The Product Feed Is Where Your Shop Is Judged

Most sellers discover their feed the way you discover a slow puncture. Sales dip, nothing obvious changed, and somewhere in an account you rarely open there are four hundred products marked disapproved with a reason written in a dialect of English nobody speaks.

The feed is not a copy of your shop. It is a second, stricter publication of the same catalogue, read by machines that will not forgive a missing field the way a human visitor forgives a typo. A shop can look immaculate and feed badly. That gap is where a surprising share of small sellers lose their shopping traffic, and it is one of the few problems where careful AI use genuinely earns its keep.

What is a product feed, and why is it stricter than your shop?

It is a structured list of your products in fields rather than in prose, and it is stricter because a machine has to compare your coat to ten thousand other coats without reading your page.

Your website can describe a jumper as a cosy something for chilly mornings and a human will understand. A feed has to say, in named fields, what the item is, who made it, what it costs, whether it is in stock, what size it is and which other listings are the same jumper in a different colour. The Google Merchant Center product data specification lists the fields every product needs to be served at all: an id, a title, a description, a link, an image link, availability and price. Miss one and the item does not compete. It simply is not there.

The limits are specific rather than suggestive. A title runs to a maximum of 150 characters, a description to 5000. The specification tells you the title must accurately describe the product and match the title on the landing page, and it names what may not appear: promotional text such as free shipping, all capital letters, and gimmicky foreign characters. Those three bans exist because sellers spent a decade trying to win attention with punctuation.

The fields that actually fail, and what to do about each

This table is assembled from the specification and the disapproval guidance rather than from one page, because the rule and the failure live in different documents. The last column is the honest division of labour between you and a model.

The image rule deserves its own sentence because it carries a date. The specification states images must be at least 500 by 500 pixels, with enforcement beginning on 31 January 2027. If your catalogue was built from supplier thumbnails, that is a photography project with a deadline attached, not a field you can fill in.

Why does a wrong barcode hurt more than a missing one?

Because a GTIN is a claim about identity that other systems trust, and a false claim merges your product with somebody else's. A missing one only means you are judged without it.

GS1, which runs the system, defines a Global Trade Item Number as a unique number identifying a pre-defined trade item that can be priced, ordered or invoiced anywhere in a supply chain. It also says the brand owner, meaning whoever owns the product specification, allocates it, and that a distinct product gets a distinct number. Read that against a catalogue where somebody put the same supplier barcode on all eight sizes of a boot, and the problem is obvious: eight products now claim to be one product.

The specification is unusually blunt on this point. Provide a GTIN only when you are sure it is correct, and when in doubt leave the attribute out. That is the rare case where doing less is the compliant answer. If you make your own goods and have no barcodes, the correct move is not to invent one. It is to declare that identifiers do not exist for the item, which the feed has a field for.

This is also the single hardest boundary for AI use in a feed. A model asked to complete a spreadsheet will complete it. Give it a column called gtin with gaps and some models will produce plausible thirteen digit numbers, because producing plausible things is what they do. We wrote about the general version of this failure in why language models hallucinate, and a barcode column is the crispest example you will ever see. Never let a model fill an identity field. Let it flag the gap.

Where does AI genuinely help with a feed?

In three places, all of them transformations of data you already hold rather than inventions.

Rewriting titles from structured fields. You have brand, product type, colour, size and material sitting in separate columns. A feed title wants them in one string, in a sensible order, within 150 characters, without your marketing adjectives. That is a mechanical job with a quality ceiling, and a model does it at a thousand rows an hour without getting bored on row four hundred. Give it the fields and the rules, not the freedom.

Mapping your categories to somebody else's taxonomy. Your shop calls a category Home Fragrance. The receiving taxonomy has its own tree with its own vocabulary. Matching several hundred of your categories to the correct nodes is the sort of tedious semantic work that used to eat a week. The safeguard is to give the model the actual list of valid values and require it to choose from that list, so an invented category is structurally impossible.

Finding the contradictions between your feed and your pages. Disapprovals frequently come from conflicting data rather than missing data: a price in the feed that the page no longer shows, an availability that lagged a restock. Comparing two representations of the same product and reporting differences is a job with a correct answer, which is exactly when automation is safe. The same logic drives a product listing compliance audit run with AI, where the model checks rather than writes.

How long does a fix take to show up?

Longer than you expect, which is why the panic fix on a Friday afternoon rarely works. Google's guidance on fixing product data quality disapprovals says it can take 3 to 5 business days to review updated products, and that review requests typically take up to 3 to 7 business days.

There is a further structure worth knowing before you need it. Account level warnings come with a 28 day window during which products keep appearing but perform less well, and one courtesy review can be requested before the deadline. After a suspension, cool down periods can apply between review requests. The practical consequence is that a warning is not a soft notice to file away. It is a countdown with one free retry, and spending that retry on a half done fix is the classic mistake.

