Lorixen.
EN

Cleaning Up Product Titles, Tags and Attributes With AI

Short answer: AI is good at the boring middle of a catalogue clean-up: proposing consistent titles, pulling colour or material out of messy descriptions, and mapping old tags to a tidy list. It is bad at deciding what your rules should be and at knowing facts that aren't in the data. Write the rules yourself, run the AI over an export in a spreadsheet, review every change, and only then import.

Most small stores don't plan their product data. Products get added over several years by different people, from supplier sheets, in a hurry before a market day. The result is titles in three styles, tags that mean the same thing spelled four ways, and filters that don't work. Here is how a clean-up runs, using a hypothetical shop.

The example: a hypothetical Nelson homewares store

Imagine a small online store in Nelson selling about 600 products: ceramics, linen, candles and a few kitchen tools. Its product data looks like this:

Problem What it looks like
Inconsistent titles "Mug - speckled blue", "Hand thrown mug (Blue Speckle) 350ml", "BLUE MUG"
Duplicate tags linen, Linen, linnen, 100%-linen
Facts buried in descriptions Capacity and dimensions only appear in body text
Variant details missing Colour is in the title but not set as a variant option

The owner wants three things: titles that read the same way across the store, a short tag list that powers collection filters, and key specs stored as fields rather than prose.

Step 1: Audit before touching anything

Export the full product list as a spreadsheet. Then count, don't guess:

  • How many distinct tags exist, and how many are used once?
  • Which product types have no consistent title pattern?
  • Which specs (capacity, size, material) appear in descriptions but nowhere else?

In this example the audit finds about 400 distinct tags, most used on only a handful of products. That number alone justifies the work.

Keep the original export untouched. Every later step works on a copy, so you can always compare or roll back.

Step 2: Write the rules yourself

This is the step owners try to hand to AI, and it is the one that has to stay human. The rules reflect how your customers shop.

A title pattern per product type. For the example store:

  • Ceramics: Item – Glaze/Colour – Size ("Hand-thrown Mug – Speckled Blue – 350 ml")
  • Linen: Item – Colour – Size ("Linen Napkin Set – Sage – Set of 4")

Google Merchant Center's title guidance (as of November 2026) supports this kind of discipline: titles should clearly describe the product on the landing page, distinguish variants (for example adding the colour), be specific and use correct grammar, within a limit of 150 characters. Even if you don't sell through Google Shopping, those are sensible rules for your own store.

A controlled tag list. Decide the 30 to 60 tags that actually drive filters or collections, and write the exact spelling. Everything else gets mapped to one of them or deleted.

Which facts become fields. Capacity, dimensions and material are better stored as structured fields than as text. On Shopify, these are metafields: Shopify's developer documentation describes a metafield as an extra field on an existing resource such as a product, and says a defined metafield gets a typed editor, validation and support for search and filtering, while an undefined one is stored as plain text. Shopify also publishes a standard product taxonomy of categories, attributes and values that is used across the platform and with marketplaces, so check whether an attribute you need already exists there before inventing your own. The set-up side is covered in using metafields for product specs.

Step 3: Let AI propose changes in the spreadsheet

Now the AI earns its place. For each row, give it the current title, description, tags and product type, plus your written rules, and ask for:

  1. A proposed title following the pattern for that product type
  2. Tags mapped to your controlled list (and nothing outside it)
  3. Extracted specs, such as capacity or dimensions, only if they appear in the text
  4. A confidence flag or a note when something is missing

If you or a developer are calling a model through an API, ask for structured output. OpenAI's Structured Outputs feature, for example, makes responses follow a JSON schema you supply, so a required field isn't silently skipped and an allowed-values list (your tags) isn't ignored. That turns a chatty answer into columns you can paste straight back into the sheet.

Two settings keep it honest:

  • No invention. Tell the model to leave a field empty when the information isn't in the source text. A blank cell is easy to spot; a plausible made-up capacity isn't.
  • Small batches first. Run 20 products, check them, adjust the instructions, then run the rest.

Step 4: Review like an editor

Add columns next to each proposal: original, proposed, approved (Y/N), edited value. Then:

  • Sort by product type and scan titles in a block; inconsistencies jump out.
  • Filter for empty spec cells and decide whether to fill them from supplier sheets or leave them.
  • Check every extracted number against the description it came from.

In the example store, the AI handles the speckled-mug variations well but tags a beeswax candle as "linen" because the description mentions a linen gift bag. That is a typical error: it read the words correctly but misunderstood what the product is. Human review catches it in seconds.

Step 5: Import carefully

  • Import a small batch first and check how the titles look on collection pages and on mobile.
  • Confirm filters now show the expected options.
  • Watch for changed URLs: changing a title shouldn't change a product's address on most platforms, but check your platform's settings, and redirect any handle you change on purpose.

Step 6: Stop it getting messy again

A clean catalogue decays within months unless adding products follows the same rules:

  • Put the title pattern and tag list in a one-page guide for whoever adds products.
  • Use defined fields with validation instead of free text where your platform allows it.
  • Repeat a short audit every quarter: new tags, titles that break the pattern.

FAQ

Can AI rewrite my whole catalogue in one go?

Technically yes, but it is a poor idea. Without written rules and human review you just swap one inconsistency for another, and invented details can slip in.

Will changing product titles hurt my search rankings?

Clearer, more specific titles usually help shoppers and search engines understand the product. The risk is changed URLs; keep product addresses stable or redirect them.

How many tags should a small store have?

Enough to power the filters and collections customers actually use. For a few hundred products, a few dozen well-defined tags usually beats hundreds of one-off ones.

Should specs live in tags or in fields?

Fields, where your platform supports them. A capacity or dimension stored as a typed field can be validated and filtered; a tag like "350ml" can't.

Got a catalogue that has grown in five directions at once? Send us an export and the rules you'd like, and we'll show you what a clean-up would involve.

Sources

WhatsAppLet’s talk