For Brands · Data
Product Feed Optimization for AI Search: From 5 Fields to 40
Your product feed was built to fill a price-comparison grid. AI assistants are asking it questions it was never designed to answer. This is the gap that decides who gets recommended, and it is very fixable.
The five-field problem
Open your product feed and count the real fields per SKU. For most brands it comes to five and a few spares: title, price, image, link, category, maybe brand and availability. That was enough for a decade of Google Shopping, because the job was matching keywords and sorting by price.
Now the buyer on the other side of the feed is an AI assistant handling a question like "a gentle vitamin C serum for sensitive skin that works under makeup, around 1,500." To answer it, the assistant needs concentration, formulation, skin-type suitability, texture, layering behaviour, price, and stock. If your feed says "Vitamin C Serum, 999, in stock," you are not in that answer. Not because the product is wrong, but because the data never showed up to the comparison.
AI assistants do not infer and do not fill gaps. They read what the data says and skip what it does not. That single fact, multiplied across the fan-out of sub-queries we described in the agentic commerce guide, is why feed depth has become the main ranking factor of AI shopping.
The four layers of an AI-ready SKU
Layer 1: identity, so AI knows what it is
Brand, model, GTIN, variant relationships, retailer ID matching, category taxonomy. Boring, foundational, and commonly broken. If your blue and black variants look like two unrelated products, or your GTIN does not match the marketplace listing, the AI cannot connect your product to the reviews and signals that exist for it elsewhere.
Layer 2: attributes, so AI can compare it
This is the 15 to 40 fields, and the right set depends on category:
| Category | What the AI needs to compare |
|---|---|
| Apparel | Fabric and weight, fit, stretch, occasion, care, size chart behaviour, season |
| Beauty | Ingredients and concentrations, skin types, concerns addressed, texture, certifications, allergens |
| Electronics | Battery life, compatibility, dimensions and weight, standards, ports, what is in the box |
| Food and supplements | Ingredients, nutrition, dietary flags, certifications, serving guidance |
Layer 3: context, so AI can reason about it
The layer almost everyone misses. "100% linen" is an attribute. "Linen breathes in humid weather, suits summer travel, wrinkles as part of the look" is context. Context is what lets an assistant connect your attribute to a shopper's situation: the persona, the occasion, the climate, the pain point. Without it your product can be sorted but not chosen.
Layer 4: trust, so AI can recommend it
Ratings, reviews, FAQs, user content, press mentions, repeat-purchase signals. Assistants stake their credibility on recommendations and visibly prefer products with proof. Much of yours already exists, scattered across marketplaces and communities. Aggregating it and attaching it to the SKU is some of the highest-leverage work in this list.
The part nobody warns you about: feeds rot
Even a perfect feed decays, because the target keeps moving. AI platforms ship new fields, flags, and schema requirements continuously. Google keeps adding conversational attributes like Q&A content, document links, related products, popularity rank, and variant options. OpenAI revises its Agentic Commerce Protocol. UCP and Merchant Centre specs evolve. Miss a newly required field and your product quietly drops out of answers, with no error message and no alert.
So feed optimization is not a project. It is an operating loop: watch every platform's spec, map changes to your catalog, fill the gaps, resync, and verify the result in your share-of-voice tracking.
Enrich once. Stay current everywhere.
Ziffi extracts your existing catalog, enriches it to AI depth, aggregates your reviews, and syncs it to every AI surface, then pushes every spec change automatically. Free integration. Ziffi earns only when it drives revenue.
A practical path from 5 to 40
- Audit your top 20 SKUs. Count real, structured fields per product. Note which of the four layers each SKU is missing. Most brands find layers three and four are simply absent.
- Mine what you already have. Product pages, images, size charts, customer service logs, and marketplace listings contain most of the missing attributes in unstructured form. The work is extraction and structuring, not invention. Never fabricate: assistants cross-check against reviews and marketplace data, and inconsistency reads as untrustworthy.
- Write context per attribute. For each major attribute, one line on who it matters to and when. This doubles as great product-page copy.
- Attach trust. Aggregate ratings and reviews from everywhere your products sell, plus FAQs answering the questions shoppers actually ask.
- Sync everywhere and verify. Push to every surface that matters, then check whether you actually surface for your target queries. If a competitor beats you, diff their product card against yours and fix exactly what is missing, as covered in why ChatGPT doesn't recommend your products.
Done by hand this is weeks per category, which is why it usually does not happen. Done with infrastructure it is one integration, with most merchants live within hours. Either way, the order of operations stands: identity, attributes, context, trust. That ladder is what turns a catalog into something an AI can sell.