The most common reason AI shopping assistants underperform isn't the AI. It's the catalog.
An AI shopping assistant takes a shopper's brief, "something for a Diwali party," "I need work clothes that aren't boring," and reasons over your product catalog to build a recommendation. How good that recommendation is depends almost entirely on what information exists in your catalog for the AI to reason over.
Most Shopify catalogs are built for navigation. They're optimized for a shopper who browses by category, filters by color or price, and lands on a product page. The metadata that supports this experience, product title, category, price, size, color tag, is sufficient for browse-and-filter. It's not sufficient for AI-assisted outfit building.
This guide is the practical version of the fix: what to add to your catalog, where to add it, and why each piece of data changes the quality of what an AI shopping assistant can do.
Why catalog quality is the rate-limiting factor
An AI shopping assistant that recommends a cocktail dress for a beach trip hasn't failed because the AI is bad. It's failed because the catalog doesn't contain clear signals about occasion appropriateness, and the AI is making its best guess from incomplete information.
The quality ceiling of any AI recommendation system is set by the quality of the data it's reasoning over. A system with a sophisticated model and poor catalog data produces generic, sometimes wrong recommendations. A system with a simpler model and rich catalog data produces specific, accurate ones.
Most AI shopping assistant implementations spend significant effort on the model and almost no effort on the catalog. The return is backwards. A 20% improvement in catalog richness typically produces a larger improvement in recommendation quality than a 20% improvement in model sophistication.
You control your catalog. You can improve it in a day. The model is someone else's problem.
The five catalog gaps that break AI recommendations
Before getting to the fixes, it's worth naming the specific failure modes. These are the gaps that produce bad recommendations even when everything else is working correctly.
Gap 1: Missing occasion metadata. Your catalog has categories (dresses, tops, bottoms) but no clear signals about which occasions each product is appropriate for. The AI is trying to answer "is this right for a Diwali party?" from a product titled "Embroidered Georgette Kurta" with tags: kurta, ethnic, orange. It probably guesses correctly. It sometimes doesn't.
Gap 2: Absent style attributes. Silhouette, drape character, formality register, print scale, fit profile, the attributes that determine whether two pieces work together aesthetically, are missing. The AI builds "outfits" by combining products that don't clash on color but have no stylistic coherence.
Gap 3: No complete-look data. Your merchandising team has editorial lookbooks. Those looks aren't in your catalog metadata. The AI can't access the human judgment that went into which pieces work together, it's reasoning from scratch on every recommendation.
Gap 4: Weak fit descriptions. "True to size" is almost useless. Does this run small in the shoulders? Is the waist structured or relaxed? How does the fabric move? The information that would resolve a shopper's fit uncertainty before it becomes a return isn't in the catalog, so the AI can't use it.
Gap 5: Generic product descriptions. "Versatile and chic, this dress is perfect for any occasion" isn't information, it's marketing. The AI needs specific descriptors: what occasions, what aesthetic register, what kind of shopper, what body type it flatters, what it pairs with.
The metadata framework
Here's the data structure that gives an AI shopping assistant enough to work with. This covers what to add, what format works, and what level of specificity produces meaningful recommendations.
1. Occasion tags (required)
The single highest-value addition. Every product should carry tags that indicate which occasions it's appropriate for.
The level of specificity matters. "Occasion: casual" is too broad. "Occasion: beach resort" or "Occasion: office casual" is useful.
Build your occasion taxonomy first. Rather than tagging ad hoc, start by listing the 15 to 25 occasions your catalog actually serves. For an Indian fashion brand, this might include:
Office formal
Office casual
Casual weekend
Date night
Wedding guest (Indian)
Wedding guest (western)
Festival — Diwali
Festival — Eid
Festival — Navratri / Garba
Cocktail party
Beach / Resort
Travel casual
Formal evening
Athleisure / Active
For each product, assign the 1 to 3 occasions it genuinely serves. Not every occasion that a creative marketing mind could stretch it to, the occasions where a shopper would actually reach for this piece.
Implementation in Shopify: Add occasion tags using Shopify's standard tag field (for existing products, this can be bulk-updated via CSV). Format them consistently, "occasion: diwali" and "occasion: Diwali" are different tags. Pick a format and stick to it.
2. Aesthetic and style attributes
These are the attributes that determine whether pieces work together and whether a piece fits a shopper's taste profile.
Silhouette: fitted, relaxed, structured, flowy, oversized, tailored. One term per product.
Drape register: stiff, soft, fluid, technical. This affects how the garment moves and what it looks like in motion, important for AI try-on accuracy and for outfit coherence.
Formality register: a scale from 1 (most casual) to 5 (most formal) is more useful than binary casual/formal tags. A 3 might be "smart casual dinner", useful for distinguishing between pieces that are similar in category but different in social register.
Print scale (for printed garments): small print, medium print, large print, pattern, solid. Small prints and solid pieces mix. Large prints generally don't.
Aesthetic family: the broader aesthetic that the piece belongs to: minimalist, maximalist, boho, contemporary, traditional, streetwear, classic, romantic. This allows the AI to understand stylistic coherence at a higher level than specific attributes.
Implementation: Add these as metafields in Shopify. If you're using a Shopify plan that supports metafields natively, add them under Product > Metafields. If you're pre-metafields, custom attributes work fine.
3. Complete-look data
If your team has produced editorial looks, lookbook shoots, Instagram styling posts, campaign imagery, that data should be in your catalog.
For each look, create a tag or metafield that groups the products that appear together. "Look: summer-campaign-03" on three products tells the AI that these pieces work together, with human editorial judgment behind that assertion.
This is cheap to add (it's a tagging exercise) and high-value to the AI (it bypasses the combinatorial reasoning the system would otherwise have to do from attributes alone).
