The Watson Weekly is sponsored by Avalara — the agentic AI platform automating global tax and compliance for leading eCommerce brands with simple plans, transparent pricing, and complete coverage across Shopify, Stripe, WooCommerce, BigCommerce, and more.

The thing I keep coming back to is not the forecast. It is the gap between what consumers say happened and what actually happened.

Holden's number is that eleven percent of consumers believe they completed a purchase inside an AI app. He says flatly that they did not, and the GMV math backs him up. Most of these apps cannot process embedded checkout. But the belief is real, and it is being reported by people who are not lying. They are describing where they think the transaction started.

We have been here before. Holden pointed at social commerce, where sixty percent of North American consumers over forty thought they had bought something through a social platform in the previous four months. The math never supported it. What actually happened is they clicked a shoppable tag and landed on a website, so in their mind the purchase belonged to the platform while in the retailer's ledger it belonged to the .com.

The retailer's ledger is accurate. It is also not the thing determining where that consumer goes next time.

There is a kind of arrogance in telling a consumer they are mistaken about their own shopping journey because your attribution model says they checked out on your site. We solved a version of this in physical retail. When Nordstrom was building the case for Nordstrom Local, one of the metrics was e-commerce lift inside a DMA, and plenty of people thought that was insane at the time because stores were supposed to be judged on same store sales. The industry accepted a more ambiguous measure because the clean one was measuring the wrong thing.

Online we never did that. We built channel silos and defended them, and AI is now walking straight through the middle of it.

Holden also gave me the number that should make executives uncomfortable. Merkle surveyed more than a hundred companies above a billion in revenue. Eighty-eight percent have deployed something powered by a large language model. Six percent could prove it moved EBITDA.

He quoted John Wooden on that - never mistake activity for achievement. What I take from it is a little different. If eighty-eight percent have shipped and six percent can show it worked, most of what you are hearing from conference stages is posturing and you are probably not as far behind as you fear. That should function as relief and warning at the same time.

His advice for what to do about it barely mentions AI. Fix your product data. Get it out of the nineteen systems it currently lives across. Build the taxonomy and the APIs and finally implement the PXM you have been arguing about for three years. Your .com remains a source LLMs privilege, so that work pays off whether or not the 2030 forecast lands.

I asked him who has the time to train a personal AI assistant with enough context to shop for them, and he did not really disagree with the objection. His answer was that the context nobody has is order history, and the companies sitting on it are the ones least motivated to hand it over. Which is the actual question. Not whether agents will shop for us. Whether Amazon ever decides it is worth letting them.