Discounts for a fan who's never bought are a classic argument. One side says a first sale is worth almost any price. The other says you've just taught him to wait. Both sound right and everyone in chatting has a story that proves their side.
So we stopped debating and split it. Every guy goes into one of two groups. One keeps the discounts as they were. The other runs stricter rules: guys who've never bought get none, whales wait longer for one, and regular buyers stay the same. We count money per group and per type of guy.
We're not quoting a result here. A few days on one account proves nothing. The point is that a question we could have argued about for months now has a way to get answered.
Why guessing is so bad at this
Chatting advice comes from memory, and memory is a terrible analyst. In our first 8 days we sent 645 paid offers and 64 were bought. You'd remember a few of the 64. Nobody remembers the rest.
Then there's the whale problem. One fan paid about 72% of the money in those 8 days. Whatever we happened to be doing that week would look like genius if you looked at the total. A good story and a big sale in the same week feels like proof. It's usually just the same guy buying.
More than once a test has gone against what we'd have bet on. That's the whole value of it. If tests only confirmed what you already believed, you wouldn't need them.
How tests turn into money
Every fan passes through the same few moments. The first paid offer. The first no. The first time he goes quiet. Improve one of those and it applies to every guy who comes through, today and next year. A clever trick on one chat earns once. A better default keeps earning.
A winner costs nothing to keep. Once you've rolled it out, it's just how she works now. No extra staff and no extra content.
A loser only hurts one group. Roll a bad idea out to everyone and you take the full hit, and you may not even notice. In a test it stays with one group, for a limited time, and you get to see it.
Why AI makes this easier
Try testing an idea on a human team. You brief the chatters, half of them forget, one does it his own way and the night shift never heard about it. When the numbers come in, you don't know if you've measured the idea or the chatter.
An AI does the same thing every time, for every guy in the group, at any hour. So the difference between the groups is the thing you're testing. With 290k+ messages and 18k+ conversations handled so far, across social and Fanvue, that consistency is what makes our tests readable.
What we test on our own account
- How high the first paid offer starts
- Where the price starts on his next script
- The price when a buyer comes back after a break
- Discount rules, and how big a cut should be
- The gap between a free tease and the first PPV
- Follow-ups written by the AI against her pre-written lines
- The opening message to a new follower
- Whether she thanks him after a purchase, and when
- Answers to objections in social chats
What we don't test
Some things are rules. The 18+ stop. Never showing a fan a price lower than he already paid. Backing off a guy with real money worries. We wouldn't run those as experiments, whatever one group might earn.
Be honest about the limits
We're one account with a limited number of buyers. Plenty of our tests will never reach anything a statistician would sign. We treat most results as strong hints, act on the clear ones and keep watching after the rollout. That still beats a guess, because you know how much you know.