Measuring tourist pressure from Wikipedia pageviews

Published 4 August 2026 · crowddodger methodology

crowddodger records daily Wikipedia pageview counts for every destination it covers: 78 crowded destinations, 19 candidate alternates, and a 100-article control set, collected daily and backfilled to July 2015 — about 2.4 million daily observations. We record three raw numbers per place per day: total pageviews, mobile-web pageviews, and mobile-app pageviews, always excluding bot and spider traffic. Everything below is computed from that one table.

This paper explains the two signals we derive from it, shows why each one is needed, and states plainly what the data can and cannot support.

Finding 1: raw volume hides seasonality in famous places

The obvious way to read pageviews is as a proxy for interest: more views in July means more people planning or taking a July trip. For mid-sized destinations that works. Santorini’s busiest month draws 1.9× its quietest; Dubrovnik’s 2.2×; Kotor’s 1.9×. The tourist season is right there in the volume.

It fails for the most famous places. Venice’s peak-to-trough ratio is just 1.26, and Reykjavík’s is 1.24 — nearly flat lines, for two destinations with strongly seasonal tourism. The reason is baseline fame: articles like Venice are read year-round by students, news readers, and the curious, and that steady readership drowns the travel wave.

Mean daily views by calendar month, averaged over all years since 2015:

Note Zermatt: its volume peaks in January — the only winter peak in the table, and exactly what a ski resort should show. The volume signal is honest; it just runs out of contrast for world-famous names.

What volume conceals, the share of views coming from phones reveals. People physically travelling — or physically there — look places up on their phones. The mobile share of the same articles moves with the season even where total volume doesn’t:

Mobile share of views (%), by calendar month:

Venice’s flat volume line hides a clean 7-point summer swing in mobile share. But the raw share has its own contamination — look at Santorini’s December (69.2%, its highest cell, in its deadest tourist month). That brings us to the correction.

Finding 2: the presence index, and the two ski towns that validate it

Mobile share has a global component that has nothing to do with any destination: everyone browses more on phones during holidays and in summer. Measured across all our destinations, December runs about three points hotter than November purely from holiday phone use. Santorini’s absurd December cell above is that effect, not tourists.

To remove it, we built a tourism-free control set: 100 English Wikipedia articles with no travel component — chemical elements, dead mathematicians and physicists, grammar terms, common animals and plants, basic mathematical concepts — chosen to avoid anything place-based, seasonal, or news-prone. Their aggregate mobile share, month by month, is the global rhythm of phone use with tourism removed:

(An earlier version computed the baseline from the destinations themselves. That was circular — most of our destinations share the same northern summer season, so we were subtracting the signal from itself. The control set fixes that.)

The presence index is a destination’s 7-day-smoothed mobile share minus the control baseline for that calendar month, in percentage points. Positive means more on-the-ground phone browsing than the world’s background rate for that time of year.

Presence index by calendar month (points vs control baseline):

Do not compare levels between rows of this table — the shape of each row is the finding, not its height. Every value is positive for a reason that has nothing to do with tourism: travel articles attract a phone-heavier readership than the academic control set at all times of year, so the offset is constant and uninformative.

The comparable form is each destination’s presence index as a deviation from its own annual mean — pure shape:

In this form the seasonal geometry is unmistakable: the four Mediterranean rows rise together through summer and collapse in winter; Reykjavík is a flat line; Zermatt is positive December through March; Queenstown is positive November through April — the southern summer — and most negative exactly when Europe peaks.

The validation is the two ski towns, which we included as controls precisely because their seasons run against the Mediterranean grain:

• Zermatt flips to a January peak. In raw mobile share, Zermatt’s best months were August and December (tied) — the global summer and holiday bumps sitting on top of its real pattern. Corrected, its best months are January–February and its worst is November, the dead shoulder between hiking and ski seasons. That is a ski town’s true shape, and the raw metric could not see it.
• Queenstown inverts with the hemisphere. New Zealand’s summer is December–February, and that is exactly where Queenstown’s presence peaks — with its deepest trough in June–August, the very months every European destination in our set peaks.

Two independent controls, two different mechanisms (season type and hemisphere), both behaving as geography says they must: that is what persuaded us the index measures presence rather than an artifact of how we built it. The same correction also delivered a negative result we kept: Reykjavík’s apparent summer peak in raw mobile share turns out to be almost entirely the global summer effect. Corrected, Reykjavík is nearly flat year-round — consistent with a city whose winter visitors (northern lights) concentrate in town while summer visitors disperse to the countryside.