The other timing lesson is about batch size. If you change six things at once and the account comes back still disapproved, you have learned nothing about which of the six was wrong. Fix the identity fields first, then images, then the rest.

Does apparel play by different rules?

Yes, and if you sell clothing the extra requirements are the most likely cause of a disapproval you cannot explain. The disapproval guidance names gender, size, colour and age group as required for apparel and accessories, on top of everything every other product needs.

The image floor moves too. The same guidance gives a minimum of 100 by 100 pixels for most products and 250 by 250 for apparel, which sits underneath the 500 by 500 requirement coming into force in 2027. Two numbers in two documents describing the same asset is exactly how a catalogue ends up passing one check and failing another.

Colour is where apparel catalogues quietly break. A shop that lists Midnight, Ink and Navy as three separate colours has three colours as far as any machine is concerned, and a shopper filtering for blue finds none of them. The fix is a colour column that holds a plain word alongside whatever poetic name your marketing uses. Models are good at this specific mapping, because the answer is drawn from a short closed list rather than invented, and the wrong answers are obvious on inspection.

Size carries the same problem with worse consequences, since a size mismatch turns into a return rather than a missed impression. If your size column contains 10, UK10, EU38 and Medium across a single catalogue, no downstream system can group them, and neither can a shopper.

What if you do not run a feed at all?

Then your pages are the feed, and the same discipline applies with fewer excuses. Plenty of small sellers never open a shopping account, sell through their own site and social channels, and assume none of this concerns them.

It does, because the machines reading your pages want the same fields. Google's documentation notes that a merchant can become eligible for merchant listing experiences by providing product data on web pages without a Merchant Center account at all. The structured data on your product page is doing the job a feed would do, with the same requirement to state price, availability and condition explicitly rather than leaving them to be read out of a paragraph.

The practical version for a seller with no feed is short. Check that each product page emits price, availability and a stable identifier in its markup. Check that variants reference a parent rather than standing alone. Check that your images are large enough to survive the next platform that sets a floor. That is the same audit, run against a different output, and it is the audit that decides whether an assistant recommending a product to someone can quote your price correctly.

Variants are where small catalogues quietly lose

Six colours of one coat, listed as six unrelated products, is the most common structural error in a small feed, and it costs twice. The listings compete with each other for the same query, and the shopper who wanted the green one lands on the blue one and leaves.

In the feed the fix is item_group_id, which ties the variants together and has a limit of 50 alphanumeric characters. On your own pages the equivalent is structured data. Google's documentation on product variant structured data describes ProductGroup, where the group carries a productGroupID, also called the parent SKU, which must match across every related variant, and a variesBy property naming the axis of variation such as colour, size or material. The documentation also says that a multi page site should carry full self contained markup on every variant page rather than a partial reference.

Most shop platforms already know which products are siblings, because you told them when you created the variant. The information exists. It just stops at the boundary of the feed. That is a mapping problem, not a content problem, and mapping problems are the cheapest ones to fix.

A sequence that works for a catalogue under a thousand items

Export everything first, including the columns you think are empty, because knowing which are empty is half the audit. Then count. How many products have no brand, how many have a title over 150 characters, how many share a GTIN with another row, how many have images under 500 pixels on the short side. A spreadsheet answers all four in an afternoon and gives you a number to report against later.

Fix the identity layer before the prose layer. Brand, identifiers, variant grouping and category are the fields that decide whether an item competes at all. A beautiful title on a product with no group id is a nicely written listing that cannibalises five siblings.

Then rewrite titles in batches, with the rules in the prompt and the fields in the input, and read the first fifty by hand before you trust the next nine hundred. Our method for writing product descriptions with AI applies directly here, with one change: a feed title has no room for voice, so strip the personality that a description earns.

Last, put the audit on a calendar. Feeds rot. Suppliers change barcodes, somebody adds a sale badge to an image, a category gets renamed, and none of that produces an alert until the disapprovals arrive. A quarterly re export and recount takes an hour and catches almost everything. If you are building or rebuilding the shop itself rather than patching a catalogue, starting from an ecommerce website builder that generates the product schema alongside the pages saves the mapping work entirely, because the fields exist from the first day rather than being reverse engineered later.

What a clean feed is actually worth

It is not a growth tactic. It is the removal of a tax you were paying without seeing the bill. Products that were disapproved start competing, siblings stop bidding against each other, and the shopper who searched for the green coat in a size 12 finds the green coat in a size 12.

The wider point is that machines are an increasing share of your readership, and they are literal. Everything a human infers from your photography, your layout and your reputation has to be stated in a field for a machine to use it. A feed audit is the least glamorous work in a small shop and one of the few with a direct line to revenue.