Sending a misspelt search to the right place

alfred#2104, the branch _mj_2104-search-ranker. 1,131 searches asked of live and of the branch, one at a time wherever the two disagreed, plus 160 asked on eight sites. 1,109 of them are searches customers really made; the other 22 are examples named on the ticket, and every row says which it is. Every figure in the table at the foot is computed in your browser from the rows underneath it.

Try it yourself

The branch is running on its own address. Type the same search into both and compare. It is a development server, so give it a few seconds, and it is restarted from time to time.

• The branch: cornwell, llandegla, north west on the Australian site
• Live: cornwell, llandegla, north west on the Australian site

What changes for a customer

Today the typed submit takes the first source that answers. Ask it for "cornwell" and Mapbox has a hamlet spelt exactly that way, so the customer gets a hamlet in Oxfordshire rather than Cornwall. Ask it for "north west" on the Australian site and it answers North West in South Africa.

The branch asks our own place search, our own campsite search and Mapbox at once, and ranks every answer against the words the customer typed. A word is worth what it is rare: "holiday" and "park" are in almost every campsite name and say nothing, while "cofton" is in one and says everything. An answer more than 4,000km from the map centre of the customer's own site keeps a third of what it scored, which is what stops the wrong continent.

One thing had to be added on top. Our own place search corrects the spelling itself and answers Cornwall for "cornwell", but the ranker scored that answer again with its own word counter, which cannot see that a correction happened. So the branch now believes our place search's own first answer, where the place is one we can believe: a place we hold at least 75 campsites in, or a country, a region or a county. A correction is believed only if it spells the whole of what was typed within two letters.

The numbers

Right means the page names the place the search was a misspelling of, or the destination the customer picked. A page that names a town inside that place counts as wrong, which is strict and is why the misspellings line reads as it does: see what it costs, below.

Before and after

Thirty searches, photographed on both sites tonight. Click a picture to open it full size.

What it costs

The searches the branch answers worse than live does. They are here first, and they are chosen by rule rather than by hand: the most typed losses on every list.

way

out of the search log · 285 searches in 90 days · on the corpus page · the customer chose Way

snowdon

out of the search log · 110 searches in 90 days · on the corpus page · the customer chose Snowdon

wensleydale

out of the search log · 59 searches in 90 days · on the corpus page · the customer chose Wensleydale

uk

out of the search log · 51 searches in 90 days · on the corpus page · the customer chose uk

cotswold

out of the search log · 18 searches in 90 days · on the corpus page · the customer chose cotswold

lakedistrict

out of the search log · 9 searches in 90 days · the customer chose lake district

lakes district

out of the search log · 5 searches in 90 days · the customer chose lake district

lakesistrict

out of the search log · 4 searches in 90 days · the customer chose lake district

debon

out of the search log · 3 searches in 90 days · the customer chose devon

Searches that reach another continent today

Nobody knows what these customers wanted, because they typed and left. Both answers are shown and the reader is the judge.

algarver, portugal

out of the search log · 3 searches in 90 days

camperplaats de buffelgaard

out of the search log · 3 searches in 90 days

cofton holiday park

out of the search log · 3 searches in 90 days

forchheim

out of the search log · 2 searches in 90 days

hopfarm

out of the search log · 3 searches in 90 days

la herraura

out of the search log · 2 searches in 90 days

liechtenstein

out of the search log · 3 searches in 90 days

litchfield

out of the search log · 10 searches in 90 days

montpelier

out of the search log · 5 searches in 90 days

national parc eifel

out of the search log · 3 searches in 90 days

Named on the ticket

Searches you, Stuart, Adam or I put on #2104 or on the merge requests, where the answer moved.

anglesea

named by matt · 4 searches in 90 days · the customer chose Anglesey

What it wins

The most typed searches the branch answers right and live answers wrong.

