How this was checked, and where it can still be wrong.

Ampelos was built with AI assistance. Claude, an AI model made by Anthropic, helped write the code and the first drafts of much of the content. I edit and check it, and I decide what goes on the site. AI models can state wrong things with great confidence, which is the main reason the checking below was done at all.

On 24 September 2026 the grape and region content was read through in full. That covered all 96 grapes then in the collection (description, history, parents, origin and growing-region notes) and, for all 163 regions, the overview, climate and terroir text, the classification tiers, the vintage notes and every name in the lists of notable producers (725 producers). A second pass the same day covered the glossary (59 terms), the four Learn guides, every History milestone, and the numeric tasting scores and flavour lists. On 28 September the same checks were made for fourteen grapes and one region added afterwards. The audit was run with Claude's help; I reviewed what it found and made the changes.

Established ampelography and wine reference works, plus a web check for every claim that those could not confirm. The web sources were mostly reference and trade sites (Wikipedia, wein.plus, national wine agencies, producers' own pages), not the original scientific papers. So “confirmed” here means “consistent with the sources I could reach”, not “verified against the primary research”. Where a claim comes only from a winery describing itself, the text now says so.

About thirty corrections, from small to plainly wrong. Mondeuse Blanche had been described as unrelated to Mondeuse Noire, when DNA shows a parent-offspring link. Toro was placed east of Ribera del Duero, when it lies to the west. A date given for Plavac Mali's DNA work was several years out. A parent drawn for Zierfandler rested on sources that only said “probably”. Each was fixed, and claims that could not be supported were softened or removed rather than left standing.

All 68 parent-child links on the lineage map, across 35 grapes with recorded parents, were re-checked against published DNA work. A parent is drawn only where the relationship and its direction are settled. Where two grapes are closely related but it is not clear which is the parent (Savagnin and Pinot Noir, Saint Laurent and Pinot Noir, the two Mondeuses), the text describes it and the map leaves them unlinked. Mutations such as Pinot Gris and Pinot Meunier are not crossings, so they are not linked either. Two links are drawn with a caveat that the grape pages explain but the map does not show: Ciliegiolo as a parent of Sangiovese, where a second study read the direction the other way round, and Gouais Blanc as one parent of Aramon Noir, whose other parent is still unidentified. DNA results are also revised as new studies appear, so any of this can change.

These are the open questions I know of.

• Blaue Zimmettraube as the mother of Blaufränkisch rests on a single source.
• Papazkarası as a cross of Prokupac and Alba Imputotato comes from one 2013 DNA report, and is worded as “genetic research suggests”.
• Where a grape first appeared is often unknown. Grapes with a disputed origin are shown as “probably” from a country on their own pages, and are left out of the origin quiz. The small labels elsewhere (search results, the style finder, the lineage map) still show the best-supported country without that word.
• The claim that roughly half of Turkey's wine comes from the Aegean could not be confirmed.

The sensory scores (body, tannin, acidity, ripening and the rest) are typical values for a grape, not measurements, and a particular bottle can sit well away from them. The detail after a producer's name, such as a founding year, an owner or a signature wine, is what is most likely to be out of date. The quiz and the blind-tasting game are built from the same data, so an error in the data is an error in them. The Studio posters were not part of this audit, and nothing added to the site after 28 September 2026 has had this full pass.

Please tell me. The contact form opens on “content correction”, and I read every message myself. A correction that comes with a source is the most useful kind.