AI shopping picks a different seller. Agents make free returns expensive.

In the fifteenth episode of Assistant to the CMO, Lexie Meskouris speaks with Will Nash and Harjot Singh about what happens when an AI shopping shelf quietly picks a different seller for a product you make.

Productrise ran the same shopping queries through Google's AI Mode and through ordinary search, in both the US and the UK. The first price shown in AI Mode averaged 21.6% higher, and where prices differed a different seller sat behind the listing about half the time. Will thinks a snapshot of a surface Google keeps changing is worth very little. Harjot stakes the opposite, that the shelf favours whoever looks like the place to buy, and that the premium attached to that is permanent.

The second half turns to Meta's new shopping agent and what Will thinks it breaks first. Generous returns policies are often priced on customers not bothering to claim, and an agent removes that assumption for free. The advice is small: check who the shelf says is selling your product, and read your refund terms the way a bot would.

Highlights of this Episode Include:

• The Shelf Chooses the Seller: Where prices differed, a different seller sat behind the AI Mode listing 49.6% of the time.
• The Headline Measured a Model That Moved: Will recalls AI Mode changing soon after publication, leaving the price number close to worthless.
• Canonical Beats Cheapest: Harjot stakes that Google optimises for the best place to buy, and that a premium for the canonical seller is permanent.
• Google's Defence Is a Non-Answer: The Shopping Graph is a data store, so it says what is available and nothing about what gets chosen.
• The Brand May Not Be the Shop: Harjot concedes the canonical place to buy is often a retailer with more evidence behind it than the manufacturer.
• Free Returns Were Priced on Apathy: Meta's shopping agent is live, and Will argues offers that assume nobody chases a refund are exposed.
• Two Things to Do This Week: Check the lead seller and price for your three best sellers in AI Mode, and read your refund terms as a bot would.

Listen and subscribe: https://AssistantToTheCMO.show

Learn more about Aloudable: https://aloudable.com/

Book a demo: https://aloudable.com/get-a-demo

New episodes of Assistant to the CMO land every Thursday.

ChatGPT ads promise not to touch the answer. Google expands AI Overviews and gives no numbers.

In the fourteenth episode of Assistant to the CMO, Lexie Meskouris speaks with Will Nash and Harjot Singh about whether anybody outside OpenAI can ever check the promise that ads do not influence the answers ChatGPT gives.

OpenAI says ChatGPT Ads reached a one billion dollar annualised run rate under 200 days after its first ad, with self-serve buying now open across Europe, India, the Middle East and North Africa. Both call that signal. The question underneath is whether OpenAI's promise that ads do not influence answers can be checked from outside. Harjot Singh calls it policy rather than a technical guarantee, and Will Nash argues a regulator can only police what leaves a trace.

The second half is Google auto-expanding AI Overviews, with no numbers given. Harjot's read is that the old length limit was economic and layout rather than retrieval. Will's worry runs further out, because removing how publishers get paid makes it harder for the models to stay current. The Monday move is to check whether your buyer sits on a ChatGPT tier that can see an ad.

Highlights of this Episode Include:

• Somebody Else's Customers: Ads run only for logged-in adults on Free and Go, so check which seat your buyer sits in before spending.
• It's Always Policy: Harjot on why OpenAI's promise that ads do not touch the answer has no mechanism under it.
• The Test That Can Never Close: Harjot's probe ad would only ever give suspicion, because targeting is the same signal as relevance.
• Plausible Deniability Is the Shape of It: Will on why a regulator can only police what leaves a trace, and a model's framing leaves none.
• Policing Something Nobody Can See: Harjot's earlier targets policed a label anyone could check, which is why this one needs a regulator.
• The Short Answer Was a Layout Decision: Harjot argues the old length limit came from economics and screen space rather than retrieval.
• Where the Answers Come From in Five Years: Will on how removing the way publishers get paid makes it harder for the models to stay current.

Listen and subscribe: https://AssistantToTheCMO.show

Learn more about Aloudable: https://aloudable.com/

Book a demo: https://aloudable.com/get-a-demo

New episodes of Assistant to the CMO land every Thursday.

Retailers can't out-build the shopper's AI. Amazon lost in court trying to block it.

In the thirteenth episode of Assistant to the CMO, Lexie Meskouris speaks with Will Nash and Harjot Singh about what a retailer can do when the shopper arrives holding their own AI, and whether either of the two obvious answers still works.

Narvar's US-only 2026 holiday report found 65 percent of shoppers plan to use AI for part of their Christmas shopping, and 8 percent of retail decision-makers feel very confident using it. Will Nash would track the 43 percent doing gift discovery instead, and adds a third source for a model's opinion of your brand: your live product catalogue. Harjot Singh argues a retailer cannot out-build the shopper's own agent, which knows the customer when the retailer's site starts from nothing.

The second half is the Ninth Circuit vacating Amazon's injunction against Perplexity's Comet, which turned on where a screenshot gets processed. Will has no view on the case. Harjot argues nobody designed that architecture for a court, GPU costs did, and that blocking agents is futile. The Monday move is smaller: put your returns and shipping terms on the product page as plain text.

Highlights of this Episode Include:

• The Number That Costs You Money: Will skips the 65 percent headline for the 43 percent of US shoppers using AI for gift discovery.
• It Knows You, Your Website Doesn't: Harjot's case that a retailer cannot out-build an assistant the shopper already carries.
• A Third Layer, and You Maintain It: Will adds the live product catalogue to memory and search as a source of what a model thinks of you.
• It Cannot See Your Photographs: Harjot on why shopping indexers mostly read text, because classifying every image costs too much.
• Amazon Lost on GPU Economics: The Ninth Circuit ruling turned on where Comet's screenshot is processed, which Harjot says hardware decided.
• Nobody Is Selling You the Future: Harjot on vendors pitching agent-ready commerce when the thing itself is still an experiment.
• Put the Returns Terms in Words: Will's move for when price is level, and Harjot's test is whether the page still reads with images off.

Listen and subscribe: https://AssistantToTheCMO.show

Learn more about Aloudable: https://aloudable.com/

Book a demo: https://aloudable.com/get-a-demo

New episodes of Assistant to the CMO land every Thursday.