On March 24, 2026, OpenAI switched off Instant Checkout, the feature that let people buy things without leaving ChatGPT. The company said the first version didn’t offer the flexibility it wanted, so merchants would go back to running their own checkout while OpenAI concentrated on product discovery.  

Six months earlier, the same feature had launched with the biggest players, such as Walmart, Etsy, and Shopify on board, and a million more merchants described as imminent. So, what went wrong? 

The usual suspect 

It would be too convenient to blame product data for all of that. The honest version is more mixed. Etsy told Modern Retail it never saw large sales volumes and that ChatGPT remains an early-stage channel for most shoppers, which is a demand problem rather than a data one.  

Analysts told CNBC that OpenAI struggled to onboard merchants and to support multi-item carts or loyalty programs, while also showing accurate product data on top of that. Forrester’s Emily Pfeiffer described the mechanism in an interview with CNBC: OpenAI was scraping some retailers’ sites for product information rather than getting them first-hand, which meant crucial data could be inaccurate or out of date. 

That detail is the whole article in a nutshell. When an agent can’t get structured and current product information from you, it doesn’t give up. It assembles something from whatever it can reach and presents the result to your customer with complete confidence. In other words, the failure wasn’t that the AI was bad at reading. There was just nothing authoritative to read. 

Look at what survived the retreat. The protocol underneath Instant Checkout is still running, but its job now is product feeds, availability, promotions, and attributes. The interface was pulled with only about a month of notice, but the data layer stayed and became more important than before.  

If you are planning functionalities such as AI-powered search, recommendations, assistants, or agentic selling for next year, treat that as your preview: the surfaces will keep changing, and every one of them consumes the same thing. 

The channel flipped, but the catalogs didn’t. 

AI referral traffic has stopped being just a curiosity. Adobe Digital Insights, working from more than a trillion visits to U.S. retail sites, reported that traffic from AI sources grew 393% year over year in the first quarter of 2026, on top of a holiday season that ran 693% ahead of the year before. In its companion survey of more than 5,000 consumers, 39% said they had already used AI to shop online, and 85% of those said it improved the experience. 

The more interesting number is conversion. In March 2026, AI traffic converted 42% better than non-AI traffic. In March 2025, it converted 38% worse. Within twelve months the channel went from the weakest performer to the strongest, and shoppers arriving from an assistant now spend 48% longer on site and view 13% more pages. Two caveats worth keeping: this is Adobe’s own analytics data rather than an audited market study, and the comparison is against all non-AI traffic combined, not against Google organic alone. 

Then comes the part that should make you uncomfortable. Adobe also scores how much of a page a language model can actually read, where a score of 50% means half the content is invisible to machines. Across U.S. retail, homepages averaged 75% and category pages 74%. Product pages came in at 66%, the lowest of every page type Adobe measured. 

Chart showing 74% AI-readable content on retail pages and 66% on product pages—demonstrating the growing importance of AI-ready product data for seamless checkout experiences, with the source cited as Adobe Digital Insights, 2026.

Read that list again in commercial terms. Your returns policy scores, on average, 82%. Your contact page scores 81%. Your FAQ scores 80%. The pages where you explain how to send things back are more legible to an AI agent than the pages where you explain what you sell. 

This gets worse in B2B 

The numbers above come from U.S. retail, which is where the measurement infrastructure exists. It would be easy for a manufacturer or a technical distributor to read them as somebody else’s problem. In our project experience, the opposite holds for three reasons: 

  1. B2B catalogs are larger and far more attribute-dependent. A buyer choosing a hydraulic fitting is not making an aesthetic judgment, but matching a specification, which means the purchase decision lives almost entirely in structured attributes rather than imagery or copy.
     
  2. Those catalogs also carry more history: product data that has passed through an ERP, one or two acquisitions, a supplier import routine written a decade ago, and a marketing team that never had authority over any of it.
     
  3. B2B sells through more surfaces at once, with the same product appearing in a distributor portal, a marketplace, a punch-out catalog, and a printed spec sheet, each with its own version of the truth. 

The retail data is a leading indicator rather than a different phenomenon. The mechanism is identical, and the exposure is higher. 

Commercial implications

Most organizations still treat product information as a back-office obligation. Someone in operations owns completeness, someone in marketing owns the descriptions. The ERP owns, well, whatever the ERP owns. It works, more or less, because a human shopper compensates for the gaps. They check another page or simply make an assumption. That creates friction, but the sale can still happen. When there’s an AI agent on the other side, there may be no second chance. 

An agent doesn’t compensate. It filters. That difference is not one of degree, and it changes what a blank field costs you. 

