How WineHQ works out its answers

The taste model underneath everything

Every grape and every style carries a profile across sixteen dimensions — body, acidity, tannin, sweetness, oak, and so on. Each is recorded as a band rather than a number: a minimum, a typical and a maximum. Nebbiolo is not “tannin: 5”. It is tannin that runs from high to very high and is usually very high.

That choice does most of the work. A single number would let the site claim a precision wine does not have, and it would make every wine of a variety identical. A band carries the variation, and the engines read the whole band — which is why a recommendation can say “usually, though it varies more than most” rather than pretending otherwise.

Profiles are editorial. They are written from the way a variety or a style behaves across its range and are not measurements, and the taste model page sets out what they can and cannot support.

Food pairing: rules, not a lookup table

A dish is described by what it does to a wine — its fat, salt, acidity, sweetness, chilli heat, umami, the cooking method, the sauce. A wine is described by the profile above. 14 principles connect the two, each one a statement about a mechanism: salt makes tannin taste harsher, chilli heat is amplified by alcohol, a dish sweeter than the wine makes the wine taste sour.

A recommendation is what happens when those principles are applied to a specific pair, and every suggestion on the site names the principles that produced it. That is the check a reader can actually make: if the reasoning is wrong, it is visible, and a pairing you disagree with tells you which rule you disagree with.

The same machinery produces the clashes. A wine that scores badly against a dish is not listed as something to avoid unless a principle fired against it — a low score on its own is an absence of enthusiasm, not an objection, and presenting it as one would be inventing a reason.

Some pairings are curated instead: a classic that tradition has settled and the rules would only approximate. Those are marked, and the curator’s reason is shown in place of the computed one. Every principle is published.

Discovery: a direction, not a list

“If you like this, try…” is a question with an implied second half. Lighter? Cheaper? Less oaky? More adventurous? A recommendation engine that ignores that half returns things that are merely similar, which is rarely what anybody wanted.

So every suggestion carries a direction — there are 28 — and a direction is a claim that can be checked. If a wine is offered as “less tannic”, its tannin band has to be lower. The site’s own tests assert exactly that, because a direction nothing verifies is a label rather than a promise.

Alongside the computed suggestions, some entities carry written hints: a specific recommendation with a reason and a caveat, saying what carries over and what does not. Those exist because profile similarity misses things a person would not — that Assyrtiko does for a Sauvignon Blanc drinker what no white with a similar profile does.

Evidence: what the site will and will not assert

Claims are typed, and the type decides what is required. A regulation must be traceable to the instrument that created it. A mechanism must be traceable to research. A taste tendency needs neither and is never written as though it did.

Sources move along a ladder: registered, retrieved, read, passage-verified, claim-linked. A source that has been registered but not read supports nothing, and the site says which rung a claim sits on rather than presenting every citation as equal. The full policy is published, along with every source consulted and the gaps that remain.

How the site checks itself

WineHQ measures how well it answers the questions people actually ask, and the measurement is deliberately unflattering. An intent does not score well because a page exists: the query is put through the site’s own search, and a page that answers a question perfectly and cannot be reached from it counts as not answering it.

What WineHQ will not compute

A method is defined as much by its refusals. WineHQ holds no price data and does not estimate prices, so “is it good value” is answered as “here are comparable wines that do a similar job” and not as a number. It holds no vintage scores. It does not rank producers. It does not review individual bottles, which would require a tasting panel it does not have.

It makes no claims about alcohol and health beyond pointing at national guidance, and it does not present drinking as necessary to a meal, an occasion or an identity. That policy is published too, and an automated check runs over every page to enforce it.

The scale this operates over

192 grape varieties, 252 wine styles and 425 places, each carrying the structured fields the engines read. None of it is generated prose: a page exists because a record was written, and a record that cannot support a useful page is held back rather than published thin.