Guide Catalog operations
Build or buy catalog AI: compare on the full workflow.
Building catalog AI in-house or on a model API is a credible choice. The decision depends on the whole workflow, not one model call: product, variant and offer identity, moderation and review, partial and failed states, language quality and the maintenance that follows. Compare any option, Youzu included, on the same sample of your data.
Published by Youzu · Editorial policy
- Published
- Reading time
- 10 min read
Key points
- Compare build, model API and vendor options on the full workflow, not on one model call.
- Use the 8 criteria to see where each option is strong for your catalog.
- Check identity, variants, offers, moderation, review, failure states and maintenance before deciding.
- Use one sample, one set of labels and the same reviewers for every option.
- Keep budget and people for the work that starts after launch.
In this guide
When building wins
Building catalog AI in-house is a credible choice, and for some teams it is the right one. It tends to win when the catalog work is a core part of what makes the business different, when the team already runs machine learning in production, and when the data, labels and review tooling exist or are worth building anyway.
- Your categories or attributes are unusual enough that no general tool models them well.
- You have engineers who can own models, evaluation and review tooling for years, not months.
- You already hold labeled data and a review team that can produce more.
- Your integration surface is internal, so there is no vendor boundary to cross.
- Your data rules require that all processing stays inside systems you operate.
Building also loses in predictable ways: the first model call works on a demo, and the rest of the workflow, from variant identity to failure handling, turns out to be most of the work.
Model APIs against a vendor
General model APIs have made the first step much easier. They can return output that follows a JSON schema you supply, so required keys are present and allowed values are respected, which makes a structured attribute record far easier to produce than it used to be.[1][2]
A model API gives you a capable model and leaves the workflow to you. A vendor of catalog tooling gives you a workflow built around models and asks you to fit your process to it. Neither is better in general. The useful question is which parts of the workflow you want to own, and what each part costs you to build, run and keep current.
Model APIs also change under you. A new model version can improve most records and still change the ones you care about, so a model update is a change to your catalog workflow and needs the same regression check as a rule change.
The worked example below shows the gap on one record. A single model call can turn the seller title 'jkt olive' into 'Olive jacket'. A usable catalog record needs 9 checks to pass, and most of them are about identity, review and rules rather than text.
Worked example
9 checks for one record
| Check | What it depends on |
|---|---|
| 1. Title is plain | Text generation from the title and the photo |
| 2. Color is an allowed value | Your list of allowed values |
| 3. Material is set or sent to review | Evidence, and a review route when sources disagree |
| 4. Style is set or sent to review | Evidence, and a review route when it is missing |
| 5. Category is a leaf of your taxonomy | Your category tree, down to the leaf |
| 6. Rows are grouped into one product | Variant identity across loose rows |
| 7. Offers belong to one product | Offer matching across sellers |
| 8. A duplicate decision is made | Duplicate matching against the catalog |
| 9. Your rules were run | Your policy rules, versioned and tested |
Worksheet
Cost the whole workflow, not the first model call
- people1 to 50.
- months1 to 36.
- hours a month0 to 400.
18 person-months to build, then 480 hours a year to keep it running.
2 of 6 parts in your estimate.
Compare Youzu with in-house builds, model APIs, PIMs and search tools on the full workflow using your data: identity, variants and offers, moderation, review, partial and failed states, and maintenance.
Read as text
8 criteria in a decision table
Judge each option in words, not numbers: what it covers, what you would have to build, and what you would have to keep running. The table lists the 8 criteria and the question behind each.
| Criterion | Question to answer for each option | Evidence to ask for |
|---|---|---|
| Output quality | Is the output right on your records, including the difficult ones? | Results on your sample against your labels |
| Identity | Are products, variants, offers and duplicates kept apart correctly? | Grouping and matching decisions on real listings |
| Review | Can people see, accept, edit and reject each proposal? | The review screens or API, used by your reviewers |
| Failure states | What happens when a step is partial, failed or skipped? | Examples of partial and failed results |
| Policy | Can your rules run before listings reach search? | A rule tested on a reference set |
| Integration | How do records get in and out of your systems? | The documented API and file routes |
| Maintenance | Who keeps models, prompts, rules and tests current? | A named owner and a written routine |
| Total cost | What does it cost to build, run and keep current? | Your own estimate on your own data |
Compare total cost on your own data. Model calls are only one line of that cost; review time, engineering ownership and the cost of wrong records reaching shoppers are usually larger.
The workflow checklist: identity, variants and offers, moderation, review, failure states, maintenance
Most build-or-buy decisions go wrong because the comparison stops at the first step. Use this checklist to make every option answer the same questions.
- Identity: each product keeps one identifier across sources and updates, and a re-sent record updates the right product.
- Variants and offers: options of one product are grouped as product variants (one purchasable version of a product that differs along defined dimensions, such as size, color or pack size), and each seller's terms stay on their own offer.
- Moderation: policy rules are versioned, tested before activation and applied before listings reach search.
- Review: proposals wait for a person when the evidence is unclear, and the original value stays beside the proposed one.
- Failure states: partial and failed results are visible per record and per step, and nothing half-finished is published silently.
- Maintenance: someone owns the models, prompts, rules, category changes and the tests that catch regressions.
