Feed Completeness Is a Management Discipline

Monday morning, trading asks why the weekend sale underperformed on Google and Amazon. Shopify had 20% off from Friday. The product page looked right. Merchant Center still had Thursday's full price. The Amazon listing had no deal. Customers who landed on the site saw the offer. Assistants comparing products did not.

Most of our members live that split every week. They run a Shopify store, where Google Merchant Center and the Shopping Graph are the owned-channel facts, and they run Amazon. Same brand. Same SKUs. Two catalogues. Two machines that will increasingly select on the customer's behalf.

Leadership still hears that shopping is "handled." The feed is live. Merchant Center is green. The Amazon listing exists. Promotions live on the product page, where a person can be closed. Someone in ops or the agency keeps the pipe running.

That was rational when a click was the job. A person saw a result, landed on the PDP, and the page did the selling: strikethrough price, gift with purchase, ends Sunday. Completeness meant GTIN, category, colour, size. The feed got you the click. The website got you the basket.

Agents do not shop that way. They compare large sets of options from what the catalogue can prove: Google's Shopping Graph on the owned path, Amazon's catalogue when the assistant is the interface. If the promotion is only on your PDP, it is not in the comparison. You are not losing a close. You were never selected.

This is merchandising, not feed plumbing. Completeness is what gets you into the comparison. Miss the record and the opportunity is gone before a session shows up in analytics.

Live is not complete

"The feed is live" answers a technical question: can systems receive rows? Completeness answers a commercial one. Are the fields that matter filled and faithful, and is the current offer in the record?

Feed is live

Rows sync. Required fields pass a minimum gate. Status is Approved. Promotions still live on the PDP. Leadership hears that shopping is handled.

Feed is complete

Required and category-conditional fields are filled. Identifiers are present where they exist. Variants group cleanly. Price, availability and shipping match the page. The live offer is in the structured record, not only on the site.

A live feed with thin attributes is an empty aisle with a functioning light. A live feed with a thin offer is a full-price SKU in a comparison that is shopping a sale.

On Google that line is already explicit. Products missing required attributes do not process. Products missing recommended attributes still process, and Google's own language is that you include them to increase the chance of appearing in relevant results. The states that matter are Approved, Limited and Not approved, plus Needs attention ranked by click potential. There is no official "feed quality score." Agencies invent those. Google shows issue cards.

Identifiers remain the first completeness pillar. If a product has an assigned GTIN and you omit it, Google documents limited visibility: throttled matching, not always a hard delist. Correct GTINs are the only completeness-adjacent lift Google still publishes in Help, a 20% increase in clicks on average. Apparel in the UK makes the category rule concrete: colour, size, gender and age group are required for clothing and shoes. That is still necessary. It is no longer sufficient.

"The feed is live" is not a KPI. Completeness of identity, category fields, faithfulness and the offer the agent will compare is.

The offer that never leaves the PDP

Picture the habit. Friday afternoon. Merch launches twenty per cent off, ends Sunday. The PDP updates in minutes. Strikethrough, banner, urgency copy. Merchant Center still shows full price. The Amazon deal console never got the coupon. Everyone in the room nods because this has worked for years. A human who clicked through from Google Shopping would see the offer and convert.

That split breaks when the shopper is an agent working a graph. Google's surfaces (free listings, Shopping ads, AI Mode, Gemini) compare what you submitted. Amazon's assistant compares what the catalogue supports: attributes, price history, deals, availability. The machine is shopping options, not browsing a page. Your trainer sits in the comparison at full price while a competitor's sale price is in the record.

You do not get a session to recover. There is no "they'll see it when they land." They were never picked.

Google already has the objects. sale_price and sale_price_effective_date put a dated discount in the product record. Promotion IDs map products to a promotions feed, and the UK is one of the markets where that mapping is required if you run promotions. None of that fires from PDP copy a crawler might notice later. It fires when you merchandise the offer into the feed on the same SLA as price and stock.

Amazon is the same discipline in a different console: deals, coupons, badges, facts the assistant can cite. The listing that is good enough for a human scroller is not the listing an agent uses to choose among fifty substitutes.

