Catalog onboarding at scale: review thousands of SKUs a day

Every product that goes live on your platform passes through a gate. Someone — or something — has to decide the listing is in the right category, that the images aren't lifted from a competitor, that the title and photo meet a quality bar, and that nothing about it violates policy. Only then can the SKU sell. That gate is not optional. It's also, at scale, the single most reliable place for growth to stall.

The math is brutal. A marketplace onboarding thousands of products a day, each needing classification, a copyright check, and a quality pass, does not have a review problem — it has a review throughput problem. And the cost of falling behind isn't abstract. On enterprise marketplace platforms, the average time from a seller signing up to their first sale is 25 days (Mirakl Seller Report, 2023, across 65,000 sellers analyzed). A large chunk of those 25 days is the SKU sitting in a review queue, not selling.

Getir's catalog queue was drowning a 25-person team

The clearest version of this problem we've worked on ran at Getir. Ninety thousand restaurants. Six thousand product images reviewed daily. And a 25-person manual QA team that was, structurally, the thing standing between a restaurant and its first order.

The sequence was familiar. A restaurant uploaded its products. Those products landed in a queue. A reviewer classified each one, checked whether the images were the restaurant's own or lifted from somewhere else, assessed whether the photo and copy were good enough to put in front of a customer, and approved or kicked it back. At roughly 90 seconds a task and 6,000 images a day, the arithmetic wrote itself: the team was always behind, and a restaurant that signed up on Monday might not be taking orders until Thursday. Every hour of that delay was GMV that never happened.

The team wasn't slow. The work was just too much of the same thing. Classification, copyright, quality — over and over, thousands of times a day.

The flow that turned 90 seconds into 1.5

We mapped Getir's catalog process and built a four-stage flow that runs on every product the moment it's uploaded. The agent moves optimistically through all four stages; a human confirms in a batch at the end rather than touching each SKU individually.

The change in throughput was not incremental. Time per task dropped from 90 seconds to 1.5 seconds. The team that had been 25 people doing manual review became 5 people running the flow and owning the exceptions. And because the quality step actively improved listings rather than just passing them, photo coverage went up roughly 30% — the catalog got better, not just faster.

The business impact stacked from there: on the order of €150K a year in savings from the smaller team, and €1M+ in upside from restaurants going live the same day instead of three days later. That second number is the one that matters. Faster review didn't just cut cost — it turned "live in three days" into "live today," and every one of those recovered days was orders that would otherwise have been lost. The full breakdown is in how Getir handles 200K product onboardings a month with 5 people.

Run the numbers against your own catalog

Getir's shape is not unique to grocery. Any platform gating SKUs through classification, copyright, and quality review has the same structure — a large routine tail that a flow clears, and a smaller set of genuine exceptions that need a person. Move the slider to your own daily onboarding volume and see how the queue splits.

The exact split depends on your catalog and your policies, but the pattern is consistent: the overwhelming majority of listings are unambiguous — right category, original images, acceptable quality — and clear without a human. The minority that are genuinely uncertain — an odd classification, a possible copyright match, a listing that's borderline on quality — are exactly where you want a reviewer spending their attention. That's the inversion. Your QA team stops being a bottleneck on every SKU and becomes a decision-maker on the few that need one.

Why "batch approval" is the part that scales

The detail that makes this work at volume is the last stage: batch approval. In a manual process, a reviewer opens one product, decides, closes it, opens the next — a serial grind where throughput is capped by human clicks. In the flow, the reviewer sees the results for many products at once: classifications made, copyright checks passed, quality assessed, exceptions surfaced. They confirm the clean ones as a group and drill into only the flagged ones.

This is what lets five people do the work of twenty-five without cutting corners. They're not reviewing faster — they're reviewing less, because the flow already did the assembly and only the genuine questions reach them. A reviewer's attention goes to the copyright match that needs a human eye, not to the four thousand listings that were obviously fine.

The human stays on the calls that carry risk

Catalog review has real stakes hiding inside the routine. A copyright miss can mean a takedown or a legal complaint. A miscategorized product is invisible to the customers searching for it. A low-quality listing drags down the whole storefront's credibility. You cannot hand those judgments wholesale to a model with no one accountable.

So the flow doesn't. The agent handles the clear cases — the products where category, originality, and quality are unambiguous — and every genuinely uncertain case routes to a person. The reviewer opens an exception with the classification, the copyright result, and the quality flags already assembled, and makes the call in seconds. Every decision is logged and attributable, so when a seller disputes a rejection or a rights-holder files a complaint, there's a clean record of what was checked and who decided.

That's the model that lets a platform onboard at scale without either drowning its QA team or letting the gate go unmanned. The agent does the volume. The human owns the last call. And the catalog stays trustworthy while it grows.

What this looks like for your onboarding queue

If your catalog queue looks like Getir's did — thousands of SKUs a day, a review team that's always behind, sellers waiting days to go live while their products sit unapproved — the playbook transfers directly. Classification, copyright, and quality are the same three checks whether you're onboarding restaurant menus, marketplace listings, or a brand's product feed. Our retail solutions page covers catalog activation alongside the support and order flows that surround it.

The setup connects to the systems you already run your catalog through — no rip-and-replace of your PIM or your review tooling. The flow slots in ahead of your reviewers and hands them exceptions instead of everything.

The fastest way to know if it fits is to run one real batch of your own listings through a mapped flow: your actual SKUs, your actual policies, and an honest read on how much would clear automatically and how much your team would still own.