AI shopping tools don’t recommend products. They recommend product data. Ask ChatGPT, Google’s AI Mode or Perplexity for “waterproof running shoes for overpronation under ₹8,000”, and the tool can only shortlist what it can find, match to each condition and describe with confidence. It usually names a handful of products, with no long list to scroll through. That makes it a competition, a bit like the first page of Google, with much less room.
Quick answer: AI shopping tools pick products whose data they can find, match to the shopper’s conditions and describe without contradiction. Products drop out at any of those three stages. No platform publishes its full ranking method, so this is our way of organising what is documented.
What each platform documents about its data
Different tools read different sources. This is what each platform says on its own pages.
| Tool | What the platform says | Source |
|---|---|---|
| Google (Gemini, AI Mode) | Shopping in the Gemini app is “powered by the Shopping Graph”, a dataset of “more than 50 billion of the world’s products”. Inputs and ranking for AI Mode: not documented. | Google blog, 7 Apr 2026 |
| ChatGPT (OpenAI) | Merchants provide “a secure, regularly refreshed feed”. The standard upload “currently targets the US”. “Eligibility does not guarantee display.” | OpenAI commerce docs (feed spec, key concepts) |
| Perplexity | Instant Buy is “currently available to all users in the United States”. Listings are ranked on “relevance and ratings”. | Perplexity Help Center |
| Microsoft Copilot | “We leverage both information found on the web and from a merchant’s feed in the Microsoft Merchant Center.” Copilot Checkout is open to “English-language merchants who sell to US buyers” for now. | Microsoft Advertising |
| Amazon (Alexa for Shopping, formerly Rufus) | Renamed on 13 May 2026. Draws on “Amazon’s extensive product catalog, customer reviews, community Q&As, and information from across the web”. | Amazon newsroom |
Last checked 13 October 2026. “Not documented” means not found on the platform’s own pages that we checked. Flipkart also has a shopping assistant, but we found no statement about its data sources, so it is not listed.
Each tool reads different sources, and each channel asks for its own format. What has to match is the facts: if one channel says a bag is 40 cm wide and another says 45 cm, a shopper asking “will it fit my laptop?” gets the wrong answer. Keep every fact correct and consistent, and present it the way each channel asks.
Stage 1: Getting in
A tool can’t recommend a product it can’t find. Three things decide that.
Be present in a source the tool reads. Google says merchants can share product data through structured data on product pages or through a Merchant Center feed. A feed is mandatory only for some surfaces, such as the Shopping tab.
Use real identifiers. A GTIN is the number under the barcode: EAN in India and Europe, UPC in the US. It helps systems see that your website, a marketplace and a reseller’s store sell the same product. Google strongly recommends one where it exists, and says a product with a GTIN submitted without one “may have limited visibility”. OpenAI’s feed spec accepts a valid GTIN too. Where a product truly has none, give the brand and manufacturer part number instead, and never invent a value to fill the gap.
Keep price and availability in sync. Google says an availability mismatch between your page and your product data “will trigger a product disapproval”, and Merchant Center lists “Mismatched value (page crawl) [price]” as a disapproval too. Google sets no update schedule. It asks for “the most up-to-date data” and offers automatic item updates that use your landing page. Make the page, its structured data and the feed show the same figures, and name one owner for price and stock, which usually live in your ERP or commerce platform.
Stage 2: Getting matched
Being found isn’t the same as being chosen. Shoppers now ask long, specific questions, and the tool has to confirm each condition from your data.
Fill the attributes shoppers ask about. For “waterproof running shoes for overpronation”, the tool needs two facts: is the shoe waterproof, and is it a stability shoe? If your listing says “all-weather comfort”, the match is much harder to make.
Give it fields, not just a description. Clearly labelled fields for price, availability, identifiers, variants and attributes remove the guesswork. A description can say “lightweight”. A field says weight: 260 g. Marketplaces and feeds have their own attribute names and allowed values, so use those first, and put anything without a field into your bullets and description, stated the same way as everywhere else.
