A travel marketplace in the mould of Expedia serves roughly 1000 searches per confirmed booking. Every search fans out to supplier APIs that charge per call and enforce contractual call quotas. Infrastructure cost per booking is climbing while cost per search falls. Which unit metric should drive architectural decisions?
Show the full answer Hide the answer
The deciding property
The revenue event is the booking, and every search exists only to produce one. A unit metric is useful when improving it improves the business; cost per booking is the only candidate here that has that property, because it carries the whole funnel's cost into the one event that pays for it.
The arithmetic makes the trap visible. If cost per search falls 20% while the look-to-book ratio rises from 800 to 1200, cost per booking has risen by about 20% while every engineering dashboard shows an improvement.
Why this is the right call
Attributing fan-out to bookings makes the look-to-book ratio an infrastructure metric rather than a marketing one. That single reframing changes what the team builds: better intent detection before fan-out, caching of low-volatility supplier responses with a freshness budget tied to booking-failure rate, progressive fan-out that queries the top suppliers first and widens only on thin results, and quota-aware routing that spends scarce supplier calls on sessions likely to convert.
It also connects to the contract. Supplier call quotas are a hard capacity limit, so calls per booking is simultaneously a cost metric and a capacity metric, and the design that improves one improves the other.
Why the other options fail
- Cost per search. The classic local optimum. Driving it down rewards making searches cheaper, which in practice means caching more aggressively and fanning out wider because each call is now cheap. Both raise look-to-book. It is the correct unit only when search itself is monetised, as in a metasearch or ad-funded model, where the conversion event genuinely is the search.
- Total spend against revenue. A board metric, not an engineering one. It moves with pricing, seasonality and product mix, so it cannot be decomposed into anything a team can act on, and a bad month cannot be attributed to a decision.
- Cost per supplier call. This is a cost driver, not a unit of value. Optimising it directly pushes towards longer cache lifetimes for supplier inventory, which trades cost for stale prices and availability - and a booking that fails at payment because the fare vanished is the most expensive failure in the funnel, costing the fan-out, the support contact and the customer.
What would flip the decision
| If this changes | Choose | Because |
|---|---|---|
| Search becomes directly monetised | Cost per search | The conversion event moved |
| Supplier quota rather than money is the binding constraint | Calls per booking | The scarce resource is not currency |
| Bookings vary hugely in margin | Cost per margin-dollar | Not all bookings are worth the same fan-out |
| A single supplier dominates traffic and cost | Per-supplier cost per booking | The aggregate hides the one that matters |
Common weak answers
"Track all of them" is not wrong and is not an answer. One metric has to be the one that gets argued about in planning, and choosing it is the decision being tested.