Our rental-booking classifier didn't speak British English
Our booking-channel study has a fixed rule: can a customer reserve a specific vehicle or boat for specific dates without a human replying? We have now run it on four markets. The fourth, the UK's holiday areas, taught us
Our booking-channel study has a fixed rule: can a customer reserve a specific vehicle or boat for specific dates without a human replying? We have now run it on four markets. The fourth, the UK's holiday areas, taught us something the first three could not: our own classifier did not speak British English.
The result first
We took 470 independent car, van, boat and water sports hire businesses from Google Maps in 26 UK holiday towns, from Cornwall and Devon to the Norfolk Broads, the Lake District and the Highlands. Chains were left out, which removed 236 listings, 119 of them branches of a single car rental network.
60.4% of the independents cannot be booked online end to end (95% CI 54.3–66.5%). For independent car and van hire it is about two in three. The full study and open data: UK rental online-booking study.
The pipeline, briefly
- One curl per homepage, looking for booking-engine fingerprints, date pickers and carts. Cheap, and it gives each site a stratum.
- A stratified sample of 150 of the 428 sites, weighted toward the strata where the curl is least reliable.
- A real browser (Playwright) on the sample: network traffic, booking links, screenshots.
- A second pass that follows the reservation link and reads the vocabulary of the final screen: checkout words against enquiry words.
- A third pass that dumps the form itself, then hand adjudication of whatever is still unclear.
- A stratified estimator to carry the verified rates back to all 470.
Where it broke: "enquiry" is not "solicitar"
Pass 4 decides with two regular expressions. One matches what a checkout says ("proceed to payment", "pagar ahora", "procedi al pagamento"), the other what an enquiry form says ("request a quote", "solicitar presupuesto", "richiedi un preventivo"). They grew over three rounds in Spanish, English and Italian.
On the UK sample the first run left 76 of 94 sites "unclear". The English in the patterns was American and generic. British hire businesses write "make an enquiry", "enquiry form", "booking request", "call to book". Their checkouts say "basket", not "cart", and "pay a deposit". Their reservation links say "hire now" and "check availability".
Adding those took the unclear count to 64. The remaining unclear sites are structural: Wix and Squarespace sites whose booking app renders inside an iframe, or sites with no booking app at all. Vocabulary cannot fix that. That is what the form dump and the hand pass are for.
The errors cancel this time, for the wrong reasons
- The curl missed 48 of the online booking flows in the sample and invented 1. Many of the misses were booking apps that inject themselves after load: eola, try.be, bookwhen, Holray, letsbook.
- The first browser pass made both kinds of mistake. It counted a basket that belonged to a gift shop selling prints as a hire checkout. It also missed booking calendars sitting in a frame it did not follow.
- Of the 71 hand verdicts, 24 overturned the first browser pass: 13 from bookable to enquiry, 11 the other way.
In Campania the second pass's overturns ran almost all one way (13 of 17 from bookable to enquiry). In the UK they nearly cancel. A study that stopped at the first browser pass would have landed close to the right UK number, but only by accident.
Measuring our own judgement
Four hand verdicts were uncertain. One site said online booking was paused for the winter, one blocked us after the first pass, and two had a booking page with nothing bookable on it. We counted all four as enquiries. Counting them as bookable moves the headline from 60.4% to 57.7%, inside the interval. We published both numbers.
Takeaways if you measure SMEs from their websites
- Localise the classifier per market, not per language. "English" was two dialects here.
- Count the iframes. Small-business booking increasingly lives in a third-party frame, and anything that reads only the top document will undercount it.
- Run a second pass that can disagree with the first, and publish where it did.
The four rounds side by side, all with open CSVs:
- UK holiday areas: 60.4% cannot be booked online
- Spain, five tourist hubs: 68.9%
- Campania: 83.4%
- Florida: 55.2%
Dataset DOI: 10.5281/zenodo.24444126.
I run Renviq, software for small rental businesses, so the obvious bias here is ours. That is why the method and our disagreements are published.
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.