The product catalogue is where a lot of eCommerce businesses lose money without noticing. Not because the site looks bad. Because products are badly tagged, filters are missing, descriptions are weak, data isn't normalised, and the advertising feed inherits all of that mess.
When a catalogue grows, manual management becomes fragile. Today you have 2,000 products. Tomorrow you have 8,000. Then come new brands, variants, formats, categories, attributes, out-of-stock products, incomplete descriptions, duplicate pages and filters that also need applying to future products.
At that point "tidying up a few product pages" is no longer enough. You need a system.
An AI agent applied to an eCommerce catalogue can become a kind of catalogue manager: it reads products, categories, brands, descriptions, filters, reviews, customer questions, feeds and SEO data. Then it proposes tags, flags errors, finds duplicates, suggests filters, spots thin descriptions and helps turn the catalogue into a stronger commercial structure.
The important distinction is this: the AI must not invent. It must help you check, connect and improve real data.
The problem: a dirty catalogue doesn't stay in the catalogue
A disorganised catalogue doesn't just annoy the back office. It spreads everywhere.
- Internal search finds fewer products.
- Filters don't show everything they should.
- Categories look thin even when the catalogue is rich.
- Filter pages built for SEO don't capture useful keywords.
- The Merchant Center feed inherits weak titles and attributes.
- Shopping and Performance Max campaigns learn on worse data.
- Product pages don't answer the customer's real doubts.
- Search Console sees pages that are similar, thin or of little use.
The catalogue is the engine of an eCommerce business. If the data is dirty, every other function works worse: SEO, UX, Ads, customer care, email marketing, internal search and chatbots.
What AI can genuinely do to a catalogue
A useful AI system does not stop at writing descriptions. That is only one piece of the job.
The most interesting areas are:
- automatic product tagging;
- filters that are missing or badly applied;
- normalising brands, formats, colours, materials and attributes;
- analysing thin or duplicated descriptions;
- identifying similar or duplicate products;
- checking for wrong categories;
- flagging data missing for SEO, feeds and product pages;
- drafting SEO copy and product page content;
- suggesting new filter pages or categories;
- supporting internal search and synonyms;
- quality control before publication.
The gain isn't magic. It's less repetitive work and more consistency.
Tags and filters are not the same thing
Before talking about automation, you need to separate tags from filters.
A tag can be an internal or semantic piece of information: "summer product", "high margin", "suitable for beginners", "natural ingredients", "premium line", "sensitive", "potential bestseller".
A filter, by contrast, is a navigation function: brand, colour, format, size, need, material, price band, product type, compatibility, availability.
Some filters can also become genuine SEO pages, with a dedicated URL, title, H1, copy and FAQ. But not all of them.
AI can help at both levels:
- it suggests internal tags that are useful for management and marketing;
- it proposes navigable filters to improve UX and conversion;
- it identifies filters with SEO potential;
- it finds products that should have a filter but don't;
- it merges duplicate or overly similar filters.
This reinforces the work we described in the article on eCommerce filters, AI and automation: the real step change comes when filters are not just created by hand, but governed by rules, data and checks.
The classic case: filters missing on products that should have them
Anyone who runs an eCommerce site knows this problem well.
You add a new filter. For example:
- grain free;
- sugar free;
- sensitive skin;
- MagSafe compatible;
- organic cotton;
- high protection;
- fast charging;
- vegan;
- made in Italy.
The problem isn't creating the filter. The problem is applying it correctly to every current product and remembering to apply it to future ones.
If you have to do it by hand, you will get it wrong. Not through incompetence, but because of volume. Some products get left out. Some are tagged badly. Some suppliers change their descriptions. Some brands use different words. Six months later the filter no longer really represents the catalogue.
An AI agent can cross-reference title, description, brand, category, technical attributes and reviews to propose which products should receive that filter. Then a rule can apply it in bulk, and to new products as they arrive in future.
Automation: existing and future products
The most useful part isn't just "select these 300 products". The useful part is: "from now on, when a product with these characteristics arrives, apply this filter".
For example:
- if the title or description contains certain terms;
- if the brand belongs to a given list;
- if the category is consistent;
- if the technical attributes confirm the characteristic;
- if the AI assigns high confidence;
- if there are no contrary signals.
That way the filter doesn't depend on the memory of whoever uploads the products. It depends on rules you can check.
