Funnel-Stage Cost Ratio

A travel marketplace in the mould of Expedia serves roughly 1,000 searches for every confirmed booking. Engineering runs a successful quarter: cost per search falls 20%. In the same quarter a wider fan-out and more aggressive retries lift the search-to-booking ratio from 800 to 1,200. Cost per booking rose by about 20% while every dashboard showed an improvement.

The funnel-stage cost ratio is the corrective. It carries the full cost of every preceding attempt into the one event that produces revenue, so the attempt count becomes part of the metric rather than a free variable outside it.

Why it matters

Unit-cost metrics are only useful when improving them improves the business, and a metric defined on a non-revenue stage fails that test whenever the stage count can move. Search, recommendation fan-out, retry, speculative prefetch, model sampling and quote generation all share this shape: cheaper attempts invite more attempts.

It matters a second time when the attempt consumes a contractual resource. Supplier APIs are commonly metered and quota-limited, so attempts per conversion is simultaneously a cost metric and a capacity metric, and the design that improves one improves the other. That alignment is rare enough to be worth building a metric around.

Implementation patterns

• Define the denominator as the revenue event - booking, order, activation - and attribute the cost of the attempts that led to it, including the ones that did not convert.
• Attribute by session rather than by request, because the attempts that produced a booking are a session's worth, not a single call's worth.
• Decompose it into attempts-per-conversion multiplied by cost-per-attempt, and track both. A falling total that hides a rising attempt count is the failure the metric exists to catch.
• Segment by the dimension that dominates variance - supplier, market, channel - because an aggregate ratio hides the one segment destroying margin.
• Set a freshness budget on any cache that reduces attempt cost, tied to the downstream failure rate rather than to a time-to-live picked by feel.
• Make the ratio a design-review input: any proposal that lowers cost per attempt must state its expected effect on attempts per conversion, or it is not a cost proposal.

Industry example

The pattern is visible wherever a cheap upstream stage feeds an expensive conversion. Search-driven marketplaces meter supplier calls; advertising systems run candidate generation and ranking over thousands of items to serve one impression; retrieval-augmented systems fan out to a vector store and a reranker for one answer. In each case the published engineering discussion focuses on cost per candidate, and the economics are decided by candidates per conversion. Absent a documented figure for a specific company, treat the 1,000-to-1 search-to-booking shape as the archetype it is: characteristic of travel marketplaces, and not a claim about any one of them.

Failure scenarios

• The local optimum. Cost per attempt falls, attempts rise faster, margin worsens, and the improvement is celebrated.
• Caching into stale inventory. Longer supplier cache lifetimes cut attempt cost and raise the rate of bookings that fail at payment because the price or availability moved. That failure costs the fan-out, the support contact and the customer, and it is the most expensive outcome in the funnel.
• Quota exhaustion mid-peak. Attempts grow until the supplier contract's call limit binds, and the system degrades on the highest-demand day of the season.
• Mix shift mistaken for regression. A marketing campaign brings low-intent traffic, the ratio worsens, and engineering is asked to explain a cost problem it did not create. Segmenting by channel prevents the wrong post-mortem.

Trade-offs

Optimising the ratio means spending effort upstream of the revenue event: intent detection before fan-out, progressive fan-out that widens only on thin results, and quota-aware routing that spends scarce calls on sessions likely to convert. All of these reduce coverage, and reduced coverage can reduce conversion. The metric does not tell you where to stop; it tells you that both terms are in the same equation, which is what cost per attempt does not.

There is also a measurement cost. Session-level attribution requires carrying an identifier across the funnel and joining to the conversion event, which is real pipeline work.

When not to use it

When the attempt is the monetised event - metasearch, advertising-funded search, an API sold per call - cost per attempt is the correct unit and this ratio adds nothing. When the attempt count is fixed by design and cannot move, the ratio is a constant multiple of cost per attempt and you may as well track the simpler number. And when conversion events are rare enough that the denominator is noisy at a daily granularity, use a longer window or the metric will generate arguments rather than decisions.

Interview question

Q: Your team's cost per request has fallen every quarter for two years and gross margin has fallen with it. How do you work out what is happening, and what metric would you put in front of the next planning meeting?

What a strong answer covers: identifying the revenue event and checking whether attempts per conversion moved; decomposing unit cost into attempts and cost-per-attempt; naming the mechanism by which cheaper attempts invite more of them; segmenting by channel to separate mix shift from regression; the stale-cache failure that trades cost for booking failures; and the honest limit that reducing attempts can reduce conversion, so the metric frames the trade rather than settling it.

Quick check

Quiz: Cost per search fell 20% and cost per booking rose 20%. How? — The search-to-booking ratio rose faster than unit cost fell. Cheaper attempts invited wider fan-out, so the funnel's total cost per revenue event went up while the per-attempt dashboard improved.

Flashcard: When is cost per attempt the right unit rather than cost per conversion? — When the attempt is itself monetised, as in metasearch, advertising-funded search or a per-call API. Then the attempt is the revenue event and there is no funnel to carry.