Four AIs said this
ad would kill it.
The fans killed the ad.
A Lynx integration in a Sidemen stadium video — 21 million subscribers, a can of deodorant written into the game itself. Textbook stuff. We gave the brief to Claude, GPT-5, Gemini and Grok and asked them to predict the reception. We asked our panel of 16,191 real Sidemen fans the same thing. Then the video dropped, and we read every single comment.
On paper, this is how you do it right.
No ad read. No cutaway. Mid-video quiz round: an assistant signals the correct answer by spraying a Lynx can on a “terribly dressed fan”, with the line that thanks to Lynx's new fragrance, even while they look their worst, they'll smell their best. The brand is woven into the game — exactly what every sponsorship playbook says to do.
Everyone saw the same brief. One saw it coming.
Same brief, same question, live web search on. The AIs reasoned like marketers: mechanic-tied integrations outperform ad reads, Sidemen fans reward creative sponsorships — 65 to 85% positive. Claude Sonnet wouldn't even give a number; it just praised the integration. Our panel was the only voice in the room below 50.
And that 33% is us being generous to the ad — it counts every “get dat bag boys” joke as a win for the brand. Count only the fans with a firm opinion and the real number is 6% positive.
We read all 3,681 comments. Every one.
No keyword filters, no sampling — every comment on the video was individually judged for whether it reacted to the sponsorship, and how. 200 did. Here's the room:
“Trying to work in the lynx ad by randomly spraying them was one of the dumbest ways to promote an ad — almost ruined the video.”
“i think i got a stroke after hearing lynx 300 times :)”
“So tired of people making the ad a part of the content. When it’s incorporated it makes everything feel fake.”
“A 1hr 13min advert for lynx 😂”
They reasoned like marketers. Our panel reasoned like fans.
The AIs weren't stupid — they applied sponsorship theory correctly. Grok even spelled it out: “creative and mechanic-tied… Sidemen viewers reward this.” True in general. Fatal here. Our panel is built from what this fanbase has actually said, and it flagged the exact things the real comments would say a day later — without ever seeing them:
“spraying some random ‘terribly dressed’ dude with lynx just to give away an answer? come on, that’s so obviously for the brand deal it’s almost painful.”
“making some random ‘fan’ look bad just to spray deodorant on them? it’s trying too hard to be ‘clever’ when it’s just an ad. feels kinda fake, y’know?”
“that ‘smell their best’ line with the terribly dressed fan is just so obviously scripted and not funny. takes you right out of it.”
Scripted-feeling bits, mocking a fan for a brand deal, tolerance for sponsors but not for being taken out of the video — three objections the panel raised blind, and the three loudest themes in the real comment section. That isn't on any web page for an AI to search. It lives in the fans, and we're the only ones holding them.
“I just want you to know this has made me less likely to buy lynx.”
“This makes me want to purposely buy a different brand to lynx.”
“Never want to see a can of lynx again.”
Real comments, real reach, a deal that cost real money. Running the same brief past our panel takes minutes and costs about as much as a coffee — and it would have come back 42%, red flag, here's why, while the contract was still a draft.
Run the ad past the audience before you run the ad.
Every AI in the world read the same brief and priced it off theory. The only system that got the direction right was the one built out of the audience itself. Concepts, creative, integrations, announcements — test them on your fans before your fans test them for you, in public, under your biggest video of the year.
The fans already know how it lands. We're just the only ones who can ask them in advance.
Rival predictions via OpenRouter with live web search, identical brief, same day. Reality: all 3,681 top-level comments individually stance-judged (no keyword pre-filter) — 200 reacted to the sponsorship: 8 positive, 57 neutral/banter, 135 negative. 33% treats banter as favourable; firm-opinion-only is 6%.
Receipts: data/validation/sidemen_live/lynx_prediction.json · lynx_judged_3way.json.