This week's news: a federal court ruled the shopping agent is the shopper, Shopify posted the receipts to prove it, stolen cards went on an AI credit spree, Washington reviewed a safety test it will not publish, and Google gave Assistant a death date.
Read the whole thing in three minutes, or go deeper where it touches your business. We talk about the market here, not about ourselves.
Every article comes with a working tool. It runs right here in the page, it is free, and we do not need your information to use it. No signup, no download, nothing to fill in first. That holds for every article in every issue, back issues included.
Send it when it hits. Every article here has its own link; the send button next to each one copies it, ready to text or email to the person who needs it. Back issues stay in the archive and their tools keep working, so nothing you forward goes stale.
On Tuesday the Ninth Circuit threw out the injunction Amazon had won against Perplexity's Comet, the browser agent that shops on a user's behalf. Amazon argued an AI agent using a customer's own logged-in account was unauthorized access under the federal anti-hacking law. The court said no: the user directs the agent, so the user is doing the accessing. The lawsuit continues, but the block is gone and the agents are shopping again.
The biggest retailer on earth just failed to keep a shopping agent out with the hacking statute, which means the question left standing is commercial, not legal: do you want software that shops for your customers inside your store or not. For a small merchant the answer is usually yes, because the agent is doing the comparison work in a venue where ad budgets do not reach. An agent that can read your price, stock, and shipping terms can recommend you.
Do not let a bot blocker you never configured make this decision for you; plenty of security defaults treat every automated visitor as a scraper. Do not copy Amazon's posture, because Amazon is defending an advertising business you do not have. And do not confuse agents shopping for a customer with bots harvesting your catalog to resell; the first is revenue, the second is the thing your terms of service are for.
Strip the case law and the holding is simple: when your customer sends an agent, the law currently sees your customer. Sites can still shape agent access with terms, rate limits, and technical controls, and the contract side of the fight is alive. But the default flipped this week. Agent traffic is customer traffic until you decide otherwise, and deciding otherwise now costs you the sale, because the same week produced hard numbers showing agent-referred buyers convert at roughly double the usual rate.
Most small merchants have never once checked what their own stack does to an agent at the door. Shopify, Cloudflare, and the security plugins all ship bot rules with defaults someone else chose, tuned for scraper abuse, not for a paying customer's assistant. The audit takes ten minutes and most owners will find they are blocking revenue with settings they did not know they had.
Shopify reported earnings Wednesday: revenue up 34 percent to $3.6 billion on $116 billion of merchandise sold, and one number that matters more than either: traffic and orders referred by AI tools both tripled year over year. Half of AI-referred sessions land directly on a product page, 2.5 times the rate of traditional search, and new buyers arrive through AI channels at nearly twice the usual rate.
The AI referral is now a measured sales channel, and it rewards different things than Google did. Agents read structured product data and match on intent, not keywords, so complete beats clever and specific beats big: 75 percent of AI-attributed purchases were outside the top 100 product categories. If you sell something particular, the machine that actually reads spec sheets is the first salesman you do not pay.
Do not buy an "AI SEO" package from whoever cold-emails you this month; the inputs are unglamorous and mostly free. Do not judge the channel by your analytics dashboard's defaults, where AI referrals hide inside "direct." And do not spend another ad dollar before your product pages state price, stock, shipping, and returns in plain text a machine can read.
The operator translation: your product page is now read by a buyer that never skims, never gets bored, and never falls for the hero image. It reads the whole page, compares it against ten others, and answers a human's very specific question. Pages that state materials, dimensions, lead times, compatibility, shipping, and returns in plain text get recommended. Pages that put the sizing chart in a photo and the price behind a quote form do not exist to this buyer.
This is the cheapest marketing work available this quarter. No spend, no agency: it is completeness, done once per product, and it compounds because every assistant reads the same page. The measurement side takes twenty minutes more: your analytics already records referrers like chatgpt.com and perplexity.ai; it just does not group them for you. Make the segment, then watch it monthly instead of arguing about whether the channel is real.
The case that made the rounds this week: a Sussex business owner's Metro Bank card was drained of 14,244 pounds in two days of Claude credit purchases, 90 to 200 pounds at a time. The bank flagged the very first charge, blocked it, then let dozens more through. Korean cardholders hit the same pattern with fraudulent ChatGPT subscriptions in June, and Microsoft has traced phishing campaigns dressed as ChatGPT, Claude, and DeepSeek notices harvesting card numbers at scale.
AI credits have become a currency fraud likes: instantly delivered, high limits, and resellable as loaded accounts. To a bank's risk model the charges look clean, since a reputable merchant selling digital goods in mid-size amounts is what normal looks like. And the noise works in the thief's favor: every business now carries a handful of legitimate AI subscriptions, so one more charge from an AI company reads as someone on the team buying credits.
