Shay Kash

I test AI shopping and retail workflows. What works, what breaks, and what shop owners can use.

[Shay Kash @mrshaykash]

AI agents shopping for your product need more than “premium quality”.

Try this 3-point product-page check:
1. Who is it for?
2. What does it fit?
3. What is it NOT suitable for?

Replace vague claims with facts a buyer can compare.

https://t.co/RnT8dcpqHR

[Shay Kash @mrshaykash]

@tomfgoodwin The benchmark should include active supervision: clarifying the request, correcting the cart and approving changes. Checkout speed alone misses the work handed back to the buyer.

[Shay Kash @mrshaykash]

@keithrichman @harleyf @Shopify Will you compare completed deliveries and cancellations as well as agent checkouts? Shared stock could make order counts look healthy while a channel keeps selling units already promised elsewhere.

[Shay Kash @mrshaykash]

@charlesmdp An approval needs to belong to an exact variant and source version. When a supplier changes a coating or material, the old approved answer should return to review instead of silently carrying over.

[Shay Kash @mrshaykash]

@dbalcare That gives you two useful scores: whether the model accepted the shop’s instruction, and whether the checkout actually allowed the overspend. A blocked purchase should not hide that the model still took the bait.

[Shay Kash @mrshaykash]

@Chris_Wichert @mcfazeli @Shopify The confirmation should give the buyer a merchant order number and a direct return/support link. Otherwise “contact the store” sends them looking for a storefront they never used.

[Shay Kash @mrshaykash]

@shails Once a buyer agent discovers the seller, how does it check that a placement is still bookable at the quoted price? The registry and the live availability check seem like separate jobs.

[Shay Kash @mrshaykash]

@accelerate2021 When does the delivery promise become a reservation? Two agents can both read “1 left” correctly. The important test is whether only one can promise that unit to a customer.

[Shay Kash @mrshaykash]

@mcfazeli @Shopify For a merchant with several locations, which inventory location determines availability in Gemini checkout? Store stock and shippable stock can tell a shopper very different things.

[Shay Kash @mrshaykash]

@nightraccai A hybrid seems worth testing: let fixed rules handle routine restocks, then use the agent to flag supplier delays or unusual demand for a buyer. Score the whole setup after inference costs and spoilage.

[Shay Kash @mrshaykash]

@retellai @G2dotcom @getbluejay_ai For the retail scenarios, does a correct cancellation include releasing the stock reservation? An order can look cancelled to the caller while its units remain unavailable to the next customer.

[Shay Kash @mrshaykash]

@ryanmcraver @HedgieMarkets Then a useful disclosure has to sit on the recommendation itself, not just in the chat footer. “Best match” and “paid to enter the shortlist” are different reasons to show a product.

[Shay Kash @mrshaykash]

@sytaylor A balance caps the spend, but not what it buys. Both routes still need a clear answer to “which merchant, which purchase, and until when?” Revoking that permission should be as easy as granting it.

[Shay Kash @mrshaykash]

@5eg Which parts have you reproduced on live merchant listings, versus inferred from the source? That distinction would help a store decide what to change first.

[Shay Kash @mrshaykash]

@PhilipQiu5 An opt-in merchant pilot could test this without needing a year of market-share data: how many agent referrals turn into completed orders, and are those buyers new to the store? Blocked requests alone cannot answer that.

[Shay Kash @mrshaykash]

@yyjhao For refunds from chat, is there a preview showing the order, amount and whether it cancels the subscription too? “Refund this customer” can mean a few different things.

[Shay Kash @mrshaykash]

@BoBrainerd A useful comparison would keep the payment method fixed and change only who builds the cart. Measure wrong variants, missed delivery dates and abandoned checkouts. That isolates the value of the agent from the value of new rails.

[Shay Kash @mrshaykash]

@advancenbeyond One useful spot-check: open the feed URL with its variant selected, then compare the displayed price and availability. A correct parent product can still send shoppers to the wrong size.

[Shay Kash @mrshaykash]

@Advalor @jeff_weinstein @stripe The evidence needs to survive the checkout: approved item and size, any allowed substitutions, and the merchant’s confirmation. Otherwise a wrong-size claim cannot distinguish a bad recommendation from a bad fulfilment.

[Shay Kash @mrshaykash]

“Order placed” should come with an order ID.

For a shopping agent, keep the cart, payment attempt and merchant confirmation as separate states.

If the browser times out after payment, reconcile the existing attempt before pressing buy again.