The healthy flow is:
- the AI proposes;
- the system shows the reasoning and the confidence;
- an operator approves or corrects;
- the rule is saved;
- future products are checked automatically;
- exceptions are flagged.
This is different from letting the AI change whatever it likes. Here the AI speeds up the work, but stays inside a governed process.
Product descriptions: generating text isn't enough
Plenty of people talk about AI for eCommerce as if the point were producing descriptions in bulk.
The risk is creating thousands of pages that read nicely but are unreliable, repetitive or simply wrong.
Good catalogue AI has to do quality control first and writing second, if at all.
It should ask itself:
- does the description contain real information?
- are materials, format, size, use and compatibility present?
- are there clear benefits for the customer?
- does the page answer the frequent doubts?
- is the copy duplicated from other products?
- are there promises that can't be verified?
- is data missing that the feed or the structured data needs?
- is the description consistent with the category and the filters?
Only then can it generate a useful draft: a summary, strengths, recommended use, FAQ, differences from similar products, important warnings and content for the SEO side.
This connects to the article on AI product pages with reviews, FAQs and customer questions: the best product page grows out of the product's real data and users' real doubts, not out of a generic description.
Normalisation: the invisible work that sells
Normalisation is boring. That is exactly why it is valuable.
In a real catalogue you will find:
- the same brand written three different ways;
- different colour names used for the same colour;
- sizes expressed in different formats;
- materials written inconsistently;
- product variants that aren't linked;
- missing EANs or GTINs;
- wrong categories;
- descriptions using supplier jargon that means nothing to the customer;
- useful attributes present in the copy but not in filterable fields.
These problems look small. But added together they degrade filters, internal search, feeds, product pages and category pages.
AI can read the catalogue and propose normalisations. For instance: "navy blue", "dark blue" and "navy" can become a single filter value; "1000 ml", "1 litre" and "1L" can be interpreted the same way; two brands written with different capitalisation can be merged.
Normalisation isn't cosmetic. It's structure.
Duplicates and near-identical products
Another serious problem is duplicates.
We aren't only talking about two identical products. We're also talking about near-identical products, unlinked variants and pages that cannibalise each other.
AI can flag cases such as:
- the same product code with different titles;
- the same GTIN on two pages;
- the same image and an almost identical description;
- a size or colour variant uploaded as a separate product with no relationship;
- an old product and a new model both live with no redirect or link between them;
- similar pages that should be merged or differentiated better.
This has an SEO impact. If Google finds a lot of near-identical pages, with thin content or duplicated products, the catalogue becomes harder to interpret.
We went deeper into this in the article on eCommerce canonicals for variants, filters and duplicate products.
AI and the Merchant Center feed
The Google Merchant Center feed is one of the places where a dirty catalogue immediately becomes expensive.
If titles, descriptions, brands, GTINs, availability, prices, categories and attributes are weak, Shopping and Performance Max campaigns work worse too.
AI can help identify:
- product titles that are too generic;
- missing attributes;
- inconsistent brands;
- wrong product categories;
- descriptions that are too short;
- availability that doesn't match the page;
- products with unstructured variants;
- products that deserve custom labels or commercial segments.
That doesn't mean the AI should send the feed on its own. It means it can run a continuous check and flag where to intervene.
If the feed is a live issue for you, there's a deeper look in the article on Google Merchant Center feed errors that block sales and campaigns.
From internal search to new SEO pages
The catalogue doesn't live on its own. It has to talk to internal search.
If users often search for terms that don't match the current filters, AI can highlight an opportunity:
- a new synonym to add;
- a new filter to create;
- a new commercial category;
- a new SEO filter page;
- a new FAQ in a category;
- a new content block on the product page.
This is the bridge between catalogue AI and SEO: don't generate pages at random, create them when the data shows a real intent.
We covered it in the guide to eCommerce internal search with synonyms, typos and AI: the queries users type are a concrete source for improving filters, pages and navigation.
Quality control before publication
One very useful function is quality control before publishing or importing products.
Before a product goes live, the system can check for:
- a title that is too short or too generic;
- a missing description;
- images that are absent or too few;
- a brand that isn't normalised;
- an inconsistent category;
- missing mandatory filters;
- feed-relevant attributes left empty;
- inconsistent price or availability;
- probable duplicates;
- copy that is too similar to other products.
