What to Check First Before Advertising in AI Shopping

Ads are showing up on the screens where AI recommends products, too. As an assistant compares items to fit a user’s conditions, it now surfaces sellers marked as sponsored, or highlights certain discounts alongside them. There’s also a new feature that lets sellers build ads directly from their catalog data — product names, prices, and inventory counts.

When I look at this shift, I think product data deserves more attention than ad format does. Even if an ad catches someone’s interest, a mismatched price or an out-of-stock item makes it hard to close the sale. And if the compatibility specs or shipping conditions an AI needs for comparison are missing, that creates problems of its own.

That said, if you lump the recommendation feature, paid ads, and checkout integration all together as one kind of “ad product,” the prep work gets tangled. Splitting it into the process where a product gets discovered, the process where it’s shown alongside ads during comparison, and the process of actually purchasing it makes it much clearer what you need to check at each stage.

Google Tests Ads Below Recommendations, Perks Right Before Purchase

In January 2026, Google announced a Direct Offers pilot in AI Mode that shows discounts and similar deals to users with high purchase intent. Advertisers set up the offers, and Google surfaces them based on the conversation context.

In February, Google said it’s testing retailer ads tagged “Sponsored” that appear below general product recommendations. It also outlined plans to expand Direct Offers beyond discounts to membership perks and bundled products. What matters in this announcement is the distinction between where general recommendations appear and where ads appear. Paying for a placement doesn’t mean an advertiser can rewrite the description of a general recommendation however it likes. Google’s advertising and commerce plans.

Not everything released around the same time is an ad. Business Agent is a feature that answers questions about a brand’s products within Search. UCP is a shared protocol that links multiple services together across product discovery, purchase, and post-purchase handling. A brand answering a question doesn’t make that single answer a paid ad product, and connecting a payment protocol doesn’t mean an ad was bought either. Google’s UCP, Business Agent, and Direct Offers announcement.

This distinction matters for budgeting too. Development costs for product-data integration, advertising spend, and payment processing fees can each have different purposes and different counterparties in the contract.

OpenAI’s product feed serves two separate masters: recommendations and ads

OpenAI announced a self-service Ads Manager beta in May. It’s a tool that lets advertisers manage campaigns and budgets. The company also laid out a principle: ads are kept distinct from answers, and advertisers don’t get access to private conversation content. Ads Manager announcement.

Product feed ads link a seller’s catalog to a campaign. The feed carries product names, descriptions, prices, stock levels, images, and sales-page URLs. When an ad runs, the system picks a target product and builds the ad from that product’s data. This cuts the burden of manually crafting ad creative for every single product. But it also means that if the feed holds a stale price, the ad will surface that wrong information too. OpenAI’s guide to product feed ads.

Three separate settings need to be distinguished here: eligibility for search/recommendations, eligibility for ads, and eligibility for checkout integration. Under the current product feed spec, ad eligibility settings are independent of search eligibility settings. Checkout requires a separate integration and must also meet search eligibility conditions. Simply flipping a setting on doesn’t guarantee visibility or checkout.

So it’s inaccurate to assume a sequence where “a product must first qualify as a recommendation candidate before it can be advertised.” Even when the same product data is reused, the participation conditions and settings for each feature have to be checked separately. OpenAI’s product feed spec.

AI shopping ad

Splitting by discovery, comparison, and purchase shows you who owns what

The reason I break this into three stages isn’t to pin down categories of ad space — it’s to clarify what sellers actually need to prepare.

Stage	What sellers need to check	Example features
Product discovery	Whether product information is delivered, and whether the market and product qualify for support	Product feeds, search/recommendation integration
Product comparison	Whether the information provided fits the customer’s criteria, and how ads are displayed	Product comparison, feed ads, Sponsored merchant listings
Purchase	Whether the price and benefits actually apply to the order, and whether checkout and post-purchase handling are connected	Direct Offers, seller checkout pages, payment integration

Take a customer looking for a laptop charger, for instance. They want to know not just the price but whether it’ll actually work with their specific model. If the product name just says “fast charging” with no wattage or compatible specs listed, that question goes unanswered. Not every platform will automatically disqualify a listing for missing that information at a given query stage, but it’s hard to provide the explanation a buyer needs to make a decision.

