Our method

Anything that can be computed is computed. An algorithm adds up a budget, compares flight fares, counts rainy days and assembles a booking link; it does so identically every time, and it can be tested. An AI model cannot promise that. So the model is confined to the two jobs where nothing else works — choosing which destinations are worth suggesting for what you described, and writing the narrative that ties a trip together — and it is kept away from every figure.

The weather figures on our destination pages are not forecasts and not model output. They come from Meteostat station records, published under CC BY 4.0: 538 measurement points, read day by day over the last ten years. For a given place and a given calendar day we count how many of those years were actually observed, how many saw more than one millimetre of rain, and what the temperatures were.

That means a page can say "in the last ten years, it rained on this day in three of them" — a fact — rather than "expect rain". Where a station's record has holes and a modelled series fills them, the page says so explicitly: partially modelled. Without measured climate, a place gets no page at all; a guide page with an empty climate section would be exactly the low-value filler we refuse to publish.

Flight prices come from our booking partners — fare search and comparison services — never from a model or an average we invented. A fare is perishable: the figure you see was true when it was collected, and it can move between your search and your booking. Hence the route, the dates and the link to the seller, rather than a promise.

When no usable fare can be found for a route and period, the proposal fails openly instead of quietly substituting an estimate. A trip whose flight price is a guess is not a costed trip.

Accommodation prices come from our booking partners' inventories for the dates you asked about, for real, identified properties. Activities come from ticketing partners, at the price the ticket is actually sold for. We do not average, round up or pad these figures; if a partner returns nothing for your dates, we say the trip cannot be costed rather than fill the gap.

Meals, local transport and small daily spending cannot be booked in advance, so they cannot be quoted. For those we use published price-level indices from Eurostat and the World Bank — an index, never a price. It tells us that a day in one country typically costs more or less than a day in another, and that is how a daily allowance is scaled in your budget. We present it as what it is: an order of magnitude, not a bill.

We use Grok, from xAI, for two things. First, choosing candidate destinations: from what you described, it proposes places worth costing, which the algorithm then either confirms against real prices or discards. Second, writing the account of the trip — why these places, in this order, at this time of year.

The model never sets a price, never invents a schedule and never writes a link. Every figure it sees has already been produced by the algorithm, and the text it writes is checked against those figures before it is displayed. Its limits are real: it can suggest a destination that turns out to be unaffordable, or a pairing that makes little sense. That is precisely why the costing, not the model, has the last word — a suggestion that does not fit your budget never reaches you.

Booking links are assembled by template from structured data: an airport code, a date, a property or activity identifier. A language model producing URLs would eventually produce plausible dead ones, and a dead link on a trip you are about to pay for is worse than no link. Built this way, a link either resolves or reveals a bug in our own data — it cannot be a hallucination.

Some booking links are affiliate links: if you book through them, the seller pays GHIC a commission, at no extra cost to you — the price is the same as going direct. This affects neither which destinations are proposed nor the order they appear in: the ranking is computed before any commercial consideration, and no partner can pay to be ranked higher.

The free version is also funded by advertising — but advertising of our own making: HTML cards we serve ourselves, pointing to travel partners, with no advertising network, no third-party script and no cookie. Subscribers see no advertising at all. The third source is the subscription itself, 9.99 € a month or 59.99 € a year, sold only on this website.

The descriptive texts of our destination guide are written with the help of an AI, then read, corrected and signed off by a human before publication — as are the 557 illustrations, generated one place at a time rather than pulled from a stock library. We state it openly: a reader deserves to know how the text in front of them was made.

What we deliberately do not publish: visa and entry formalities. They change without notice, they depend on your nationality, your route and your length of stay, and a wrong answer could cost someone a flight. We would rather send you to the relevant consulate than be confidently wrong.