How WhatDrink builds and validates drink information

The guidance on the site is assembled from a few kinds of source:

• • Structured recipe records
• • Normalized ingredients and components
• • Reviewed reference mappings between drinks and their attributes
• • Public venue information, including official venue websites
• • Google Places fields, where applicable
• • Manual corrections
• • Approved reference data

We do not claim sources we do not use. If a fact cannot be traced to one of these, it does not appear.

Every specific claim on the site follows the same evidence rule:

• • Derive facts through calculations over real, structured data.
• • Extract facts from identifiable source evidence.
• • Use generated text only for supported interpretation and wording.
• • Never fabricate missing specifics.
• • Omit the claim when the evidence is insufficient.

AI-assisted wording helps present what the data and sources already support — the wording itself is never a source of fact. A number, a pairing, or a classification only appears when real data or a cited source backs it; otherwise we leave it out.

Objective operations are handled in code, the same way every time. This includes:

• • Identifier resolution
• • Slug validation
• • Component matching
• • Normalization
• • Metric calculation
• • Filtering
• • Deduplication
• • Output caps
• • Route validation
• • Rejecting unsupported values

AI may help with judgement-heavy or wording tasks, such as:

• • Summarization
• • Classification judgement
• • Interpreting candidate information
• • Editorial wording

To be clear about the guardrails:

In other words, AI helps organise and phrase — it does not decide, unchecked, what is true. Not every page is individually written by a person, but every published page is constrained by approved data and validation.

When we rank drinks or compare them, we combine a handful of objective signals rather than a single score. Depending on the guide, that can include:

• • Numeric threshold filtering
• • Balanced trade-off scoring
• • Ingredient and component inclusion
• • Base-spirit classification
• • Normalized flavour matching
• • Nutrition comparisons
• • Alcohol-strength comparisons
• • Ingredient-count comparisons
• • Data-completeness handling

The aim is to surface useful trade-offs — lighter versus stronger, simpler versus more complex — not to declare a single "best" drink.

A guide that says it ranks by a measurement — for example "lowest calorie" or "highest ABV" — orders drinks strictly by that measurement, in the direction the title promises. Only estimates that pass our quality checks are ranked; drinks whose estimate is incomplete may still be listed afterwards, without a rank.

This page is the site-wide view. See how individual WhatDrink decision guides are ranked and evaluated, including the nutrition, ABV and confidence-level detail behind the drink-guide system.

Recommendations such as "similar drinks" or "try next" follow a reviewed pipeline. Generated candidates are never published automatically:

For the specifics of how these figures are derived, see how WhatDrink calculates cocktail nutrition and strength.

A figure existing in our data is not, on its own, enough for us to repeat it as an answer. An estimate can be incomplete — for example when not every ingredient quantity in a recipe could be read, or when a quantity looks implausible. We may still show such an estimate in a table, labelled as an estimate, but we do not use it to rank drinks, to declare which of two drinks is lighter or stronger, or to state it in an answer.

Ranked guides and the data-based answers on drink comparison pages already follow this rule, and we are extending it to the rest of our answer text. Where a value is missing altogether, we say it is not available rather than guess.

Substitute suggestions come only from substitution records that have been reviewed and approved; unreviewed candidates are never shown. Each record states how the swap changes a drink — for example its sweetness, acidity, bitterness, strength, body or colour — only where that effect has been reviewed, and gives adjustment guidance only where it exists. Some recipes have their own reviewed exceptions, which take priority over the general suggestion.

A substitute is a practical swap, not an equivalent: it changes the drink, and the page says how. Substitution notes are not allergy, dietary or medical advice.

Bottle (product) pages show attributes published by the producer or by a regulatory source, and list those sources on the page. Tasting notes are quoted from the producer. When a producer has not published an attribute, we leave it out rather than infer it from the category or brand.

Each ingredient has at most one definition. Definitions are checked against a cited source — such as a regulation or a producer specification — before they are published, and the same text is used wherever the definition appears, including in structured data.

Internal links only point to resolved, supported public routes. If a recommendation would link to a page that does not exist or is not published, we suppress the link rather than send you to a dead end.

If you spot something wrong, our corrections page explains how to tell us, and data updates records material changes we have made.