You help creators sell pay-per-view (PPV) content through one-to-one conversations with fans. AI is taking on work previously handled by human chatters.
I believe in building AI-first. But the dashboard should test that belief, not celebrate it.
The whole company needs to answer three questions:
| Team | Daily question | Decision it supports |
|---|---|---|
| Product | For comparable fans, does AI generate more PPV revenue than humans—and where does it still need help? | What should we fix or test next? |
| Business | Are we growing tomorrow’s revenue, or just selling to today’s biggest spenders? | Which fan relationships should we develop or recover? |
| Marketing | Which campaigns bring fans who keep buying and become more valuable? | Where should we invest the next acquisition budget? |
Build one view around each question. Put the answer first, the evidence underneath, and the next action last. Keep definitions and fan-level details one click deeper.
1. Product: does AI sell better to comparable fans?
The main number: net PPV revenue per originally assigned fan.
For each group, divide PPV purchases minus refunds by every fan originally assigned to that group who has completed the measurement window. Include zero buyers and inactive fans. Compare the same elapsed period: first 30 days against first 30 days, not an older human cohort against a younger AI cohort.
What the view should show
| Measurement | Why it belongs here |
|---|---|
| AI versus human net PPV revenue per assigned fan | Measures the revenue outcome, including fans who never buy |
| Mature fan counts and fans still maturing | Shows how much evidence is available |
| Frozen starting spending tier | Stops differences in fan value from masquerading as better chatting |
| Actual AI availability | Separates access problems from selling performance |
| Human takeovers and affected fans | Identifies work the product still cannot complete independently |
Where possible, randomly assign eligible fans within each creator, balancing pre-assignment spending tier and relationship age. Hold the offer and pricing rules comparable when testing chatter performance. If the test includes AI’s longer availability, say that you are comparing the complete operating models.
Freeze the starting tier before assignment. A minnow who graduates to whale stays in the original minnow experiment group. A later human takeover stays in the original AI-assigned outcome. Otherwise, the successful fans and the difficult work disappear from the group being evaluated.
Fans excluded before assignment are outside the eligible experiment. Fans excluded afterward remain in their original assigned result. Record both cases explicitly.
Across creators and starting tiers, compare like with like. Show results within those groups or standardize both arms to the same starting mix. Disclose groups that lack a counterpart. Statistical adjustment does not turn an observational comparison into a randomized experiment.
Keep attribution separate from the experiment
The experiment asks which assigned approach produces better outcomes. Transaction attribution asks who did the selling.
Put each sale in exactly one category: AI only, human only, AI + human, or unattributed. Link the purchase to its offer and conversation history. Count shared revenue once; do not invent a 50/50 split.
If AI sends an offer on Monday and the fan buys after AI is disabled on Tuesday, Tuesday’s switch must not rewrite Monday’s contribution. If a human takes over a valuable AI-assigned fan, preserve the assignment while attributing the later human work accurately.
Show creator-level AI-on minutes separately from per-fan exclusions. “AI on 100%” can coexist with the highest-value fans being excluded. Drill into exclusion reasons, affected fans, observed revenue and dated switch intervals. That revenue is not an estimate of what AI would have earned.
Track explicit takeover events. Shared selling alone cannot tell you whether a human rescued AI, reviewed an offer, or intentionally collaborated.
Next action: inspect the recurring takeover or availability problem, assign one fix, and test it on comparable fans. Do not declare AI the winner from a daily revenue fluctuation. Set duration, sample requirements and uncertainty reporting before the experiment starts.
2. Business: are we growing tomorrow’s revenue?
The main comparison: newer acquisition cohorts at the same age.
A rise in total revenue may come from a handful of established whales. That does not tell you whether the next generation of fans is developing.
Follow fans from acquisition at day 30, 60, 90 and 180. Keep each original cohort intact, including fans who never buy or stop buying.
What the view should show
| Measurement | Question it answers |
|---|---|
| Net PPV revenue per original fan at the same milestone | Are newer cohorts becoming more valuable? |
| Repeat-purchase rate | Are fans returning to buy? |
| Graduation into higher spending tiers | Are more relationships developing? |
| Regression in recent spending tiers | Are valuable relationships weakening? |
| Revenue share from the top 1% of fans | Does one small group explain the apparent improvement? |
Define repeat buying precisely. For example: fans with at least two paid purchases divided by all original fans. Disclose whether refunded purchases count, and apply that rule consistently.
Show graduation as a count and a rate: “20 of 200 starting minnows moved up,” not just “20 graduates.” Record how long it takes to reach a tier; fans who have not reached it are still being observed, not silently removed.
Compare AI and human original assignment within acquisition cohorts where possible. Also show the overall business trend. Differences between months can reflect acquisition sources, offers and starting mix—not just the chatter.
Recent status is different from accumulated value
Keep these three concepts separate:
| Grouping | Definition |
|---|---|
| Current tier | Net PPV spending in the trailing 30 days |
| Starting tier | Spending tier frozen immediately before assignment |
| Acquisition cohort | Fans who joined during the same stated period |
Track lifetime spending separately. A fan can have €2,000 in lifetime purchases and no purchase in the last 30 days. That relationship has accumulated value but may need attention now.