The same-sample method
The fairest comparison runs every option on the same records and judges them the same way. It is also the cheapest way to avoid a decision based on demos.
- Choose the sample. Take real records from the workflow you care about, at random, and add the difficult ones on purpose: catch-all categories, near duplicates, missing images, mixed languages.
- Write the labels first. Let your team decide what correct looks like for each record before any option runs, including which fields should stay empty.
- Run every option. The current manual process, an in-house prototype or prompt your team writes, and each vendor, all on the same sample.
- Judge blind where you can. Reviewers should not know which option produced which record.
- Report errors by kind. Measure accuracy on your own sample against your own labels, with critical errors reported separately, and list the records each option could not handle.
Compare Youzu with in-house builds, model APIs, PIMs and search tools on the full workflow using your data: identity, variants and offers, moderation, review, partial and failed states, and maintenance.
What changes after launch
The decision does not end at launch. Categories change, sellers find new ways to describe products, channels change their requirements and models are updated. Whichever option you choose, plan for these from the start:
- A regression sample that is re-run after any change to models, prompts, rules or the category tree.
- A review queue that stays staffed, because uncertain cases never reach zero.
- A route for sellers and reviewers to report wrong records, and an owner who acts on it.
- A record of which version of each rule and model produced each decision.
If you buy, ask how each of these works and who does it. If you build, budget for them as ongoing work, not as part of the first release.
Ownership decides more of these outcomes than technology does. A build with a named team that owns it for years can serve a catalog well, and so can a purchase with clear responsibilities on both sides. What fails is a system nobody operates after the project that delivered it ends.
Boundaries
What Youzu does not do
- Youzu publishes no accuracy figures. Measure accuracy on your own sample against your own labels, with critical errors reported separately.
- Youzu makes no cost claims against other approaches. Compare total cost on your own data.
- Youzu does not replace your ERP, search or storefront; those systems stay authoritative for their own data. It can replace your PIM, or work alongside it.
Checklist
Build, API or vendor comparison sheet
The 8 criteria and the workflow checklist as rows, with one column per option for your notes in words. It has no weights and no totals.
A CSV file of 10 rows, named youzu-build-vs-buy-comparison-sheet.csv. No form to fill in first.
| Criterion | Question | In-house build | Model API | Vendor |
|---|---|---|---|---|
| Output quality | Right on your sample against your labels? | |||
| Identity | Products, variants, offers and duplicates kept apart? | |||
| Variants and offers | Options grouped; seller terms on each offer? | |||
| Moderation | Rules versioned, tested and applied before search? | |||
| Review | Proposals reviewable with originals kept? | |||
| Failure states | Partial and failed results visible per record? | |||
| Integration | Documented routes in and out of your systems? | |||
| Language quality | Reviewed by readers of each language? | |||
| Maintenance | Named owner for models, prompts, rules and tests? | |||
| Total cost | Build, run and keep-current cost on your data? |
FAQ
Questions buyers ask
Short answers to the questions teams ask about this topic.
Yes, for teams that already run machine learning in production and can own evaluation, review tooling and maintenance over years. The first model call is rarely the hard part. Identity, variants and offers, moderation, failure handling and keeping everything current are where most of the effort goes, so estimate those before deciding.
It can be enough for one step, such as drafting titles or extracting attributes into a schema. A catalog workflow also needs matching, variant grouping, review queues, policy rules and handling of partial results. Decide whether you want to build and maintain those around the API, or adopt them from a tool.
Use the same-sample method: one realistic sample with difficult records added on purpose, labels written before any option runs, the same reviewers for every option and blind judging where possible. Measure accuracy on your own sample against your own labels, with critical errors reported separately.
Compare total cost on your own data. Include model or license costs, engineering time to build and maintain, review time per record and the cost of wrong records reaching shoppers or channels. Per-call prices alone hide most of the difference, because the work around each call is usually larger than the call.
Compare Youzu with in-house builds, model APIs, PIMs and search tools on the full workflow using your data: identity, variants and offers, moderation, review, partial and failed states, and maintenance. Evaluations start with a sample of your own catalog and expand to more workflows once the results are reviewed.
Last reviewed
Sources
Structured model outputs, OpenAI API documentation. developers.openai.com/api/docs/guides/structured-outputs. Accessed . Back
Structured outputs, Gemini API documentation. ai.google.dev/gemini-api/docs/structured-output. Accessed . Back
About this guide
Published by Youzu · Editorial policy
- Published
Every guide defines its terms, works through one example and cites the public sources it relies on, with the date each source was checked.
Guides explain methods. They do not rank Youzu above other approaches, and an example is never presented as a customer result.
More in all guides and the catalog glossary.
Related
Keep reading
Product
Catalog Intelligence for marketplaces and retailers
Catalog Intelligence maps products to your category tree, extracts category attributes, keeps variants and offers apart and flags issues for review.
Evaluation
Benchmark Youzu on a sample of your own catalog
We run a sample of your catalog and report results against your own labels. You set the acceptance criteria; critical errors are counted separately.
Guide
PIM vs catalog intelligence: which do you need?
A PIM stores product information you have authored; catalog intelligence builds the structured record from raw feeds and images. When to use each.