Is your catalogue ready to go live? A product data quality checklist
A catalogue can look finished and still let you down on its first day: a filter that lists one value three ways, images a channel will not accept, figures nobody can trace. These are the checks to run before it goes live, and how to run each one in the spreadsheet you already have.
Have a supplier file like this? Send it exactly as it arrived and see the same products before and after.
What ready to go live means
A catalogue is ready to go live when its data is complete, consistent and traceable enough for every channel it is going to. A product data quality checklist turns each of those into tests you can run, so the answer to “is it ready?” is a set of results rather than a feeling.
- Completeness: the values a shopper filters on. Not every column in the file, but the ones that matter for each kind of product.
- Consistency: same unit, same word, everywhere. One way of writing each value, across every supplier.
- Confidence: how much is traced to a source. The share of values someone can check, rather than take on trust.
They are the three measures the guide to product data enrichment ends on, and they come before any channel’s rules for a simple reason. Channel rules decide whether a catalogue is accepted. These three decide whether it works once it is.
Completeness: the values a shopper filters on
Completeness is the share of products that carry every value their category needs, and it has to be measured one category at a time.
A blank energy rating matters on a fridge freezer and means nothing on a door seal, so a count of blanks across the whole file mixes real gaps with irrelevant ones. Start by listing the key specifications for each category. The filters on your own category pages are the best first draft of that list, because they are exactly the values shoppers use to narrow a range.
Then count. COUNTBLANK on each key column, filtered to one category, gives the obvious gaps. It will not find the blanks in disguise: values such as N/A, TBC or absent look filled in to a spreadsheet and empty to a shopper. Filter for them separately and count them as missing.
Channels add a floor of their own. Google Merchant Center lists seven attributes it requires for every product: an ID, a title, a description, a link to the product page, an image link, availability and price. Without them a product cannot be served in ads or free listings.1 That floor is the least a catalogue needs, not the target. A fridge freezer with all seven and no capacity is accepted by the channel and useless to the shopper comparing three of them.
Consistency: same unit, same word, everywhere
Consistency means each attribute is written one way across the whole catalogue, and it is the check most likely to fail when a catalogue is built from more than one supplier.
Three checks cover most of it. For units, filter each measurement column for every unit in turn, and for numbers with no unit at all, because a width of 60 and a width of 595mm are one fact written twice. For compound values, search the specification columns for cells holding several measurements, such as 1850x595x655, which have to be split into height, width and depth before a filter can use any of them. For spellings, run a pivot table on each attribute column and sort it A to Z, so near-duplicates land side by side.
Spellings are where filters quietly break. On one bathroom retailer’s catalogue of 329 baths, a single Type column held twelve spellings covering eight actual bath types, so shoppers got twelve filter options for eight real things. The fix is one canonical word per value, set once; the post on normalising attribute values works through it.
Identifiers need the same discipline. A barcode column should hold digits only, 8, 12, 13 or 14 of them, which are the lengths Google accepts for UPC, EAN, JAN, ISBN and ITF-14 numbers.2 Look out for spreadsheet damage too. Excel removes leading zeros and converts large numbers to scientific notation, like 1.23E+15,3 so a barcode that has passed through it can arrive a digit short, or as 5.01235E+12, and neither identifies a product any more. The guide to GTINs, EANs and MPNs covers how to check and repair them.
Confidence: how much is traced to a source
Confidence is the share of values you can trace to a named source, and it is the measure most quality checklists leave out.
A catalogue can be complete and perfectly consistent and still be wrong. A capacity ten times too big, or a title that contradicts the specification underneath it, reads as data like any other value. The only defence is being able to ask where each value came from, and getting an answer.
Two checks make it measurable. First, count the products that have no source recorded against their values. Second, take ten values at random, open the source for each, and compare. Any disagreement in a sample that small is worth chasing before launch, because it is unlikely to be the only one.
What cannot be verified should be listed, not hidden. On a catalogue of 901 products for a percussion and pro-audio retailer, we checked 874 against a named source and listed the other 27 as unchecked rather than guessing at them. Knowing what we could not verify is what makes the number at the top of the report worth reading.
The last check in this group is the simplest. Nothing flagged along the way should still be open: every flag is either fixed, or accepted on purpose by someone who read it.
The checklist, check by check
These ten checks can all be run in Excel or Google Sheets on the file you already have. None needs anything more than a filter, a pivot table or a simple formula.
| Check | How to run it | A pass looks like |
|---|---|---|
| Key specs present on every product | COUNTBLANK on each column the category’s filters use, one category at a time; then filter the same columns for N/A, TBC and absent. | No blanks and no placeholders in key columns |
| Units consistent inside each field | Filter each measurement column for every unit in turn, and for numbers with no unit. | One unit per column |
| Compound values split into their own fields | Search specification columns for an x between numbers, a slash or a semicolon, as in 1850x595x655. | One measurement per cell |
| One spelling per value | A pivot table, or UNIQUE, on each attribute column, sorted A to Z so near-duplicates sit together. | Each real option listed once |
| Every product has a usable image | COUNTBLANK on the image column, then open a sample of links and note each image’s size in pixels. | Every link opens an image at or above the channel’s minimum |
| Titles and descriptions within limits | LEN on both columns, filtered to anything over the channel’s limit. | Nothing over the limit, nothing cut mid-word |
| Barcodes that could be valid | LEN on the barcode column, a filter for anything containing E+ or a space, and a check digit test on the rest. | Digits only, 8, 12, 13 or 14 of them |
| No duplicate IDs | COUNTIF each ID or handle against its own column, filtered to counts above 1. | Every product appears once |
| Every value traced to a source | COUNTBLANK on the source column, then open ten sources at random and compare the values. | Unverified products listed by name |
| Nothing left awaiting review | Filter for rows flagged during checking. | Every flag fixed, or accepted on purpose |
Run the list on the supplier file when it arrives as well as on the finished catalogue. On the raw file the results size the job; on the finished one they tell you whether the job is done.
