Canonical benchmark · Flights + Accommodation · 4 scenario families · 40 requests · September 2026

A smaller shopping space that still holds the right options.

Discovery removes 67.7% of the logical shopping space across flight and accommodation discovery while retaining 67.5% Recall@10 in the two-track mean. Reduction and relevance are reported together because neither number means anything alone.

Internal measurements across 4 scenario families, Flights and Accommodation tracks, with 5 repeats each (40 headline requests).

Why relevance is the constraint

The reduction is only as good as what survives it.

Reduction alone proves nothing

Returning no candidates reduces the shopping space by 100%. Any reduction figure published without a relevance figure beside it is unfalsifiable.

Recall@10 shows how much relevant supply survived

At the ten-candidate cutoff Discovery retains a mean 67.5% of relevant options across the equally sized Flights and Accommodation tracks, including 69.5% on Flights.

NDCG@10 shows they are ranked, not just present

Mean NDCG@10 is 67.8% across the two reported tracks, including 76.3% on Flights. Relevant options appear near the top of the shortlist rather than at its tail.

These measure agreement, not fit quality

The reference ordering is price-led while the production objective is multi-factor, so recall and NDCG measure agreement with that reference rather than traveller-fit quality directly. 100% agreement is neither expected nor necessarily desirable, and the gap holds both correct fit-over-cost decisions and genuine ranking errors.

Reduction on its own is not a result: returning nothing reduces the shopping space by 100%. These reductions are reported only alongside the recall and ranking quality measured on the same runs, which shows the reduction preserves substantial reference relevance rather than simply returning fewer options.

The reference ordering is price-led, while the production objective is multi-factor: seasonal suitability, reachability from the origin, stated preferences and affordability together. Recall and NDCG therefore measure agreement with that reference rather than traveller-fit quality directly, and 100% agreement is neither expected nor necessarily desirable. The gap contains both cases where Discovery is right to prefer fit over cost and genuine ranking errors, and these metrics do not separate the two. Read them as evidence that relevant options survive pruning and rank highly, not as a score to drive toward 100%.

Headline metric

Logical shopping-space reduction

1 - selected_destination_date_units / broad_destination_date_units

Numerator

Exact destination/date units returned by Discovery.

Denominator

Unique destination/date units actually exposed to the broad provider-backed lifecycle. Not planner expansion, not the candidate limit, and not provider HTTP calls.

Aggregation

Four scenario families across two reported tracks, Flights and Accommodation, with five repeats per scenario (40 requests). Mean figures are the arithmetic mean of the track metrics because each track contains the same number of scenarios. The separate Combined track is not included in these figures.

Per-track results

Full results across both shortlist cutoffs.

Results use the ten-candidate cutoff because that is the shortlist size Discovery is typically integrated at. Recall measures whether relevant options survived; NDCG measures whether they are ranked near the top.

Flights

Recall@10

69.5%

NDCG@10

76.3%

Shopping-space reduction

67.9%

Accommodation

Recall@10

65.5%

NDCG@10

59.2%

Shopping-space reduction

67.5%

What we do not publish

Provider-side reduction is measured but withheld.

Provider-side reduction — fresh HTTP requests, fresh provider request units, total provider units and configured estimated cost — uses the same formula but is not published. It is reported only when both arms have comparable cache state and complete accounting; unknown request counts remain unavailable and are never counted as zero.

Run it on your own traffic

The only benchmark that matters is yours.

A 4–6 week pilot measures recall, ranking quality and shopping-space reduction against your own scenarios rather than ours.

See what you would save

MyEscapePlan Business

AI travel decisioning for travel businesses: connect customer intent with your catalogue, supplier policy and commercial rules across digital and advisor workflows.

MYESCAPEPLAN LTD · Company no. 16395758 · Registered in United Kingdom

Discovery API, Advisor, website widgets, tour operator and Meetings & Events workflows are all available now, on the same controlled decisioning layer.

Discovery benchmark figures, their formula and their limitations are published in full.

Read the benchmark methodology ↗

Tour operator and Advisor pilots can start with your existing catalogue and supplier policy.

Products

Consumer app ↗

Legal

Terms & conditions

API Terms

Privacy