Anatomy of one fleet decision.
A 5-percentage-point utilization gap turned into a Q3 sale calendar, a $14k revenue forecast, and a Volt acquisition target. The full analyst loop in one read: identify, triage, test, ship.
A 5-point yield gap on one vehicle was costing the fleet about $5,400 per quarter.
The question wasn't whether the data was bad. It was: keep operating, intervene, or sell. This page is the loop I ran to decide, end to end. Numbers are real fleet numbers. The decision was real. So was the follow-through.
- Hypothesis ranking
- Evidence triage
- A/B style listing test
- Before/after measurement
- Scenario forecasting
- Decision card / pre-mortem
- Executive synthesis
- Interactive Plotly viz
- Identify. Yaris was the bottom-quartile vehicle by yield: 39% utilization vs fleet median 51%, and $58/booked-day vs fleet median $115. Carrying cost stayed the same. The slot was leaking.
- Triage. Four candidate explanations (price too high, listing too weak, demand mismatch, vehicle wrong-fit). Scored each against the existing trip data before spending money on a test.
- Test. 30-day controlled change on the two hypotheses the data still supported: better photos + a $7/day price cut + an airport-pickup blurb. Single vehicle, no other variables touched.
- Result. Utilization moved from 39% to 44%. Revenue per booked day held flat (the price cut paid for the volume lift). The lift was real but did not clear the long-term retention threshold.
- Ship. Keep the Yaris through summer travel cycle (high demand for cheap airport rentals), list for sale at the end of Q3 with a $7,500 floor, and reallocate capital to a second 2013 to 2015 Volt with a +$14k/yr revenue forecast.
01 / IDENTIFYThe slot was leaking.
Every Sunday I rank the fleet in Vantage Fleet OS on a single composite score: 50% revenue, 30% mean rate, 20% rate stability. The Yaris had been ranked fifth out of five for three consecutive weeks. That alone would not normally trigger a decision (small fleets have a worst-in-fleet by definition). What triggered this one was the gap.
The Yaris is half the fleet on every metric that matters.
The carrying-cost line is what makes a sub-fleet vehicle interesting and not just unfortunate. Each car in the fleet costs the same to insure, list, photograph, message about, and maintain. The slot is the constraint, not the hardware. A vehicle that earns $8.3k in a slot that another vehicle could earn $19.4k in is not breaking even. It is paying me to operate it at a loss relative to the alternative.
A vehicle is flagged for review if it sits in the bottom quartile of fleet yield (revenue per booked day) for two consecutive months and its annual revenue is below 60% of the fleet median. Both conditions were true for the Yaris in Mar 2026. This is the threshold I use to separate noise from signal at five vehicles. With more vehicles I would tighten the quartile cut.
At this point I knew the metric was moving in the wrong direction. I did not yet know why. The next step is the part most operators skip: rule out hypotheses with the data I already have before paying for a test.
02 / TRIAGEFour hypotheses, two surviving.
I wrote down every plausible explanation I could think of for the gap, then asked what the existing 12 months of trip data would have to look like for each one to be true. Hypotheses that did not match the existing data got dropped before they cost me a single dollar of test budget.
Price is wrong (set too high for the segment)
The Yaris was priced at $58/day. Comparable Yaris-class vehicles within 15 miles averaged $48 to $62. Yaris was at the top of that band, not the bottom.
Evidence: Turo competitor scrape, n=14 nearby Yaris-class listings, Mar 2026.
Listing is weak (photos, copy, search ranking)
The Yaris had three photos. The Tesla Model 3 SR+, in the same fleet, had eleven. The Yaris listing had no airport-pickup mention. Booking conversion (views to bookings) sat at 1.8% vs Model 3 SR+ at 4.1% on roughly comparable view volume.
Evidence: Turo host dashboard, listing-level conversion ratios, last 90 days.
Demand is wrong (no one wants this segment locally)
Bay Area Turo demand for sub-$70/day vehicles, especially around airport pickups, is the strongest single demand pocket in the market. Two competing Yaris-class listings within 5 miles were averaging 65%+ utilization. If demand were the issue, they would not be.
Evidence: competitor listing scrape, calendar block-out density as proxy for utilization.
Vehicle is wrong (model is structurally not a fit for the lane)
Possible long-term answer, but not what the data was saying right now. A Yaris with better photos and tighter pricing should be able to clear at least 50% utilization based on competitor benchmarks. If it could not after a real test, then H4 becomes the answer, not the starting hypothesis.
Evidence: comparison to nearby Yaris-class listings already at 65%+ utilization.
Two hypotheses survive the existing data. Two get ruled out for free.
After triage I had two surviving hypotheses (H1, H2) that the data still supported. H3 was ruled out because nearby comps in the same segment were doing fine. H4 was deferred because it is the answer if and only if H1 and H2 fail. That last part matters. Designing the test this way means H4 gets answered for free as a byproduct of testing H1 and H2.
