How we count regret
Every verdict on this site is one number: the share of reviewers who paid for something and regretted it. This page explains where that number comes from, how sure we are of it, and what we leave out.
1. One page, one paid decision
We don't review attractions. We review purchases: the Express Pass on top of the park ticket, the tower on top of the basic Sagrada Familia entry, the London Pass instead of paying at the door. If a question doesn't involve paying for something, it doesn't get a verdict page.
2. Where the reviews come from
- Google Maps reviews of the place itself. This is the main source.
- Reddit threads for decisions that have no single place, such as rail and city passes.
- Trustpilot for companies that sell passes. These reviews skew towards customer service, so each page says which sources it used.
We collect public reviews through a scraping service, without reviewer names or photos. We don't use TripAdvisor. Raw reviews stay in a private archive and are never published.
3. Which reviews count
A review counts only if it talks about the decision, not just the place. For the Express Pass, a review that says "great rides" is ignored; one that says "the express line was still 40 minutes" counts.
It also has to come from someone who paid. Many reviews say "we should have bought the pass". Those people didn't buy it, so they are not part of the regret figure. Where there are enough of them, we report them separately.
Only reviews from the last 24 months count. Prices, ride lists and booking rules change, and a 2019 complaint about a policy that no longer exists shouldn't move today's verdict.
4. How each review is classified
A language model (Anthropic's Claude) reads each review and records, using a fixed list of labels:
- whether it is about the decision and whether the reviewer paid;
- the verdict: regretted, mixed or glad they paid;
- complaints from a closed list: price, still had to queue, crowds, not much to see, felt like an upsell or scam, over too quickly, booking hassle, rude staff, closed or under renovation, not what was advertised, weather ruined the view, accessibility;
- praise from a closed list: the view, actually skipping the queue, the value, the experience, the organization, the upgrade itself;
- a short quote of up to 12 words. We check that each quote appears word for word in the original review; if it doesn't, it is thrown away.
5. Checking the model
Before a page goes live, 50 randomly chosen reviews are read again with their labels, and each verdict is marked right or wrong. The target is at least 85% agreement; below that, the instructions change and the reviews are classified again. For our first pages this second read was done by a more capable AI model (Claude Opus), not by a person. Gonzalo Gutierrez will repeat it by hand, and this page will say when that has happened.
- Top of the Rock vs Empire State Building: 92% agreement on 50 reviews, checked by a second AI model (2026-11-03).
6. From reviews to a stamp
| Regretted paying | Stamp |
|---|---|
| 15% or fewer | Worth it |
| Between 15% and 30% | Only if (the page says who the "if" is) |
| 30% or more | Skip it, plus the cheaper way |
Every figure comes with a 95% confidence range (the Wilson interval). If that range crosses one of the thresholds, the stamp drops to "Only if" and the page says so: we would rather be cautious than confidently wrong.
A page needs at least 150 reviewers who paid to be published. Situations inside a page (for example "busy day" or "travelling with small kids") need at least 30 reviewers to show their own figure.
7. Words, prices and money
- The written verdict is drafted from the numbers by Max, our house voice, and checked: every number in the text must appear in the data.
- Prices, opening hours and policies come only from official sources, with the date we checked them. If we haven't verified something, we don't mention it.
- Some links earn us a commission. The data sets the stamp; a commission never does. When the answer is "skip it", we say so and link to the cheaper option. See our affiliate disclosure.
8. Updates and corrections
Each page is refreshed with new reviews once a month. The "updated" date changes only when the data does.
The figures for every published page are free to download from our data page (CC BY 4.0).