Finding 3: Venice, down by both signals at once

The point of carrying two semi-independent signals is that they can confirm or contradict each other. For the 30 days ending 2 August 2026, compared with the same 30 days in 2025:

• Venice’s total pageviews: 83,753 vs 103,653 — down 19.2%.
• Venice’s mean presence index: +11.2 vs +15.1 — down 3.9 points.

Interest in Venice and evidence of people physically in Venice both fell, measured from the same raw table by two different constructions. Either signal alone could be an artifact; both moving together, by that much, is the closest this data comes to saying “Venice is quieter this summer than last.” (Presence changes are reported in points, not percent — the index sits near zero, and percentages of a near-zero number mislead.)

Validation against official ground truth

In August 2026 we began collecting the official measure of tourist presence: monthly nights spent at tourist accommodation, published by Eurostat at regional level for 26 of our European destinations (and by national agencies for others). This lets us test the presence index against numbers produced by governments counting actual guests.

For each destination we correlated our monthly presence index with the official bednights of its statistical region, across every month where both exist (36–72 months per destination, 2020 onward — the regional monthly series starts there).

Median correlation is 0.54 over all months and 0.37 when 2020–2021 are excluded. We show both because the first number flatters us: part of the full-period agreement is simply that both series collapsed and recovered together through the pandemic — a shared shock, not independent confirmation.

The pattern in the post-2022 column is the real finding. The correlation is strong exactly where the comparison is fair: where the statistical region is roughly the destination itself and its tourism is overnight-based. Athens (0.90) and Kotor (0.72) are the clean cases, and both hold up when the pandemic years are removed. The correlation weakens for two identifiable reasons, and naming them matters more than the headline number:

1. Region mismatch. Official statistics cover regions, not destinations. Venice’s bednights are all of Veneto’s; Salzburg city sits inside a region whose winter is ski season. Where the region is much bigger than the destination, the two series measure different places.
2. Day-trippers are invisible to bednights. A visitor who arrives at nine and leaves at six never appears in accommodation statistics. For destinations whose crowding is day-trip-driven — Bruges, Hallstatt, Plitvice — a presence signal should diverge from bednights, because the crowds are made of people bednights cannot count.

So the claim this table supports is narrower than “validated,” and we state it as such: where official ground truth measures the same population we do, the presence index agrees with it, up to r = 0.90; where they diverge, the divergence points at day-trip pressure — the form of crowding official statistics see least. That second property is not a consolation prize; it is the reason to compute a presence signal at all.

Limitations of this validation, plainly: the overlap window is short (2020 onward, and only 36–48 clean post-pandemic months); monthly averaging discards the daily structure the index is built from; correlation measures shared seasonal shape, not level accuracy; and the official series lags by roughly two months, so this check can never be current. We will re-run this table as history accrues, and publish the numbers whether they improve or not.

How this is used on the site

An alternate destination qualifies as “quieter” only if its summer peak-month mean daily views are under 0.15 of its paired crowded destination’s — computed from this table at verification time, never taken from a third-party description. Alternates recommended for absorption capacity rather than quietness (large cities that can take the overflow) are labelled as such and are never described as quieter.

Limitations, stated plainly

• Pageviews measure interest, not density. A town with a seventh of Dubrovnik’s readership can still feel packed at 11am in its one narrow street, because felt crowding is people divided by space, and Wikipedia knows nothing about space. Our own re-check made this concrete: two candidate alternates that pass the volume test were described by visiting journalists as busy with day-trippers. Both things are true.
• The 0.15 threshold is a reasonable line, not a derived one. It sits comfortably below every destination we consider crowded and above the noise, but nothing in the data makes 0.15 correct and 0.20 wrong. We state it so it can be argued with.
• Article choice moves the numbers. We track the article matching the catchment people visit (the town, not the island; the lake, not its gateway village). Where the choice is genuinely ambiguous — Ischia the island vs Ischia the town — the measured volume can differ by 20× . Every mapping choice is recorded and reviewable.
• English Wikipedia skews anglophone. Our counts under-represent domestic and non-English-speaking visitors. A destination mainly visited by, say, Italians will look quieter to us than it is.
• The control set has a school-year rhythm. Its articles lean academic (math, grammar, scientists), so its mobile share dips when school starts; September presence readings run roughly a point hot as a result.
• Small articles are noisy. Below about 500 views/day — which includes 21 of our 78 destinations and most candidate alternates — day-to-day swings of 30–45% are routine noise, and we only read those series through a 7-day rolling average.
• Presence index levels are not comparable across destinations. Different articles attract different device mixes at baseline. The index is meaningful within one destination over time; ranking destinations by it is not.