clarach bay

out of the search log · 203 searches in 90 days · on the corpus page · the customer chose Clarach Bay

wast water

out of the search log · 152 searches in 90 days · on the corpus page · the customer chose Wast Water

flamingo land

out of the search log · 143 searches in 90 days · on the corpus page · the customer chose Flamingo Land

trunch

out of the search log · 137 searches in 90 days · the customer chose Trunch

camber sands

out of the search log · 138 searches in 90 days · on the corpus page · the customer chose Camber Sands

bibury

out of the search log · 115 searches in 90 days · on the corpus page · the customer chose Bibury

south milton

out of the search log · 111 searches in 90 days · on the corpus page · the customer chose South Milton

wasdale head

out of the search log · 106 searches in 90 days · on the corpus page · the customer chose Wasdale Head

pendle hill

out of the search log · 100 searches in 90 days · on the corpus page · the customer chose Pendle Hill

craster

out of the search log · 92 searches in 90 days · on the corpus page · the customer chose Craster

The eight sites away from Europe

Every site sends Mapbox its own map centre, and eight of them sit more than 4,000km from every campsite we sell: the American, Canadian, Australian, New Zealand and South African sites, and the Russian, Swedish and Chinese ones. The same twenty searches were asked of all eight, on both sites.

The losses have one cause and it is worth stating plainly. The 4,000km discount is measured from the customer, so on the American site a Mapbox place in Massachusetts is near and keeps everything it scored, while our own Beverley in East Yorkshire is 7,000km away. Beverley is spelt one letter from what was typed and Beverly is spelt exactly, so the American place wins, and we hold nothing there to sell. The English site never sees this, because there the discount falls on the foreign answer instead.

What that says is that distance from the customer is the wrong measure. The question a ranker should ask is whether we have anything to sell near the answer, and that is the change I would make next. It is not in this branch.

How much traffic the eight carry

140 typed searches of 43,677 in those 90 days, which is 0.32% of them. That is the size of the loss, and it is worth stating before anybody decides what to do about it.

Out of the same GA4 export as everything else, but read straight rather than through the corpus page: that page's query folds every site outside twelve named languages into en-gb, so it cannot tell an American customer from a British one. The raw export keeps the site. The Chinese site does not appear at all, so its row is nought rather than missing.

Judged by hand, because no metric can do it: an Australian typing "anglesea" wants Victoria and an English customer wants Anglesey. The pair is printed beside each judgement so you can disagree with it.

The eight EU markets, where the traffic actually is

The eight sites above carry 0.32% of typed searches. Our EU markets carry 7.95%, which is 3,471 searches in the same 90 days and twenty-five times as many. France, Germany, Italy, Spain, the Netherlands, Ireland, Poland and Portugal were measured live against the branch, each market asked its own twenty most typed searches out of the same log.

Read the place, never the wording. The branch is served from a development machine that renders a French page with English chrome round it, so our own Devon page comes back saying "Best Campsites in Devon, England" where live says "Devon, Angleterre : Meilleurs campings". Same page, same stock, and judged the same. Only rows naming a different place are in the tables.

These customers type their own language and their own countries: Bretagne and Lac Leman in France, Gardasee and Suedfrankreich in Germany, Lago di garda and Toscana in Italy, Huesca and Costa brava in Spain, Balaton in Poland, Algarve in Portugal. An earlier version of this page asked all eight markets the same twenty English searches and reported 19 better against 16 worse. That measured a premise nobody had checked, and the table below replaces it.

The wins are worth having. The Irish site answers Galway for "galway", where live answers Silloth in Cumbria, three hundred miles away. The Spanish site answers Gijon where live answers Madrid, and Bilbao the city where live answers a bank building. The Portuguese site answers the Algarve where live answers Garve in Scotland. The Dutch site answers Lake Garda for "gardameer" where live answers one town on it.

What an English name does on a site that is not English

Our place names come back translated, and the ranker scores the typed words against the name it is given. So a customer typing an English name on a non-English site scores nothing against the right page, and any answer whose name is not translated covers the search outright and wins.