Get variants right. Each size and colour is its own variant, and each needs its own price, stock status, image and attributes. A shoe marked “in stock” is no use to a shopper if their size 10 in black has sold out. Inside your own product data (spreadsheets, supplier files and feeds), write every option the same way: S, M, L or Small, Medium, Large, never a mix. Google’s size guidance says “Use a consistent format for size” and warns against submitting “S”, “Medium” and “Lrg” together. A platform’s dropdown accepts only its own values, so map yours to them.
Use precise, consistent words. “Waterproof” and “water-resistant” are different claims, “wide fit” and “2E” are related but not interchangeable, and UK 8 isn’t US 8. A controlled list of allowed values, a dropdown instead of free text, stops teams drifting apart.
Stage 3: Being described correctly
Once a product is shown, the tool describes it, and that depends on whether your sources agree.
Conflicting facts create doubt. If one listing says “wide fit” and another says “regular fit”, the tool has conflicting evidence. With only a handful of slots to fill, a product whose facts disagree is an easy one to skip. Regulators expect the facts to match too. In India, FSSAI’s December 2024 advisory to e-commerce platforms says product claims online “must be fully aligned with the information provided on the product’s physical label.”
Markets need different versions. The same product can need a different name, label or certificate in each market. Indian marketplaces must show each food seller’s valid FSSAI licence or registration number. In the UK and EU, the 14 main allergens must be emphasised in the ingredients list, under EU Regulation 1169/2011 and UK Food Standards Agency guidance. A shoe sold as “sports shoes” in India is “trainers” in the UK and “sneakers” in the US.
Images may count as data too. AI tools may read images as well as text, so a photo that contradicts the specification is a risk.
What nobody can promise
No platform publishes its full ranking method, and none promises placement. OpenAI’s feed spec says “Eligibility does not guarantee display.” But platforms do say richer data helps. OpenAI says recommended attributes “improve ranking, relevance, and user trust”, and Perplexity says its answer engine is “more likely to recommend merchants who provide deeper product details”.
The three stages are our way of organising it, not a documented pipeline, and we have not measured how often fixing product data changes what an AI tool shows. Treat any promise of “guaranteed AI visibility” with caution.
Start here: a ten-product check
Pick your ten best-sellers and work in this order.
- Run the platforms’ own checks. Open the product issues report in Google Merchant Center and each marketplace’s listing-quality or suppressed-listings report. Fix flagged products first.
- Check identifiers. For each product, note the GTIN or EAN and confirm it is a genuine number issued through GS1. If there is none, make sure the brand and manufacturer part number are filled in.
- Compare price and availability. Open the page, the feed and each marketplace listing on the same day. Do the price and the stock status match?
- Fill the mandatory fields. Check each marketplace template’s required fields. In India these include tax details such as the HSN code, which Flipkart’s listing API makes mandatory, and for food, the seller’s FSSAI licence or registration number.
- Answer the shopper’s questions as fields. Write down the three questions shoppers ask about this category. Is each answer in a field, not only in the description?
- Match variant options. List the option values across a variant group. Is the format identical?
- Compare facts across channels. Put fit, material and weight from your website, marketplace and distributor sheet side by side. Does each channel state the same facts, even if the format differs?
- Check images. Does each variant’s image show that variant, and does it meet each channel’s image rules?
Fix in this order. Mismatched price or availability can get a product disapproved on Google, and a missing GTIN can limit its visibility.
Why shared spreadsheets break
Spreadsheets are not the problem on their own. A shared sheet with dropdown lists and one named owner can work well for a small catalogue. It starts to break as three things grow: the number of products, the number of channels and the number of teams who add to it.
At GridPXM, we see the same typical case: the right answer exists, but in another team’s file. Production knows the material, the warehouse knows the shipping weight, compliance holds the certificates, and the person writing the marketplace listing sees none of it. So each channel gets its own copy, the copies drift apart, and nothing enforces a standard. One team writes “water-resistant”, another writes “waterproof”, and the sheet accepts both. When a platform adds a field, someone searches twenty files for the answer.
That is a process problem first, and it is hard to fix without one record per product, agreed rules and named owners.
Where GridPXM fits
GridPXM is PIM software that helps you get the facts right at the source, before anyone copies them.
- Required attributes by category. You define which attributes each product category needs.
- A check before it’s saved. Every product is checked for missing or invalid values, and shows what it still needs before it’s ready.
- One record, several teams. Production, listing and warehousing add to the same product record.