The AI story of the week is that the AI story didn't ship. Google's Gemini 3.5 Pro — the launch I told you last Sunday to have a test ready for — missed its third deadline. Meanwhile, in the same seven days, Amazon published four dated, unglamorous line items that will each touch a real number on your Q4 P&L: a peak fulfillment surcharge, a rebuilt launch-incentive program, a delivery-rate mandate with deactivation teeth, and an attribution upgrade for anyone selling beyond Amazon.
Nobody is posting threads about any of those four. Everybody is posting threads about the model that doesn't exist. Here's the week, ranked by what it means for someone running $50K–$500K/month on Amazon.
1. Gemini 3.5 Pro missed its third deadline — and the people who waited for it lost again
What happened: July 17 came and went. TechTimes reported on July 16 that the rebuilt model failed Google's internal reliability bar — frequent hallucinations, benchmark results short of GPT-5.6 — and that Google is now weighing a stopgap release. As of July 18 there is still no model card, no pricing, no API listing.
Operator implication: Last week I wrote that if July 17 held, the move was to run your prompts head-to-head and keep the winner as a one-line config change. It didn't hold, and here's the part worth internalizing: if your model is a config variable, this cost you nothing. You've been running on GA models — GPT-5.6, Sonnet 5, whatever wins on your SKUs — the entire time. The only operators who lost this week are the ones who put "wait for Gemini" on a roadmap, and they've now burned a month waiting on a vendor promise for a model that has missed June, early July, and July 17. Three misses plus a benchmark shortfall against GPT-5.6 also tells you something bigger: the gap between frontier labs is narrow enough that a full rebuild couldn't clear it. Whatever ships eventually, the price pressure on your AI tooling is already here from the models that exist. Schedule work around those. Test launches the day they're real, never the day they're promised.
2. Amazon set the 2026 peak fulfillment surcharge — and it's keyed to ship date, not order date
What happened: On July 15, Amazon announced in Seller Central that a peak fulfillment surcharge averaging +$0.32 per unit applies to every FBA shipment leaving fulfillment centers between October 15, 2026 and January 14, 2027 — stacking on the 3.5% fuel and logistics fee added earlier this year (Nova Analytics has the rundown).
Operator implication: The number is small per unit and large in aggregate: 10,000 peak-season units is $3,200 straight off contribution before you've discounted anything. The detail most sellers will skim past is the trigger — ship date, not order date. An order placed October 12 that ships October 16 eats the fee. What I'd actually do this week: rerun your Q4 deal math with the surcharge and the fuel fee baked into unit economics before you submit deals. Prime Big Deal Days submissions are due in early August; a deal you grade at July's fee structure and run at October's is a deal you priced wrong on purpose. If you're on the fence about a marginal SKU's holiday promo, this $0.32 is exactly the kind of line that flips it.
3. The New Selection Program (2026) launches July 30 — the cheapest launch window of the year just got cheaper
What happened: Amazon announced in Seller Central that an expanded New Selection Program takes effect July 30: reported benefits include referral-fee credits on the first units sold of eligible new branded parent ASINs, free storage on the first 200 units for up to 120 days, larger inbound-placement credits, and — new this cycle — relief on coupon and Vine costs. The IPI eligibility floor drops from 400 to 300. Currently enrolled sellers get the new benefits automatically for branded FBA ASINs launched July 30 through October 31; ppc.land reports you must confirm enrollment to keep benefits after October 31.
Operator implication: I've been telling clients the launch-by-mid-August-or-wait-for-January rule for weeks. This changes the math on the first half of that sentence. A launch that goes live in August now stacks the honeymoon window, pre-peak CPCs, and a fee-credit package that directly subsidizes the two most expensive parts of launching — early storage and review velocity. Vine enrollment relief in particular hits the exact budget line I tell every launching brand to protect. The specific benefit structures vary by account, so read your own terms in Seller Central rather than a blog's summary — including mine. Then put October 31 on the calendar, because auto-applied benefits that require a manual confirmation to continue are how sellers quietly lose free money.
4. Amazon Business orders get a 90% delivery-rate mandate — with offer deactivation attached
What happened: Also announced July 15: starting September 30, professional sellers must maintain a 90%+ delivery rate on Amazon Business seller-fulfilled orders, or face deactivation of those offers.
Operator implication: If you're pure FBA, skip to item five. If you're hybrid — FBA on heroes, FBM on the long tail or as a stockout backstop — this is the quiet one that bites, because for most brands the B2B slice is invisible. Business customers buy through the same listings, in bigger quantities, and almost nobody breaks them out in reporting. A deactivation on those offers won't announce itself as a policy event; it'll look like a soft month on SKUs you weren't watching. You have ten weeks. Pull your seller-fulfilled delivery performance now, check it against the 90% bar, and if your FBM operation is riding close to it, fix the carrier promise or move the exposure to FBA/MCF before the enforcement date does it for you.
5. Amazon DSP now shows which retailer closed each offline sale — attribution gets sharper exactly where Amazon sells you ads
What happened: On July 13, Amazon's DSP Omnichannel Metrics added a per-retailer offline breakdown — attribution now shows which specific retailer closed each offline sale, replacing the category-level grouping it launched in April.
Operator implication: If you're one of the brands expanding beyond Amazon — Walmart, Best Buy Marketplace, wholesale into physical retail — the perennial question "did my Amazon ad spend just sell a unit at someone else's register?" now has a real denominator. That's genuinely useful for deciding whether DSP is a brand-level investment or an Amazon-channel expense. But notice the shape of the week: Amazon shipped precise attribution for the traffic it sells you, while the traffic its own AI assistant routes organically stays blended into your sessions with no native split. Measurement follows the money. Amazon will instrument the paid path to the decimal point; the unpaid AI path is still yours to instrument yourself — weekly, manually, on your top ASINs — because nobody's building that dashboard for you.
What I'd ignore
The DeepSeek V4 countdown content. V4 graduates from preview to stable July 24, and the takes are already flowing. Same rule as always: don't rebuild on a preview, and don't rebuild on a stable release either until it wins on your prompts. The 2M-token context debates. Speculating about the specs of a model with no model card is fantasy sports. GPT-Live voice demos — still a real Alexa-for-Shopping tailwind eventually, still nothing you'd change on a listing this quarter. Claude for Teachers (July 14) — a good program that has nothing to do with your catalog. And the UK VAT headlines — HMRC's proposal to have marketplaces collect VAT directly would cut UK payouts meaningfully if adopted, but it's a proposal. One caveat: if the UK is a real chunk of your revenue, that's the one item in this section you should actually read rather than ignore. Watch it; don't reprice off it.
Three weeks ago I wrote that the frontier had become a spectator sport. This week was the cleanest demonstration yet: the loudest AI story of the week produced nothing you can use, and the retailer you actually sell on published four dated changes with dollar signs attached. The frontier missed its deadline. Amazon didn't miss any of theirs. Keep score where the numbers are.