The place search returns our North West England page as "Angleterre du Nord-Ouest" on the French site, while North West Romania keeps its name, so "north wes" answers Romania. Our Cornwall page is "Cornualles" on the Spanish site, so "cornwall" answers St Austell, a town inside the county whose name never changes. On real traffic it costs "Brittany" on the French site, typed seven times in 90 days, which answers a village called Pleumeur-Gautier, and "England" on the Polish site, typed four times, which answers Coventry.

The same twenty English searches asked of all eight markets show the whole family at once. This is the sweep that produced the earlier 19 against 16, kept because it is the clearest picture of one fault rather than a picture of the traffic: 19 better, 16 worse, 19 sideways.

What it costs

Eighteen searches are answered worse than live answers them: fifteen misspellings and three of the most typed. Every one is here, classed by hand, with what each site says.

8 of the eighteen land on a town inside the place that was typed, which the measure counts as a loss and a customer would not: "lakes district" answers Penrith, in the Lake District, and "essx" answers Westcliff-on-Sea, in Essex. 9 are the wrong place. Six of those are a Mapbox point spelt exactly as the customer typed it, where the place they meant is a letter away. The other three are answers of our own: Angles in the Vendee, a campsite called Scar Close in Lancashire, and one search that names nothing at all. 1 is a missing space, which is the next change.

The most typed of them, "way" with 283 searches, is not in this list at all: it is on the list where the customer picked a destination, and live answers Waytown in Dorset while the branch answers Clun in Shropshire. Neither is a place called Way.

Where these searches came from

Every search on this page is labelled with where it came from and how often customers really made it. 1,109 of the 1,131 are real searches out of the log. The other 22 are examples somebody typed to test a rule, and all 22 of them are on the ticket list.

The search log is GA4, 9 June to 6 September 2026, every view_search_results event. A search counts as typed where the page still carried ?q=, which is the customer typing and pressing enter rather than picking a suggestion. That is the path this change alters.

526 of these searches are on the published corpus page, search-terms-corpus.html, which is the same export cut a different way: the top 250 terms per site and per set. The lists here are cut at 300, so about fifty of the most typed and fifty of the picked fall past that page's own line. The misspellings are on none of it, because that page carries popular searches and a misspelling is by definition rare.

The table at the foot carries the same label on every row, so a reader can sort a claim by whether anybody ever made the search.

The five lists

Each one was chosen a different way, and the differences matter more than the totals. Tap one to filter everything below.

The wins worth reading

Each of these sends a customer to another country today.

What it still gets wrong

Every search

How it was measured

Every search was asked of www.pitchup.com and of the branch over /search/?q=, which is the submit a customer makes when none of the suggestions is the place they mean. The answer is the place the page names: the title where the search landed on one of our own pages, and the summary line, "Showing results for", where it landed on a campsite's coordinates. Reading only the title counts every campsite landing as a loss, because that title names a price and no place.

Six requests at once against a development server do not get the branch's answers, so every search where the two sites disagreed was asked twice more, one at a time, and a search that answered differently the second time is marked rather than counted. None of them did.

One search, "galway", was cut from the log twice and so was measured twice. An earlier version of this page counted it twice as well and said 147 of 162 landed near. It is 146 of 161, which is what the merge request has always said.

The eight sites were read differently again, because neither the title nor the words "Showing results for" survive translation. The line the page marks as its own search summary is read instead, which is markup rather than words and works in Russian and Chinese.

"Another continent" cannot be read off a page at all, because "North West, South Africa" answers "north west" and reads as a hit. It comes from an offline copy of the ranker, over answers fetched the same night, which holds every candidate's position. Those answers are not the same fetch as the page, and Mapbox moves: on 5 of 461 searches the two name different places, so read the flag as the shape of the answer rather than as this page's own.

The pictures were taken with a headless Chrome on the same night, at 1100 pixels wide, with the cookie banner hidden. Each one waits for the map to finish: every tile loaded and the fade over, because on a development server a picture taken too early shows a grey box where the price pins and the search radius should be.