A current whale can regress as recent spending falls. Lifetime spending alone will hide that deterioration. Compare trailing-30-day tiers at two complete milestones—for example, day 30 and day 60—using fixed boundaries.
A cohort that is only 30 days old has no day-180 result. Display “not mature,” never zero. Either wait for a complete monthly cohort to mature or clearly label a partial mature population. Do not switch denominators silently.
Next action: inspect fans who stopped buying or moved down, then test a relevant follow-up or offer against a comparison group. Learn from repeat buyers without assuming their behavior will generalize to everyone.
“Tomorrow’s revenue” is the business question, not permission to present observed lifetime spending as a forecast. Any forecast needs separate assumptions and validation.
3. Marketing: which campaigns bring fans who keep buying?
The main comparison: revenue per acquired fan and ROAS at day 30, 60, 90 and 180, split by original AI/human assignment.
One campaign may acquire an immediate whale. Another may acquire many minnows who become reliable buyers. First-purchase revenue alone cannot distinguish their long-term value.
What the view should show
| Measurement | Calculation |
|---|---|
| Net PPV revenue per acquired fan | Matched cohort net revenue ÷ all mature acquired fans |
| ROAS | Matched cohort net revenue ÷ its allocated acquisition spend |
| Repeat-purchase rate | Repeat buyers ÷ original mature acquired fans |
| Graduation | Fans reaching higher fixed spending tiers ÷ the stated starting population |
| Revenue concentration | How much of the result comes from a few fans? |
Compare AI and human assignments within each campaign where possible. Otherwise, better traffic can look like better AI.
Revenue and acquisition spend must refer to the same fans, attribution rule and observation horizon. Include the spend used to acquire non-buyers. Show no ROAS for organic traffic, missing spend or zero allocated spend. ROAS is not profit.
For a 30/60/90/180-day curve, hold the fan population fixed across the displayed milestones or make changes explicit. For example, fans mature through day 90 can supply their day-30, day-60 and day-90 results; their day-180 result remains unavailable.
Next action: inspect the leading campaign’s repeat buyers and revenue concentration before increasing spend. Expand only when the mature result is sufficiently reliable; keep measuring younger cohorts instead of treating their missing future revenue as failure.
A campaign with the best observed ROAS is a candidate for investigation, not an automatic budget decision.
Shared definitions: one set of rules for every team
Start each report with creator → cohort → measurement → breakdown → dates.
Count a fan’s relationship with a creator. One person following two creators is two relationships. Define “new” with an editable acquisition range; “established” means acquired before its start. A sales window and an acquisition window are different filters.
Use explicit spending brackets. For example:
| Tier | Net PPV spending in the stated window |
|---|---|
| No spend | €0 |
| Minnow | More than €0, below €100 |
| Dolphin | €100 to below €1,000 |
| Whale | €1,000 or more |
These are illustrative thresholds, not industry standards or official Mixpanel definitions. Choose boundaries from your distribution, then keep them fixed during comparisons. Use the same currency and refund policy across creators. Handle negative balances from adjustments separately.
Default current tiers to trailing 30-day spending. Allow lifetime through an explicit date and custom windows for other analyses. Always display the dates. Editing today’s brackets must not rewrite a previously frozen experiment.
Mixpanel supports behavior- and property-based cohorts, and documents historical properties for values that change. Ordinary profile properties contain the latest value, so applying today’s whale label to earlier events can distort the analysis. The framework here uses those capabilities; its thresholds and experiment rules remain our choices. Mixpanel cohort documentation
Save useful event cohorts such as “acquired at Venus Berlin” or a specific Meta campaign, with event dates and tags. Keep a fixed experiment membership snapshot separate from a dynamic marketing segment.
Record acquisition date, campaign, original assignment, actual actors, offer and transaction IDs, availability intervals, exclusions, and spending tier immediately before each interaction. Version the definitions.
Reconcile PPV revenue to the payment ledger or warehouse. Deduplicate purchases and incorporate linked refunds, chargebacks and cancellations once. State whether late adjustments restate the purchase period or appear on their adjustment date. Mixpanel describes warehouse-backed revenue reporting including ARPU, LTV and ROAS. Mixpanel Revenue Analytics
Keep these operational views focused on revenue. Add service costs and profitability separately when the data is reliable; revenue alone cannot establish profitability.
The daily habit
Product: review the comparable revenue result, then inspect the biggest recurring availability or takeover problem.
Business: check same-age cohort development, then inspect regression or concentration before celebrating growth.
Marketing: review mature campaign economics, then decide which campaign needs investigation, more time or a controlled budget test.
Each view should end with one action, an owner and a measurable follow-up. Make every fan count clickable. Give sales a creator-level report using the same definitions, with observed results clearly separated from measured lift and forecasts.
Review daily. Change direction when the evidence is ready—not because yesterday fluctuated.