How to scan wine labels with your phone, and keep the cellar in sync

A mixed case arrived and sat in my hallway for nine days. Not because I did not care — because a mixed case is twelve separate lookups. Twelve producers to spell, twelve vintages to squint at, twelve rows to type. One bottle is thirty seconds. Twelve unfamiliar bottles is an evening, and I never had that evening.

So the case stayed by the door, and the cellar record quietly became wrong — not by twelve bottles, but by however many I opened out of that box before it was ever entered. By the time I got to it I was reconstructing history from memory.

This is the failure mode that matters. Not typing being slow, but the activation cost being high enough to defer — and a deferred entry becoming a wrong record rather than a missing one.

It is worth being unromantic about the mechanism, because knowing it tells you exactly when to trust the result.

The phone takes a photo. The image goes to a model that reads the text on it and works out which parts are the producer, which is the wine name, which is the vintage, and which is the appellation, region or grape. It returns those as fields. That is the whole trick — it is reading, plus judgement about what each piece of text is.

What it is not is a lookup against a database of every wine ever made. That distinction matters: a scan can read a label from a four-hundred-case producer it has never encountered, because it is reading rather than matching. And equally, it can misread a famous label if the photo is poor, because there is no catalogue to fall back on.

It is the obvious question, and the answer is a genuinely interesting one about how wine differs from almost everything else you might scan.

A barcode identifies a product in a catalogue. That works beautifully for a tin of tomatoes, because every tin is the same tin. Wine is not: the same producer’s same wine in two consecutive vintages is, for your purposes, two different things — different window, different value, different taste — and plenty of producers reuse a barcode across vintages or never register one at all.

Below a certain scale, most wine has no meaningful barcode presence. Small estates, restaurant allocations, direct-from-winery bottles and anything from a market that does not require it will simply not resolve. And the failure is silent: a barcode lookup either returns the wrong vintage confidently or returns nothing.

Reading the label sidesteps all of it, because the label is the one place the producer is contractually obliged to tell the truth about what is in the bottle. It is slower than a barcode and it works on everything, which is the correct trade for a cellar.

Most failures are photographic, not linguistic. In rough order of how much they matter:

• Fill the frame with the label. The single biggest factor. A bottle photographed at arm’s length gives the model a label a few hundred pixels wide, and small type is the first thing to go — which is usually the vintage.
• Kill the reflection. Gloss labels and foils bounce ceiling lights straight back. Turn the bottle a few degrees, or move so the light is behind you rather than above the bottle.
• Flatten the curve. A label wraps, so the left and right edges distort. Shoot the centre square-on rather than trying to capture the whole wrap.
• Get the vintage in shot. On plenty of bottles the year is on a neck band, not the main label — and if it is not in the photo, it cannot be read. This is the most common single miss.
• Steady beats bright. A slightly dim sharp photo reads better than a well-lit blurry one. Cellars are dark; brace against something.

It will get things wrong. The useful skill is knowing which errors are likely so you check the right field rather than re-reading everything.

• Vintage is the field to check. Small digits, often on the neck, sometimes absent entirely (and correctly so — NV Champagne has no vintage to find). If one field is wrong, this is it.
• Producer versus wine name gets swapped on labels where the cuvée is set in larger type than the estate. Common in Italy and the Rhône, and easy to spot because the result reads oddly.
• Importer text confuses things. US back labels in particular carry an importer name in prominent type; it is not the producer, but it looks like one.
• Non-Latin scripts — some German, Greek and Georgian labels — read less reliably. Worth a look rather than a glance.

The design rule that follows: a scan should fill a form, not submit one. Anything that writes straight to the cellar without showing you the fields is trading a small amount of your time for a permanent error in your record, which is a bad trade. In CellarPals the scan populates the add-bottle form and you confirm it — label scanning is an input method, not an authority.

The instinct is to treat the photo as scaffolding: scan, extract fields, discard image. Do not. The photo is frequently more valuable than the fields it produced, for a reason that only becomes obvious later.

Text fields are a lossy summary chosen by whatever read the label. The image carries everything nobody thought to extract — the importer, the bottling detail, the alcohol, the cuvée in small type, the shape of the thing you are looking for on a shelf. When a record turns out to be wrong, or a question arises that the fields cannot answer, the photo is the only way back to the truth without going to the cellar.

It is also how you recognise the bottle. Reading “Domaine X, 2019” on a screen and finding it in a dim rack are different tasks, and one label photo does more for the second than any amount of text.

The practical note: shoot the back label too when there is anything on it. It takes a second and it is where the alcohol, the importer, and frequently the only honest statement of what is actually in the blend live.

The hallway problem is specifically a batch problem, so it needs a batch habit. What works for me:

• Photograph all twelve first, then review all twelve. Do not alternate between shooting and correcting — the context switching is what makes it feel long. Twelve photos takes about two minutes.
• Enter them unplaced. Do not make the record wait on the physical job of racking. Get them into the cellar as bottles that exist, then give them a slot on the cellar map when you actually put them away.
• Accept “good enough” on the first pass. Producer and vintage right, everything else approximate. A case fully entered with four correct fields beats six bottles entered perfectly and six still in the box.

This is the same principle as the cataloguing guide: the record only has to be good enough to be trusted, and trust comes from completeness far more than from detail.

Buying depth creates a small problem the scan does not solve on its own: six identical bottles are one wine and six separate objects, and which of those you record determines what the cellar can tell you later.

Record them as six. It feels redundant — one row with a quantity of six is obviously tidier — but quantity collapses exactly the information you bought depth for. Six bottles can sit in three different places, leave on three different evenings, and carry three different notes as you learn how the wine is developing. A quantity field cannot express any of that; it can only count down.

The one time quantity is right is an unopened case you are storing as a case and intend to move as a unit. Treat it as one thing until you break it, then split it into bottles. That mirrors what is physically true, which is the general rule for all of this.

Scan once, then duplicate. There is no reason to photograph six identical labels, and the entry that takes six seconds instead of six scans is the one that actually gets done on delivery day.

Entry is the easy direction. The record rots at the other end — bottles leave without ceremony, and nobody opens an app to announce a Tuesday night.

Two things help more than discipline. First, make consumption as cheap as entry: if marking a bottle drunk takes one tap from a screen you are already on, it happens. Second, do the reconciliation physically and rarely — once a season, count a shelf against the record rather than auditing the whole cellar. Discrepancies cluster, and a shelf is a job you will finish.

And if you thought something about the bottle while you were drinking it, that is the moment to catch it — which is the journal guide, and the reason the scan is worth having at all.

Everything above is how CellarPals works, because it is the cellar I keep. Fourteen days free, then $12.99 a month or $119.99 a year.

If you are moving from somewhere else, the CellarTracker comparison is honest about what carries over and what does not.

How to track a whisky collection — Bottles, values and drinking order.

All guides