Secondary benefit: Complete-look data improves your SEO and on-site cross-selling too, it's not only for the AI.
4. Fit notes (specific, not marketing)
The standard Shopify size guide is not enough. What the AI, and the shopper, needs is fit character data at the product level.
For each product, write one to three sentences that answer:
Does it run true to size, small, or large? And in what dimension (length, shoulders, waist, chest)?
What body types does it flatter? What is the waist treatment? Is there stretch?
What's the layering compatibility? Does it work under a blazer? Over a t-shirt?
Example of what not to write: "True to size. Model is 5'7" and wears size M."
Example of what to write: "Runs slightly long in the torso — petite shoppers may prefer sizing down. The waist is relaxed, not fitted — if you want definition, pair with a belt. The fabric has no stretch; size up if you're between sizes."
The second version gives the AI something it can actually use to answer "will this fit me?" questions in a way that's specific to this product.
5. Product description rewrite
Most fashion product descriptions are written for SEO keyword density or brand voice expression. They're not written to give an AI the information it needs to make recommendations.
The description format that works for AI-assisted recommendations:
Lead with the specific occasion or use case. "This kurta set is best suited for festive occasions — wedding celebrations, Diwali, Eid events where you want to look put-together without being overdressed."
Describe the aesthetic decision. "The embroidery pattern is traditional in motif but contemporary in scale — minimal enough to read modern, detailed enough to read special."
Name what it pairs with. "Pairs well with: block heel sandals (not stilettos — the silhouette is relaxed, not structured), minimal gold jewelry, dupatta optional depending on formality."
Acknowledge the limits. "Not ideal for: casual weekend wear, outdoor events where you'll be walking a lot, or occasions requiring western formal dress."
This structure gives the AI specific, directional information, not just a description of the product, but a description of what it's for, who it's for, and what works with it.
Implementation priority
You won't be able to enrich an entire catalog in a day. Here's the sequence that produces the most improvement fastest.
Priority 1: Occasion tags on your top 100 SKUs. Your top SKUs by traffic or by revenue are the ones most likely to appear in AI recommendations. Get those tagged first. This can be done in a few hours with a CSV upload.
Priority 2: Rewrite descriptions on your worst-returning SKUs. High-return SKUs usually have fit or occasion mismatch problems that better descriptions would address. Fix these first, the return rate impact is fastest here.
Priority 3: Complete-look data from existing editorial content. Go through your last three to six months of Instagram content and lookbooks. For every look you can identify, tag the products that appear together. This is mostly a data-entry exercise and can be done by a junior team member.
Priority 4: Style attributes on new products. Apply the attribute framework to all new products as part of your standard product listing workflow. Don't try to retroactively tag the whole catalog at once, do new products correctly and work backwards on existing catalog incrementally.
Priority 5: Fit notes on high-traffic PDPs. Product pages that get significant traffic but have high exit rates are likely losing shoppers to fit uncertainty. Adding specific fit notes to these pages directly reduces both abandonment and post-purchase returns.
What good looks like
Here's the difference between a catalog entry that gives an AI shopping assistant almost nothing to work with and one that gives it everything it needs.
Before: A product titled "Floral Midi Dress," categorized simply as Dresses, tagged only floral, midi, blue, women, with a description reading "A versatile midi dress featuring a beautiful floral print. Perfect for any occasion. Available in blue," and a size guide that just refers shoppers to the general size chart.
After: The same product retitled "Blue Floral Midi Dress — Relaxed Silhouette," categorized under Dresses > Midi, with occasion tags for date night, weekend casual, beach resort, and wedding guest (daytime). Style attributes specify a relaxed silhouette, fluid drape, formality level 2, medium floral print, and romantic aesthetic. It's grouped into a complete-look tag, summer-campaign-04. The description reads: "A relaxed-fit midi dress suited for casual social occasions — beach gatherings, daytime garden events, warm-weather dates. The floral print is medium-scale and blue-toned: it mixes well with solid whites and creams, less well with other prints or bold colors. The fabric drapes softly and moves well, which makes it comfortable in heat but means it doesn't layer under structured pieces easily. Best styled with flat sandals or low block heels; high heels shift the vibe toward more formal than the silhouette supports." And the fit notes add: "Runs true to size on the waist; the shoulder seam sits slightly wide — narrow-shouldered shoppers may want to size down. No stretch. Length hits at the mid-calf on a 5'6" model; petite shoppers will find it slightly longer."
The second version isn't just better for AI recommendations. It's better for every dimension of product performance, SEO (more specific terms), conversion (more useful information), and returns (better fit guidance).
The one-time investment with compounding returns
Catalog enrichment is front-loaded work. Tagging 500 products, rewriting 100 descriptions, and creating complete-look groups from your editorial library takes time.
The return on that time compounds. Every AI recommendation that uses the enriched data is more accurate than it would have been on the bare catalog. Every more accurate recommendation converts at a higher rate and produces fewer returns. The catalog work doesn't expire, it keeps improving every recommendation that references it.
The alternative, deploying an AI shopping assistant on an unenriched catalog, is possible, and vendors will tell you it works fine. It works fine the way a search engine works fine with thin content. The floor is functional. The ceiling is low.
Your catalog is the data layer. The AI is the reasoning layer. A good model on bad data produces mediocre recommendations. A good model on rich data produces the kind of recommendations that produce 3 to 4x conversion lift.
Start with the catalog.
When Elara indexes your Shopify catalog, it ingests and enriches your product metadata to make it AI-ready. But the richer your starting point, the better the recommendations from day one. Here's how to start your pilot with a catalog that gives Elara everything it needs.