This isn’t something you fix by rewriting ad copy. The product team has to manage accurate specs, the operations team has to keep prices and inventory current, and the ad team has to decide which products and offers to surface. When return or warranty terms change, that has to be reflected in the product information too.

I think it’s better to settle who owns this information before increasing ad spend. Even when testing a new ad format, the underlying information has to be correct — otherwise you can’t tell what’s actually causing weak performance.

The draft looks accurate and complete. No Hangul characters, all numbers preserved (February (translated correctly), seq=34353 in URL unchanged), heading and link counts match. No changes needed.

Korean sellers need to look at both target countries and distribution channels

Conversational shopping has already begun in Korea, too. In February, Naver launched a beta shopping AI agent in its Naver Plus Store app. The early version focused on summarizing product information, making comparisons, and analyzing reviews. For example, if you enter a condition like looking for a sofa that fits a newlywed home shared with a dog, it pulls up product specs alongside customer reviews. Naver’s launch announcement.

Listing products on a domestic platform and connecting a catalog to a global AI service aren’t the same undertaking. Sellers first need to check what product information each channel actually accepts, which countries’ buyers it targets, and whether their account is even eligible to participate.

OpenAI’s standard product feed upload is currently built for the US market as well. Simply adding a country field to the file doesn’t get your products exposed in other markets. Using a global service doesn’t automatically mean your products become discoverable in every country.

Brands that sell overseas need to check supported markets and the scope of shipping and returns together. Brands that sell only domestically would do better to start by examining the features actually offered by the distribution channels their real customers use. What matters isn’t sending your feed to as many places as possible — it’s connecting to the conditions under which customers can actually buy.

Order records and recommendation reasons are two different things

We also need to look at where the purchase process actually ends. In March, Shopify announced product discovery and purchase features through ChatGPT. On mobile, the merchant’s checkout page opens inside an in-app browser; on desktop, it moves to a separate tab. So even though it looks like the shopping continues inside ChatGPT, the entity operating the checkout page doesn’t actually change.

Shopify explains that merchants retain the transaction and the customer relationship, and can check in the admin dashboard which orders originated from ChatGPT referrals. So it’s not the case that AI-mediated orders leave no conversion record for merchants at all. Shopify’s agentic storefront announcement.

Still, knowing which channel an order came from is different from knowing why the AI recommended that particular product and what the customer compared it against. Order and referral metrics provided by the platform alone can’t capture the entire conversation.

This distinction matters when evaluating ad performance. Even if sales rise through an AI channel, we need to break down how much came from paid ads, how much from organic recommendations, and how much was simply existing customers switching purchase paths. Where possible, comparing against periods or products without ad spend, and checking repeat purchases and returns, gets us closer to the actual profit picture.

Before you set an ad budget, check five things

I’ve noticed that product data has become the shared foundation connecting recommendations, ads, and purchases. That doesn’t mean every function runs on a single rule, or that ad spend alone can buy you the outcome you want.

If you run a commerce business, I’d start by checking these five things:

Where is the source-of-truth for product, price, inventory, and compatibility information managed, and who updates it?
What data does the service use to describe products, and how are errors corrected?
How do readers distinguish organic recommendations from ads, and what does the ad contract actually guarantee?
Among traffic, order, and conversion records, which ones can the seller actually access?
After checkout, who handles returns, refunds, and customer inquiries?

You don’t need to allocate a big budget the moment a new feature launches. Start with your top 20 best-selling products and cross-check their actual prices, inventory, compatibility specs, and shipping/return terms. Then run small tests on the supported channels — that makes it much easier to tell whether the problem is with the data, the ad proposal, or the purchase flow.

This is exactly what I want to check first, too: not just where our products show up, but whether we’re giving customers accurate information all the way through to understanding the product and completing the order.

What to Check First Before Advertising in AI Shopping