Where the record contradicts the site

Google continuously reconciles what you submit with what it can fetch: structured data on the product page, live price and stock, shipping, crawl access. Amazon reconciles Buy Box, landed price and deal eligibility against its own catalogue rules. A feed can be live while that loop is broken: warnings, preemptive item disapproval on price or availability mismatch, a deal that does not match the PDP, or silent overrides via Google automations, which are on by default.

Four failure modes keep turning up. None of them need a new AI programme to see.

Identifier gap. Hero SKUs are Approved. Barcodes never left the supplier file. Matching is limited. Sessions on the site still look fine.

Price or stock mismatch. The site moved this morning. The feed did not. Preemptive disapproval hits the SKU.

Offer only on the page. The campaign is live. Merchant Center still shows full price. Amazon has no deal. Humans who land convert. Agents comparing the graph never pick you.

Variant fracture. Colourways lack a shared group identity. Machines treat siblings as unrelated products.

Titles still matter. Titles without structured attributes are how you get rewritten and still fail filters. Completeness is both.

Taxonomy (wrong Google Shopping category, title tricks that drift the map) is the sibling problem. That is Comparing Apples with Oranges. This piece owns coverage, identifiers, faithfulness, the offer, and ownership.

If the CMS cannot carry a missing attribute this week, patch it through a supplemental source without changing the product ID. Change the id and you wipe history. The overlay cannot join.

Who owns catalogue truth

In too many organisations nobody owns this. Merchandising owns the human story. Technology owns the connector. Growth owns media. The Amazon team owns the listing. The agency owns "shopping." Google's automations own a silent patch. When the weekend sale never reached the feed, detection is accidental. Growth checks spend. Ecommerce checks the PDP. The Amazon team checks the listing. The green tiles all looked fine.

Give one merchandiser responsibility for the range, offer and product facts that machines receive. Review it in the same forum that looks at margin and stock. Start with the SKUs that fund the month. Check Needs attention, price and stock mismatches, missing identifiers, and promotions that are live on Shopify but absent from Google or Amazon.

Amazon's own assistant is already shopping that catalogue. In May 2026 Amazon introduced Alexa for Shopping to US customers: questions in the main search bar, comparisons, a year of price history, deal-finding and cart-building. UK founders should not pretend the search bar on Amazon.co.uk already works that way. They should also not wait. Incomplete attributes, stale price and missing deals are how you lose a shortlist you never see.

You already review margin by SKU because averages hide weak products. Feed gaps deserve the same scrutiny. "Are we approved?" is a comfort metric. The uglier question is what share of SKUs in the top categories carry the attributes, identifiers and live offers this category needs. Private-label is the honest exception: if there is no GS1 barcode, do not invent one. Completeness there is attributes, variants and a truthful offer.

Before you approve the next shopping or Amazon work package, ask for coverage by category, identifier honesty, and whether the current offer is in the record on the same SLA as price and stock. You are not buying a pipe. You are buying merchandising that machines can use.

You do not need Universal Cart in London, or Alexa in the Amazon.co.uk bar, to justify this. Free listings, paid Shopping and every agent that already compares structured offers will punish a catalogue that still merchandises only for the human who lands. For UK operators the local fires are already here: limited performance on missing identifiers, apparel size and colour, VAT-inclusive price matching the page, shipping completeness, and promotion IDs if you run promotions.

Five questions for the ELT

1. Who owns catalogue truth? Not the website. Not "the feed." Coverage, faithfulness, and whether the live offer is in the record, in the same forum that reviews margin and stock.
2. What share of SKUs in our top categories carry the identifiers, attributes and live offers this category actually competes on? Which lines are structurally thin even when the PDPs look strong?
3. When a time-limited promotion goes live on the site, is it in Merchant Center and the Amazon catalogue for the same window? What happens when it is not?
4. What is our SLA for price, stock and offer sync across Shopify, Google and Amazon, and who gets paged when they diverge?
5. When listings show full price while the site is on sale, or an agent compares us without our promotion, who finds out first?

If those answers are vague, you do not have a feed problem. You have a merchandising problem the board is misreading as a green Merchant Center tile.

Brands that stock the machine shelf with the same care they give the PDP stay in the comparison as discovery splits between humans and agents.