Do not assume the bank catches it; Metro flagged charge one and paid out fourteen thousand more. Do not let AI subscriptions ride on cards nobody reconciles, which is where these hide best. And do not shrug off the first small unknown AI charge as a teammate's trial; the Korean pattern started at 19 dollars a head.
Nothing about the defense is sophisticated, which is the good news. Real-time alerts on every business card, set to a low threshold, turn a two-day drain into a two-charge loss. A written list of the AI services you actually pay for, with the card and the owner next to each, turns "probably someone's subscription" into a thirty-second check. Same-day disputes preserve your chargeback rights, and the bank's own conduct in the Metro case is the argument for speed: the customer who reports at charge two is a different case file than the one who reports at charge forty.
If you take cards online, you are on the other side of the same trade: fraudsters test stolen numbers against small merchants before spending them. Your processor ships defenses for this, velocity limits and checkout challenges among them, and most sit half-configured. A burst of small declines in your processor dashboard is the tell.
The White House convened Meta, Nvidia, Microsoft, OpenAI, Anthropic, and a run of smaller labs on Tuesday to review a draft framework for evaluating frontier AI models before public release, the product of a June executive order. Labs would submit models to the government up to 30 days ahead of launch. Participation is voluntary, and the administration says the framework itself will not be published.
Less than the headlines suggest, which is the point worth having. A voluntary, unpublished testing arrangement between the government and the labs creates no obligations for your business and no protections either. The rules that actually bind you are unchanged and already live: state disclosure laws, the EU transparency rules, and ordinary consumer-protection enforcement. And it carries a signal: after a July of models escaping test environments, the people with the most information decided pre-release containment testing is necessary.
Do not wait for a federal standard to settle before writing your own AI rules; the one under discussion is secret, voluntary, and aimed at the labs, not at you. Do not accept "passed government review" as a vendor safety claim, since nobody outside the room can verify it. And do not read the absence of federal rules as the absence of liability; regulators enforce today's laws against today's AI conduct.
You cannot comply with a rulebook you cannot read, so treat the framework as information rather than instruction. The information is this: the organizations that know these models best now route them through containment testing before the public touches them. Your business runs a miniature of the same problem every day, with none of the staff. The equivalent control at your scale is not a framework, it is a card on the wall: what happens in the first hour after an AI tool does something wrong.
Every incident in this run of issues, escaped models, indexed chats, drained cards, was survivable for an operator who could answer one question fast: who turns it off, and what do we keep. An incident card names the owner, the kill switch for each tool, what gets preserved (logs, screenshots, timestamps), who calls the vendor, and what you tell customers if it touched them. Writing it takes twenty minutes. Reaching for it beats improvising on the worst morning of the quarter.
Google announced Wednesday that Gemini replaces Google Assistant starting September 4, rolling out over several weeks across Android phones and tablets, Wear OS watches, headphones, and Android Auto, with Google TV, speakers, and smart displays to follow. Once a device switches there is no way back. The retirement was planned for 2025, slipped a year, and now has a date and a user email.
If your business runs on Android, the voice layer underneath your workday changes on someone else's schedule: kitchen timers, hands-free calls and navigation in the truck, the front-counter speaker, routines that announce and remind. Gemini converses better and follows deterministic commands differently, so a routine that fired identically for six years may now interpret instead of obey. You get about a month of notice, and older devices reportedly get a slower Gemini permanently.
Do not find out on September 5 from a confused crew; inventory the "Hey Google" habits first, they were never written down. Do not build elaborate new voice automations on any assistant right now, since the layer is mid-churn. And do not assume watches, headphones, and car units transition cleanly; test one of each before the fleet updates.
Voice assistants entered businesses without a rollout plan: someone said "Hey Google, set a timer for twelve minutes" in 2019 and a kitchen depended on it by 2020. Because these workflows were never deployed, they are not documented, and because they are not documented, nobody knows what breaks on September 4. The inventory is a five-minute standup question: what do you say "Hey Google" for? Timers, calls, navigation, reminders, music, intercom announcements, door locks, thermostats. Write the list; it is shorter than you fear and more load-bearing than you think.
Then stage the test instead of taking the update blind. The rollout runs for weeks, which is a gift: update one phone, one watch, one speaker early, run the list against Gemini, and rebuild or simplify what breaks before the rest of the fleet turns over. The strategic move is the same one the AWS story taught: fewer dependencies on any vendor's assistant layer, not more, until the churn settles.
The Briefing arrives weekly: what moved, what it changes, what not to do, and a working tool for each. Every issue gives the tools away, free and built for the week they cover. No product tours, no victory laps. If it stops earning the read, unsubscribe.
This is back issue No. 004. Every tool in it still runs. Read the current issue · Open the archive
If something in here lands on a problem you are already working, and you would rather not work it alone, write to hello@goudegroup.com and tell us what you are exploring. You get a reply from a person, not a sequence.