Google Announced “Honey” Dipped in AI

by Rian

Just last year, browser extension Honey suffered an infamous scandal. Now owned by PayPal, the extension promised to find the best deal, including finding the best coupon codes. The exposé revealed Honey’s business model focused more on being a man in the middle in both security and transaction. And it also reveals a gap, a bias between what buyers think they do and what they actually do.

To understand why, it helps to look at how Honey actually operated — and MegaLag’s three-part investigation is still the most thorough account available. In security terms, Honey was a man-in-the-middle. It inserted itself between buyer and seller, presenting itself as the buyer’s advocate while quietly extracting value from both directions. The extension replaced affiliate codes from creators and publishers with its own, claiming commissions it had no part in earning. It harvested coupon codes from its own users and redistributed them without consent. From the consumer’s side, none of this was visible — if a discount applied, Honey got the credit; if it didn’t, you just paid full price. But for the merchants, the exposure was harder to escape. Honey maintained a database of over 180,000 websites. Only 35,000 had formal agreements. The remaining 150,000 appeared without authorization — and some were told to become paying partners when they asked to be removed.

To understand how and why, it helps to look at Honey’s MO. MegaLag’s three-part coverage is still the most thorough account available. Honey was a man-in-the-middle. It inserted itself between a buyer and a seller, presenting itself as the buyer’s advocate. The extension replaced affiliate codes for influencers with its own, claiming their commissions. It harvested coupon codes from its own users and redistributed them without consent. The true genius of this insidious scheme is it is invisible to the user — if it works, Honey gets the credit; if it didn’t, price remained unchanged. But for the merchants, Honey did not let go of them so easily. Honey maintained a database of over 180,000 websites. Only 35,000 were affiliate partners. The remaining 150,000 appeared without consent — and some were told to become a partner when they asked to be removed.

What Honey shares with many AI products that followed is less a coincidence and more a pattern. Users remain largely unaware of what runs beneath the surface. Data — or in Honey’s case, codes shared unknowingly by its own users — gets absorbed and redistributed for the extension’s gain. And when the visible cost fell disproportionally on the laps of the producers: the sellers who are suddenly hit with fees and unauthorized discounts, the influencers who used the ad space only to be taken. Consumers, meanwhile, felt nothing. The scheme was invisible by design. And it is no easy feat to sue over supposedly invisible harm.

Honey’s model worked because it offered the satisfaction of “finding” the deal. But experienced online shoppers know price is rarely the whole story. Amazon — the largest e-commerce platform in the US — is not particularly known for coupons. It’s known for fast shipping, vast selection, and a return policy that removes risk from the transaction. Every platform has its own version of this trade-off. The cheaper option usually means absorbing more risk yourself. For first-timers, these are the perks that are hard to appreciate without learning the hard way.

Let me offer an example — one Google itself touched. At I/O 2026, Universal Cart was demoed catching a socket incompatibility during a PC build. Any PC builder would read the product page instead of opting to ask an AI. Published specs and details are least of their worries. It’s the ones that are not documented: whether the GPU clears the case, whether the CPU cooler clears the RAM, or whether the AIO pipes reach the radiator mount. Those answers are found in trial-and-error and live in community threads. To truly innovate shopping on the level the marketing literature is flirting at, the AI will need to deduce the clearance and physical dimension of each parts. Otherwise it will be yet another dumb crawler.

People do not simply go online and shop for an “item.” Even a roll of toilet paper comes with preferences. The idea that AI can meaningfully handle the research part of purchasing assumes that product descriptions and specs are what drive decisions. They rarely are. The merit of a given product is often impossible to convey in text. Take the PC building analogy. It’s easy to say “find me the most silent cooler.” But any builder would immediately ask: silent how? Liquid or air? What case clearance? What TDP? The question contains multitudes that a search query flattens. Just like AI, it’s about asking the right questions — and knowing which ones you don’t yet know to ask.

Honey is still going through several lawsuits. But one thing stands out — the majority of plaintiffs are influencers and creators, not the sellers or small businesses whose codes were scraped without consent. This is not unlike how AI models are trained: data absorbed without asking, and near impossible to remove once it’s in. For sellers who found themselves in Honey’s database without agreement, the options were limited. For those who find their data in an AI’s training set, the options are fewer still. If Universal Cart is truly AI-driven as Google says it is, there is nothing in the announcement to suggest sellers would fare any better. Honey never asked merchants to partake. AI is even less likely to. After all, dead businesses and bad deals tell no tale.