03 / TESTOne vehicle. Three changes. 30 days.
I wanted the cheapest controlled test that could move the metric inside one Bay Area rental cycle (28 days). The constraint was that I run the fleet solo on weekends, so the test could not depend on me being available to message guests at unusual hours. That ruled out things like dynamic pricing experiments and pushed me toward changes that would set themselves and run.
2. $51/day base rate (was $58), -12%, putting it mid-band rather than top-band
3. Listing copy rewritten to lead with "5 minutes from SJC, free pickup, fits 4 + bags"
Util 42 to 49%: partial, dig further
Util < 42%: H4 is correct, plan exit
A purist would test photos, price, and copy independently. I did not, for two reasons. First, the sample size at one vehicle in 30 days is too small to attribute lift cleanly to one variable. Second, this is an operating decision (keep or sell), not a marketing optimization (which lever moved it most). I needed to know if the bundle worked, not which lever to credit. If the bundle had failed I would have been comfortable acting on H4. If it had passed strongly I would have run a follow-up to attribute. As it happened, it passed weakly, which is its own answer.
04 / RESULTThe lift was real, just not enough.
Over the 30-day window the Yaris took on 4 more bookings than the prior 30-day baseline. Booking conversion moved from 1.8% to 2.7%. Utilization moved from 39% to 44%. Revenue per booked day held flat at $58 (the volume lift roughly paid for the price cut, which is what I was hoping for, not a great outcome).
Lift was real on three of four metrics. The fourth tells the story.
The lift built gradually, not in a single week.
H1 and H2 were both real. H4 was also real. The Yaris listing was underperforming and the vehicle is structurally not a fit for the slot.
This is the unsexy answer. There was not a single root cause. Better listing work moved the needle 5 percentage points, which is meaningful. It was not enough to clear the long-term threshold I use to keep capital deployed in a vehicle (annual revenue ≥ 70% of fleet median). That threshold is set deliberately above survival because at five vehicles I cannot afford to keep capital in a vehicle that is just not losing money. The slot needs to earn its keep relative to the next best deployment.
This is a one-vehicle, 30-day test. I have no control group. I cannot rule out that spring travel demand alone would have moved the metric somewhat. The 5 pp lift is the upper bound of what the changes contributed; the true effect is probably 3 to 5 pp. I am comfortable acting on the upper bound here because the decision (sell at end of Q3 with a $7,500 floor) is robust to the true effect being lower, not higher. If the decision had been the reverse (commit more capital based on a strong test), I would have wanted a control vehicle.
05 / DECIDE + SHIPThree operating changes with dates on them.
The analysis is worthless if it does not become a calendar entry, a checkbook entry, or a listing change. Below is what I actually shipped to operations on April 28, 2026.
Keep the Yaris through Q3 travel cycle. List end of Q3 with a $7,500 floor. Source a second Volt by Q3 close.
The Yaris stays on the fleet through summer because Bay Area airport demand for cheap rentals peaks Jun through Aug and the upgraded listing now captures it. The improved photos and copy stay live. The $51/day rate stays. End-of-Q3 listing for sale, $7,500 floor based on KBB private-party + condition adjustment.
The freed capital goes into a second 2013 to 2015 Chevy Volt. The existing Volt is the fleet's #1 ROI vehicle (191% on capital vs M3 Performance at 28%, full math in the data story). A second Volt at $7k to $9k acquisition forecasts to +$14k/yr at the same operational footprint. Payback under one year.
Two scenarios. The keep-through-Q3 path wins by ~$5.4k.
06 / WHAT I'D DO DIFFERENTLYHonest retro.
Three things I would change if I ran this loop again on a different vehicle.
- Set the retention threshold before the test, not after. I had a vague "I want to see at least 50%" rule. Writing it down up front, with both util and rev conditions, would have made the decision easier to defend later when the lift was in the awkward 42-to-49% middle band.
- Pull a control vehicle from the fleet, not from comps. The Volt is in the same lane (cheap, ICE, airport). Tracking its util in the same window would have given me a within-fleet baseline to subtract demand seasonality from the Yaris lift. I did not do this and I should have. Cost: a 30-minute spreadsheet.
- Quote the test cost up front. The photo shoot was three hours of my time. The price cut, in expected revenue forgone over 30 days, was about $90. The listing copy rewrite was 45 minutes. Total cost: ~$200 of opportunity cost. I should write that on the test plan next time so the lift is read against a real denominator, not zero.
The bigger lesson is that the analyst loop only pays back if the last step happens. Most fleet operators I know stop at the chart. I keep the loop honest by forcing every analysis to end in a date on a calendar.
See the rest of the loop.
This decision came out of a Sunday review using Vantage Fleet OS. The dashboard, the case study on building it, and the 12-month data story that informed the Volt acquisition forecast are all linked below.