Channel rules to check against
Each channel sets its own floor, so run the checks against the strictest channel the catalogue is going to.
- Google Merchant Center. Titles of up to 150 characters and descriptions of up to 5,000.1 A GTIN wherever the manufacturer assigned one, with a correct check digit and never guessed; products with missing or incorrect GTINs may have limited visibility.2 Images at least 100 by 100 pixels, or 250 by 250 for clothing,4 rising to 500 by 500 for every product on 31 January 2027, with warnings in Merchant Center since 14 April 2026.5
- ChatGPT. OpenAI’s product feed asks for titles of at most 150 characters and plain-text descriptions of at most 5,000, a real brand rather than a placeholder, and any GTIN as exactly 8, 12, 13 or 14 digits with a valid check digit and its leading zeros kept.6
- Shopify. A product CSV must be UTF-8 encoded, needs a unique handle for each product, and every image URL must be a publicly accessible direct link. Two variants with identical option values are rejected on import.7
The image change is the one with a date on it, and the one most catalogues built from supplier thumbnails will feel. The post on Google’s 500 by 500 pixel minimum covers what to check before January.
How a quality report should read
A quality report is worth reading when it shows its working: what was checked, what passed, what failed and what could not be checked at all.
RefynData’s quality score is built that way. It gives one score per catalogue over 20 checks, broken into completeness, consistency and confidence, plus a grade on every product, worst first, so you know where to start. It recalculates every time the data changes. Checks that found nothing are shown too, because a score you can read is worth more than a score you have to trust.
It also warns rather than blocks: exporting a catalogue that scores 61 is your call to make. A launch can carry known gaps, such as an optional specification missing on a slow-selling line, as long as the gaps were seen and accepted rather than missed.
Some checks are worth running as late as possible, because the data changes without anyone touching the file. Supplier image links rot quietly. On one musical instrument retailer’s catalogue, we fetched all 12,015 image links before the file was built, and 3,566 came back broken or blocked. Nearly a third of them. The import never saw one.
When to run the checks
Run the checklist at four points, because a catalogue that passes at one can fail at the next.
- When the supplier file arrives. To see the size of the job before anyone commits to a launch date.
- After enrichment. To confirm the gaps were filled and every added value has a source.
- After the copy is written. To hold each title and description against the specifications it claims.
- Just before import. To rerun the channel limits and fetch every image link again, since links can stop working between enrichment and import.
If you would rather hand over the whole job, from the supplier file to a scored catalogue ready to import, that is what RefynData’s product data enrichment service does. Either way, the checklist is the same, and so is the question it answers.
Questions
What should a product data quality checklist include?
At minimum, checks for completeness (every product has the values its category’s filters use), consistency (one unit, one spelling and one identifier format per field) and confidence (values traced to a named source). Add the rules of each channel the catalogue is going to, such as title and description lengths, barcode formats and minimum image sizes, and a final check that nothing flagged is still open.
How do I measure product data completeness?
Measure it per category rather than across the whole file. List the attributes that matter for each category, usually the ones its filters use plus any a channel requires, then count the products missing each one. Count placeholders such as N/A or TBC as missing too, because a spreadsheet’s blank count treats them as filled in.
What is a good product data quality score?
There is no universal pass mark. A useful score is one you can read: it shows which checks passed, which failed and what could not be checked, and it grades each product so the worst can be fixed first. Whether to launch with known gaps is a business decision, and the score should inform that decision rather than make it.
Should failing data quality checks stop a catalogue going live?
Not usually. Some gaps are acceptable at launch, such as an optional specification missing on a slow-selling line, and a check that warns and lists exactly what is missing lets you decide with the gaps in view. The exception is a channel’s hard rules, such as a required field left blank, because the channel will reject those products anyway.
How often should product data quality be checked?
Whenever the data changes: when a supplier file arrives, after enrichment, after descriptions are written and before every import. Some checks go stale on their own. Image links stop working without anyone touching the file, so link checks are worth rerunning before each export rather than only at launch.
Sources
- Google Merchant Center Help. Product data specification. Read .
- Google Merchant Center Help. GTIN [gtin]. Read .
- Microsoft Support. Keeping leading zeros and large numbers. Read .
- Google Merchant Center Help. How to fix: Image too small. Read .
- Google Merchant Center Help. Merchant Center product data specification update 2026. Read .
- OpenAI Developers, Agentic Commerce. Products: product feed specification for ChatGPT. Read .
- Shopify Help Center. Solutions to common product CSV import problems. Read .