The summer of agentic whiplash

Good evening.

When I wrote a few weeks ago about the scandal at Phia, the AI-powered shopping startup co-founded by Phoebe Gates, the one remaining question was whether the deceitful tactics in question were the result of top-down directives or bottom-up chicanery.

“I am finding it harder to write off Phia's actions as accidental,” I wrote about the evidence at the time. “I'm now more interested in who knew what internally, and when.”

As a reminder, the startup, co-founded by Gates, the daughter of Bill and Melinda, and her college classmate Sophia Kianni, makes a browser extension and shopping app that helps online shoppers find lower prices and coupon codes for merchandise they want to purchase. But last month, the venture-backed company was caught using underhanded engineering techniques to earn affiliate commissions from online shopping sites even when a shopper hadn’t knowingly interacted with Phia’s tools before purchasing something from a retail or brand partner. That’s a big no-no in e-commerce.

Most of the feedback I got at the time was that it would be unheard of for the founders of a small startup to not, at a minimum, be aware of these types of revenue-generating tactics. Based on my 15-plus years covering startups in some form, I agreed.

Sure enough, Bloomberg published a damning follow-up on Tuesday alleging that not only did the two young founders direct some of these activities uncovered in the first investigation, but were also allegedly aware of several additional technical hacks also designed to make sure the startup got credit for sales from partnering e-commerce websites even when it hadn’t earned them.

This is obviously most embarrassing for the young founders who—at best—have seemingly violated contracts with their partners. It’s also a bad look for their venture-capital investors—including Kleiner Perkins, Khosla Ventures, and Notable Capital—as well. My educated guess, though, is that both sides are simply going to try to move forward, put their heads down, and “keep building”; the startup says it’s paying back money from misattributed sales, and hiring a head of compliance, though I wouldn’t blame you for questioning how much autonomy that person will have.

Either way, I just don’t know how partner brands and retailers continue to work with the startup. You can believe in second chances for young, dumb decisions. Or even believe this scandal is overblown because of the Gates name. Or maybe you’re one of the people arguing that this would be much less of a story—maybe even a non-story—if these founders were dudes, instead of two young women. I’ll entertain those arguments.

Still, for online retailers and consumer brands, is it worth the risk that you might be giving up revenue for no good legit reason when doing business with this startup? I’m skeptical the just-keep-building option will work. But in our current AI era, I honestly don’t know if that opinion is an obvious or naive one.

The fear now for other startups in the space is that this situation makes it harder for brands and retailers to trust any young AI-fueled e-commerce company generating revenue through affiliate cuts. Affiliate revenue is perhaps the easiest cash spigot to turn on as a young company in the space, so the collateral damage could extend well beyond Phia. Now the affiliate marketing industry has another black mark against it.

There’s a reason why deep business relationships are hard to make, but easy to burn. I feel this is especially true in this AI-everything era.

Another story on my radar:

To comply with new EU legislation, Anthropic has said that content touched by its Claude AI models may soon carry digital, machine-readable watermarks that link it to the AI service. Google and OpenAI have already agreed to similar initiatives. Yes, that means that these newsletters I write by hand—but run through Claude and other AI services for spelling and grammar checks—could carry an AI version of the scarlet letter.

The human eye can’t see these marks, but there will be computer tools that can detect them. There will be exceptions of Claude outputs that won’t get marked—such as ones that have been heavily edited, or others that are very short passages. My take? Most people will continue to judge content, including mine, on the characteristics they always have: does it inform, entertain, or make them think differently about something or someone. I don’t think people hate AI-generated writing just because AI is behind it. I think many people hate it because it’s not very good. And that distinction will remain, AI tag or not.

Now on to the good stuff…

Over the last several weeks, executives from Shopify, Amazon, Google, Etsy and DoorDash were all asked on earnings calls about, or offered comments on, how product discovery inside AI experiences is changing online shopping. The answers ranged from “not much” to transformative.

This is the summer of agentic whiplash. Or, more diplomatically, perhaps what a market looks like before there's enough behavior to make anyone's broad claims verifiable.

Over at Etsy, CEO Kruti Patel Goyal was asked point blank about traffic from AI sources and said it accounted for less than 1% of visits. She added that it does convert at better rates and with higher order values.

DoorDash CEO Tony Xu was also asked if he was seeing any significant AI-driven traffic. “In short, no, the traffic is quite low,” he said. But he offered a reason nobody else did, and one that helps reframe why some of the numbers in this piece look the way they do.