape.
Authenticity, and the part we cannot prove
The framework asks whether the person had an honest link to the category before anybody paid them. For eight competing athletes, the answer is yes and it is not interesting, since training and hydration are their job.
What we cannot prove from our data is the harder half of that question, which is whether they used this kind of product before the cheque. We hold announced deals, not private habits, and a brand that tells you it verified sincerity is selling you a feeling. What we can check is whether the public record contradicts the story, and here it does not.
The reputation scan, including whether they are alive
Four names were blocked out of the wider pool before this shortlist existed. Conor McGregor for a civil assault verdict and a competing whiskey he owns, Floyd Mayweather for repeated domestic violence convictions, and Lance Armstrong for the doping case that cost him seven titles.
The fourth is the one that matters most. Angus Cloud ranked at the top of three of our candidate pools until a person wrote a line taking him out, because he died in 2023, and nothing in the data knew that.
A scoring model will happily rank somebody who cannot sign a contract. That single line is the whole argument for keeping a human in the loop, and it came out of our own file rather than a cautionary tale from a conference.
Who is actually left for Celsius
Emma Raducanu and Gabby Thomas come through the strongest, on the metric that survives hidden likes and viral outliers. Both carry a shelf-substitute conflict from 2021 that would need checking for whether it still runs.
Allyson Felix is the interesting one, because she is cleanest on conflicts, healthy on comments, and she is the person the automated screen would have thrown out first.
Christen Press is the one honest fail. Jalen Green is unknown until somebody collects data on him.
And the benchmark answer is the uncomfortable one, since the person Celsius already ran with sits fifth of seven on comment depth.
What we could not check
Our deals map covers drinks only, from 2021 onward, so a candidate could hold nine deals in other categories and still look clean here. That means our overexposure reading is partial, and we are saying so rather than presenting it as a full sweep.
The announcement year is not proof a deal ended, so every 2021 conflict above needs a live check before anyone signs.
This was a ratio screen on engagement, not follower forensics. We do not hold audited bot detection, so it catches dead and lopsided audiences and would not catch a well-run purchased following. If that distinction matters to you, why fake follower counts still fool most brand teams covers what the deeper check involves.
Our engagement snapshots run 12 posts per account, and 3 for Hopkins.
Where We Come In
The pattern across all three layers is the same, which is that the automated part is quick and confidently wrong in both directions, and the value sits in somebody opening the file. We do that reading for brands, and we would rather tell you a candidate is unknown than hand you a score that looks certain.
We also fixed a fault in our own pipeline because of this run, which is the sort of thing you only find by checking your own work in public.
If you are weighing a famous name right now, it is worth knowing what the fee actually buys before the vetting question even comes up, and the $42,000 versus $2,000 per post maths is the place to start. If the plan involves ownership rather than a fee, only 44 of 351 drinks deals gave the person a piece of the company, and the ones who got paid at the exit all held equity.
Send us the shortlist and we will run this against it, and tell you which names we cannot score. Speak with us when you want the reading done properly.
Open the file, then decide.
Reading loop
Frequently asked
What does a three-layer celebrity vetting process actually catch?
Less than the name suggests. Across nine candidates we screened for Celsius, the automated engagement pass flagged five people, and only one of those five flags survived a manual check of the underlying posts. One flag was caused by a fault in our own stored numbers, one person had no data at all, and one was flagged on likes while having the second strongest comment depth in the group.
Why do engagement rates on celebrity accounts often look wrong?
Because Instagram lets an account hide its like count, and it reports that as -1. If the pipeline storing the number treats -1 as a value and averages it in, the account's engagement rate collapses. Allyson Felix hides likes on 9 of the 12 posts we hold, which is what turned her stored rate into 0.11 percent. Comment volume is the more dependable signal, since it survives hidden likes.
How do you check whether a celebrity endorses a competing product?
Split the question in three. A shelf substitute sits in the same fridge, like Evian or Olipop against an energy drink. A positioning contradiction points the audience the other way, like an alcohol appearance for a fitness brand. Live exclusivity is the only one that is a legal blocker. Of the eight candidates we scanned, all eight tripped one of the first two, and none tripped the third.
Can automated vetting replace a person reading the file?
No, and the clearest reason sits in our own records. Angus Cloud ranked at the top of three of our candidate pools until somebody wrote a line excluding him, because he died in 2023 and nothing in the data knew that. A scoring model will rank a person who cannot sign a contract.