Data problem  Traditional search  Agent-driven discovery 
Missing attribute  Lower ranking, still findable  Excluded from the result set 
Inconsistent values across systems  Human reconciles them  Product dropped or misdescribed 
Stale price or stock  Corrected at checkout  Wrong answer given, no click ever happens 
Where you notice  Analytics and rankings  Nowhere 

 The last row is the one that should worry you. A ranking drop shows up in a report, but exclusion from a candidate set produces no data point at all. You can’t see the revenue you aren’t getting, which is exactly why this problem stays unfunded. 

In conversations with eCommerce leaders, we see the same pattern again and again. The problem is rarely a lack of product data. It’s that nobody has decided: 

  • Which information actually matters for a buying decision?  
  • Who owns it?  
  • Where does it need to be consistent?  

Those answers have typically been established years ago, and with the context of human readers. Once AI starts making those buying decisions on behalf of the customer, everything starts to fall apart. 

Why “fix the data” is the wrong project 

Nobody needs convincing that their product data has gaps. However, “our catalog is incomplete” is a statement about hygiene, and hygiene loses every budget round to initiatives that arrive with a number attached. Diagnosis is the easy part. Where it gets difficult is showing which gaps cost revenue, and roughly how much. 

The standard answer at this point is roughly this: launch a data quality program, establish governance, unify the source of truth, clean the catalog. It’s not wrong, but in our experience, it’s also why so little changes. A full remediation across a large assortment is an eighteen-month program with no visible outcome until late, and nothing inside it can be traced back to a single euro of revenue while it runs. We’ve watched several of these start with real commitment and quietly lose their sponsor at the first reprioritization, while agents kept reading the current feed every day. 

There’s another problem with this approach. Generative AI makes it trivial to fill blank fields at scale, which feels like progress, but frequently isn’t. Generating attributes without a validation layer produces variance faster than any human team could: category logic splits, naming diverges between regions and channels, and, as a result, the catalog gets larger and less trustworthy at the same time. You can’t generate your way out of a governance problem, and you can’t govern a catalog that you haven’t scoped. 

The question we would ask a client is a commercial one: which attributes and buying questions decide whether a sale happens at all? Most of your catalog doesn’t need to be perfect. You need to know which part of it is costing you money right now, because that is the part that gets funded. 

How to narrow it down 

This is the sequence we use with clients. It runs in weeks rather than quarters, and each step produces something you can put in front of your CFO. 

1. Start by asking the machines what they already say about you.  

Take your ten highest-margin products and put the questions your buyers actually ask to the major assistants: 

  • What is this made of?  
  • What does it fit?  
  • What size do I need?  
  • Is it in stock? 
  • What’s the difference between this and the model above it? 

Record the answers verbatim. Expect to find products described with a competitor’s specification, and prices that haven’t been correct since last quarter, or even products that don’t show up at all. This costs a couple of days of one person’s time and turns an abstract data quality argument into a list of specific commercial failures the commercial side of the business will act on. 

2. Scope by margin rather than SKU count.  

Prioritize the products carrying a disproportionate share of contribution margin rather than the ones with the largest number of variants. In most assortments that set is small enough for remediation to be a matter of weeks. Starting with the long tail is how these programs quietly die. 

3. Map buyer questions to attributes, then check whether the attributes exist at all.  

For that product set, write down the ten questions a buyer asks before purchasing and trace each one to the field that answers it. The exercise usually reveals that the highest-value attributes are precisely the ones nobody was ever asked to maintain, because no human shopper needed them written down. 

4. Build the control layer last.  

Governance is what stops a remediated catalog from decaying back, and it’s worth close to nothing before you decide what you’re governing. Sequenced this way, it’s a maintenance decision with a defined scope rather than a transformation program that needs its own budget line. 

“But we already have a PIM” 

This is one of the most common responses I hear from clients. Having a PIM is not the same as having product data that’s ready for AI. Your PIM stores what the business previously decided was important. The question now is whether that information is enough for a machine to understand your product and compare it with others. 

This is why the third step above matters more than the platform choice. The output of that exercise is a decision about which attributes are mandatory for which products in which channel. A PIM is where that decision gets enforced rather than merely documented. It’s also the only sensible place to put a validation layer over AI-generated content. It will ensure that machine-written attributes and descriptions become structured and checked product information instead of variation that spreads faster than anyone can review it.  

Across the client conversations I’ve been part of, the technology is rarely the hardest part. The harder question is deciding which product information the business is actually willing to own and be accountable for. 

AI readiness beats AI adoption 

You don’t need to fix your entire catalog to become AI-ready. You need to know which products matter most and which questions your customers need answered. Then, you need to verify whether your product data can answer them consistently. 

That is where I think the conversation needs to move: from adopting more AI to making sure the business is actually ready to benefit from it.