This week I sat down with Ashwin Menon, the former Voice of the Seller at Amazon, to talk about something I think most sellers are still underestimating: what happens when your customer asks ChatGPT what to buy instead of typing it into an Amazon search bar.

We also get into why ChatGPT ads aren't working yet, the Amazon Ads release this week that connects your PPC data to AI agents, what one of our AI for Ecom cohort alumni built to stop burning money on AI tokens, and why the early bird for the Japan Mastermind closes Aug 31, 2026.

Here's the 80/20 of Ecom for this week.

When someone asks ChatGPT what to buy, will it recommend your product?

I recorded this interview with Ashwin Menon, co-founder of Nile and formerly the Voice of the Seller at Amazon FBA. His job at Amazon was being the seller's advocate internally. Now he's building the same kind of thing for AI agents.

He opened the call with news I hadn't seen yet. Ashwin told me ChatGPT had just brought back in-chat checkout, so a shopper can debate a product with GPT and buy it right there without ever clicking through to your site.

His line stuck with me: "Your product display page, your advertisement, and your lander all became ChatGPT's surface."

Here's the part I think sellers are getting wrong. Most brands ask "how do I get seen in AI?" and stop there. Ashwin says that's just top of funnel thinking, and it isn't enough if you're going to treat this like a real sales channel.

Because AI can hallucinate about you. It can link to a reseller instead of your site. It can pull some random Reddit post from an unhappy customer and read that back to your buyer as fact.

His advice is a mindset change before it's a tactic: stop thinking of AI as a casino and start treating it as a sales channel.

He also went out on the street in San Francisco and asked people whether they use AI to shop. About half said yes, and it's the first place they go. The other half said they buy from brands they know or on Amazon — but when he asked those people whether they check with ChatGPT first, they said yes, of course.

Then he said the thing that explains why this matters more than another ad platform. Shoppers don't file it as marketing. They think the AI is on their side.

I see it in my own family. My mom is 70 years old and she trusts ChatGPT. She's asking it whether she should write a book about our family. It's not a search engine to her. It's a friend.

Ashwin puts it this way: ChatGPT is your hip younger sister who knows what's cool, and Claude is your nerdy older brother who gives you one answer and no small talk.

TAKEAWAY: Go ask ChatGPT the 3 questions your best customer would ask before buying in your category. Right now, before you finish this email. If it recommends a competitor, sends them to a reseller, or repeats something wrong about you, that's your work list for next week. It costs nothing to find out.

What you can actually do about it on Monday

The second half of the conversation was the practical part. Three things Ashwin says are working right now.

YouTube. It's the number one cited domain in Google's AI Overviews. Ahrefs put it at about 21% of all AI Overview citations as of June 2026, and other studies put it higher. Every sentence spoken in a video is citable text. Ashwin's advice is to run two tracks: your own channel doing explainers, comparisons and tutorials, plus real reviews from creators.

The change is small but it matters. Script the video answer first. Lead with the answer to a buying question instead of building up to it, because that first line is what gets quoted back to your customer.

LinkedIn. It's one of the most cited domains in AI answers, and most ecom sellers ignore it completely. It's organic and it's cheap. Worth publishing where your buyers and your suppliers both are.

Reddit — be careful. This is where he says brands make their most expensive mistake. They go bot heavy on Reddit with a pile of AI content. Once you get flagged as illegitimate, it's hard to come back. He also thinks Reddit's weight is going to drop because it's picking up a reputation as a source of bot traffic in training data.

Now the question I know a lot of you have been asking.

Do ChatGPT ads work? Ashwin's answer was blunt. Not right now. He says he's never met anyone who got the ChatGPT ads manager to be ROI positive, and he's seeing customer acquisition costs at 5 to 6 times what they are on Meta.

TAKEAWAY: Don't move ad budget into ChatGPT yet. Do change who you're writing for. The next video and the next post should answer a real buying question in the first line, because that line is what the AI reads and repeats.

This one went live this week and I don't think it got enough attention.

Amazon Ads put its MCP Server into global open beta for partners with active Amazon Ads API credentials. In plain English: you can point Claude, ChatGPT or Gemini at your live advertising data through one integration, and ask for what you need in normal language instead of writing API calls.

It covers running reports, creating and updating campaigns, account settings, and billing and financial data.

But here's the part I want you to notice.

Amazon didn't just open the pipe and walk away. They built multi-step tools on top of it, because they say raw API access on its own doesn't guarantee a reliable outcome. And their Sponsored Products workflow stops at a ready-to-launch campaign that still requires your review and approval.

Amazon built the approval gate in on purpose. That should tell you something about how much they trust an agent to spend your ad budget unsupervised.

TAKEAWAY: Start with reports and campaign drafts. Keep launches, deletions and budget changes behind a human. If the biggest advertising platform in ecommerce won't let an agent push the button by itself, don't build your setup so yours can.

What an advanced AI operator in Ecom actually looks like (and what he's still stuck on)

One of the sellers in my AI for Ecom cohort showed us his version of his AI setup on our call yesterday. I want to share it because it's the most useful example I've seen of doing this properly, and because the second half is just as important as the first.

He pulls Amazon reports that hit API rate limits. Instead of paying an AI model to sit there and wait, he wrote a Python script with the timing built in and let it run overnight on a schedule.

The AI doesn't watch it. It checks the folder Monday morning, confirms the data landed, and only starts spending tokens if something went wrong.

In his words: "That's the only time I burned a single token."

That's the lesson. If your AI is counting, waiting, or retrying, you're paying a model to do work a free script does better and more reliably.

He also built a shared wiki so his agents can see what the others are doing.

I asked him, "Do your agents talk to each other?"

He tried letting them talk to each other directly and turned it off. His reason: "You introduce a lot of garbage. You basically pollute the knowledge base."

He's spent weeks trying to make AI video work for his product which has 3 functions. 

The problem is the AI video models keep drifting.

They swap in someone else's product or invent a different one entirely.

The only thing that gets close is feeding it reference images frame by frame.

He hasn't solved it. This is a guy running Python scripts, cron jobs and multiple agents, and he's still stuck.

TAKEAWAY: Being further along doesn't make running AI workflows for your business easy. It just makes your problems more specific.

If you're frustrated that AI isn't working perfectly for your business yet, you're not behind. That's what this looks like for everyone right now.

Online is where you learn. Offline is where you decide.

I've been thinking about this all week.

I was in Shanghai recently and sat in a room with about a dozen sellers for two hours.

One guy had agents running a real operation.

A couple of others were just trying to get better listing images.

Somebody asked the security question everyone's thinking but nobody posts publicly.

None of that comes out of a webinar chat box.

You get it when people are sitting across from each other with no reason to posture and no comment section to perform for.

That's the whole reason I run the Japan Mastermind the way I do.

7 Figure Japan Mastermind — Tokyo, Oct 13–14, 2026

Two working days in Tokyo with a small group of ecommerce operators, focused on AI agents and workflows you can actually run in your business.

Here's what we're doing:

  • Building AI agents and workflows that save time and make money — without giving up control or doing anything unsafe

  • Working sessions where you can bring your own workflow, get feedback, and leave with implementation notes

  • A room small enough for honest operator conversations with people who are actually building

  • Welcome mixer on Oct 12, 2026, and an optional Fun Day on Oct 15, 2026

10 seats. Early bird pricing is US$499 and it closes Aug 31, 2026.

In case you missed it:

Talk soon,

Gary