This stops the site slowly filling up with problems. Because a catalogue doesn't get dirty in a day: it gets dirty one product page at a time.
Yes to AI, but with rules and checks
AI applied to a catalogue is powerful, but it has to be designed properly.
There are three risks to avoid:
- hallucinations: the AI invents attributes, benefits or compatibility that aren't in the data;
- blind automation: the system changes too many products without review;
- flat content: thousands of similar descriptions, formally correct but not much use.
That's why the best flow is:
- real data as the basis;
- clear rules;
- a confidence score;
- human approval in the delicate cases;
- a change history;
- rollback;
- spot checks;
- priority on the products that matter most.
AI shouldn't replace commercial judgement. It should make it scalable.
Which products should you check first?
You don't have to start with the whole catalogue.
The most sensible priorities are:
- products that generate the most revenue;
- products with many impressions but few sales;
- products that matter to Google Ads campaigns;
- strategic SEO categories;
- high-margin products;
- pages with short descriptions;
- products without key filters;
- products that show up often in internal searches;
- the main brands;
- newly imported products.
That makes the work more concrete: first fix where the economic impact is greatest, then widen the checks.
How we build this into a custom eCommerce site
In our custom eCommerce builds, AI can work on the catalogue to help the people running products, SEO and marketing.
The system can:
- read every product and its attributes;
- propose tags and filters;
- apply rules in bulk;
- check future products;
- flag weak descriptions;
- create content drafts;
- find duplicates;
- connect internal search, filters and SEO pages;
- support the Merchant Center feed;
- feed chatbots and product pages.
The interesting part isn't "let's add AI because it's fashionable". It's using AI to take friction out of work that would otherwise become unmanageable.
This is also where our AI and automation for digital processes service comes in: the goal isn't a toy, it's a system that saves time and creates a competitive advantage.
Why this has SEO value
A clean catalogue helps SEO because it makes the structure of the site clearer.
With better data you can have:
- more complete categories;
- more useful filters;
- stronger canonical filter pages;
- more informative product pages;
- fewer duplicates;
- more consistent internal linking;
- more reliable structured data;
- a cleaner advertising feed;
- more effective internal search;
- a better user experience.
Google doesn't reward a site for "using AI". It rewards useful pages, an understandable structure and content that helps the user. AI is worth it if it improves those things.
For categories, the reasoning also links back to the guide on eCommerce category page SEO: if the catalogue is properly governed, categories become far more solid pages.
Checklist: what a useful AI catalogue manager has to do
- Understand the existing categories, brands, filters and attributes.
- Propose tags with an explanation and a confidence level.
- Find filters missing on products.
- Apply rules to future products as well.
- Merge duplicate values or synonyms.
- Flag thin, duplicated or inconsistent descriptions.
- Identify similar or duplicate products.
- Check the fields the Merchant Center feed needs.
- Use internal searches to suggest new filters or pages.
- Create drafts, not invented content.
- Keep a history of changes.
- Leave human review where it is needed.
If a system only writes automatic descriptions, it isn't a catalogue manager. It's a text generator.
Useful Google sources
For this article we cross-referenced Google's official guidance on Product structured data, the Merchant Center product data specification, helpful, reliable content and eCommerce site structure.
FAQ: AI for an eCommerce catalogue
Can AI tag products automatically?
Yes, it can propose tags by reading title, description, brand, category, attributes and other data. The best flow includes a confidence level, an explanation and human review for the delicate cases.
Can AI apply filters to future products?
Yes, if it is turned into rules you can check. For example, the system can apply a filter when title, description, brand and category all confirm a characteristic.
Is generating product descriptions with AI safe for SEO?
Yes, if the AI starts from real data and the copy is checked. The risk arises when the AI invents benefits, compatibility or details that aren't on the product page.
What is the difference between tags and filters?
Tags can be internal or semantic information. Filters serve navigation and, in some cases, can become genuine SEO pages with their own URL, title, H1, copy and dedicated products.
How does AI help the Merchant Center feed?
It can flag missing attributes, weak titles, inconsistent brands, wrong categories, thin descriptions and data that doesn't match between product page and feed.
Does AI replace the person managing the catalogue?
No. The best use is as an operational assistant: it finds problems, proposes solutions, automates checks and speeds up the work, but the important decisions stay governed.