Stuck at 100,000 Views for 100 Days. 1 Thumbnail Later, It Hit 408,000.
A spreadsheet formula found the break point by hand. Reflare claims it can find it for 600 videos at once.
A video doesn't die all at once. It goes flat first, and somebody has to notice.
I'm looking at a spreadsheet with 1 column that matters, ELO, sorted top to bottom. It's not fancy math, just videos ranked against each other by views, and read as a curve it does 1 useful thing: it shows exactly where a channel's back catalog stops moving. There's a bend in that curve, and right after it, the numbers go to almost nothing.
The question worth asking is what happens when a machine reads that same bend instead of a person. Does it catch it the way an eye does, or does it just process numbers.
The Spreadsheet Before the App
Before any of this got automated, Hardisk found that bend by scrolling 600 rows on a Sheet and trusting his own read of the drop-off. No XP bar, no quest marker, just a gut call on a scatter plot. His channel, 104,000 subscribers, hundreds of videos going back to 2023. Top of the sheet, a documentary about the Olympics sits at 9,816.24. Near the middle, a video called La face cachée de l'ASMR sits at 6,823.95, 1 entry among a long stretch that's slowly losing altitude. The plot next to the sheet shows it clean: a tight cluster of dots up near 8,000 to 10,000, a steep fall, then a second cluster that just sits flat near zero. That gap between the 2 clusters is the whole diagnosis. Above it, a video is worth touching. Below it, you're wasting a thumbnail.
I've built the same kind of sheet before, for cron job failures, ranking scripts by how often they silently died. Got obsessed with the sorting for about a week, then forgot the sheet existed for 4 months. It's probably still running somewhere, ranking jobs nobody's looked at since (building the cheap version before paying for automation), the same way most creators try the manual version once, then quietly stop.
Reading the bend by hand doesn't scale past a few hundred rows. Hardisk kept doing it anyway, for months, before he turned the sheet into a product.
The Bend the Dashboard Confirmed
That ASMR documentary is the founder story, and the proof is sitting right there in the YouTube Studio dashboard. Late 2023, before Reflare existed as a company, the video sat stuck at 100,000 views for 100 straight days. Same 24 minutes of audio, no reshoot, just a new thumbnail. A tooltip on the growth curve marks day 358 since publication, 360,427 views at that point, the bend already visible on the graph, sharp, right where the flat plateau ends. Total views today: 408,693. Still climbing.
That's the moment that convinced Hardisk the game was real, and it's the same video the dashboard is showing me now, further along, still adding views months after the fact.
What Reflare Actually Does
On a Lego video, Reflare's own attribution model credits +23,000 views to 1 thumbnail swap. On a Boeing video, +6,000. A test still running when I checked it: the original packaging held a 2% click-through rate (CTR) at 187 impressions, 9 views. The new variant is running 6.5% CTR at 232 impressions, 31 views, and hasn't locked yet.
Those numbers come from a demo on the founder's own catalog, not an outside audit. Worth saying plainly before going further.
The product runs 3 loops once you connect a channel over OAuth, and they read like a boss fight with 3 phases, except nobody respawns, the video just sits at zero. Scan ranks videos by retention, CTR, and impressions over the last 7 days, the same signal that spreadsheet ELO score was trying to approximate by hand. Generate studies the channel's existing art direction (faces, fonts, color choices) and proposes new thumbnail variants, 2 clicks, no prompt box, nothing goes live without approval. Test rotates the variants on the real video, 3 days if it's still getting impressions, 7 if it's quiet, and locks a winner once the gap is statistically real. The original stays archived and comes back if the test gets cancelled.
That's a lot of catalog to hand over (handing catalog monitoring to a tool instead of you), and somewhere between a genuinely useful spreadsheet macro and a system quietly making calls a human used to make, this stops being just a thumbnail tool.
\
If your own catalog has a plateau sitting somewhere in the last year, Reflare's channel audit is free to run first, and the code WELCOME knocks 10% off the first month if you go past the audit.
The Order Hardisk Gave Me
I asked the obvious dumb question: 216 videos flagged in 1 week, where do you actually start.
- Open the attention filter first, sort by impressions over the last 7 days. Videos still getting shown are the ones that can still move.
- Launch several tests at once, the first 5 to 10 on that sorted list, not 1 a week.
- Keep the defaults when impressions are low. 7 days, 1 variant. The app already recommends conservative settings, and it warns you if you push for more variants than the data can support.
Hardisk's own math on a 600-video catalog: roughly 5% of a test wave "sticks." Some early customers churned after 2 weeks expecting 20,000 extra views immediately, while their tests were still running and the thumbnail was already sitting at plus 1 or 2 points of CTR. Not nothing. Just not a slot machine.
That patience requirement is the easy part to sell and the hard part to actually practice.
What It Won't Touch, and What It Costs
No titles, ever. No thumbnail on a video that just went live. Hardisk cites a Seoul University study putting roughly 70% of click intent on the thumbnail alone, and he refuses to mix title changes into the same test, too much statistical noise to tell a client which lever moved the needle. You can disagree with the 70. The product already made its choice and isn't touching titles until it's confident enough to say exactly what moved.
Pricing sits in 3 tiers. Creator at 49 euros a month covers 1 channel and 60 AI thumbnails. Pro at 99 covers up to 3 channels and 200 thumbnails. Studio at 299 covers up to 10 channels and 600. Compare that to a freelance designer at 50 to 150 euros per thumbnail, and the math works if the tool revives a handful of videos a month. On a 10-video channel, it doesn't, there isn't enough catalog to revive.
If the catalog is big enough for the math to work, start with the free audit and the code WELCOME for 10% off. If it's a 10-video channel, save the 49 euros.
What a Script Doesn't See
Here's the part that nagged at me reading that plot next to the spreadsheet: the person who built it wasn't just finding the bend, he was also noticing everything around it. Which videos clustered near the bend but for the wrong reason, a bad title rather than a stale thumbnail. Which ones sat in the dead zone because the content itself never had legs, no amount of repackaging fixes that. A ranked list tells you where the cliff is. It doesn't tell you why a particular video is standing on the edge of it, and that context lived in Hardisk's head, not in the ELO formula.
I think the CTR-first approach the tool is built around gets that trade-off mostly right, though I'm honestly not sure how it holds up once the catalog gets big enough that nobody's eyes have touched most of it in years.
Reflare is explicit about the danger it can't fix: a thumbnail that pulls a higher click-through rate but tanks average view duration hurts the video twice, gets it shown, then gets it dropped from recommendations. YouTube is a game everyone plays for the same 12 slots on somebody's homepage, 1 click. A packaging change that pulls the wrong audience gets the video seen and then buried, and the tool has no way to know that happened until the retention numbers come back and confirm it. It executes the test. It never asks why the pattern is there in the first place, the same way HAL never asked why the mission mattered, just whether the numbers lined up.
The algorithm finds the bend. It has no idea why the video is standing on it.
The person who built the spreadsheet still checks it sometimes. Not because the automation needs supervising. Because he can't quite stop reading the curve himself.
Sources
- Reflare.io
This post may contain affiliate links — if you click them, I might earn a small commission, costs you nothing, and helps me keep shipping quality articles every day for your reading pleasure.
The bend in your video performance curve tells you exactly which videos are worth touching, but spotting it by hand doesn't scale past a few hundred rows. The demo vs product checklist in the welcome kit shows you the infrastructure difference between a spreadsheet you check once and a system that actually catches the signal.