Chew on this.
On AI, automation, software and search. Written from what I build and test, with the dates and the numbers.
- Grab says AI agents can do your groceries. Fix your product data now, not your agent strategy.
Grab thinks AI agents can handle grocery shopping, but analysts point to thin retailer margins and costly AI tokens. Whether agents scale in Southeast Asia is unsettled. A small grocer can still publish precise, current product, price and stock facts today.
- Microsoft's outages followed its AI layoffs. Cut the work AI does, not the people who check it.
Outlook, Teams, Copilot and more than a dozen other Microsoft services have gone down after the company's AI-linked layoffs. If you use AI to do more work, keep the people who verify and own each process.
- The FAA built an AI tool without its controllers' union. Involve the people who run the process before you roll out AI.
Airlines praised the FAA's AI congestion tool, but the controllers' union wasn't involved. A business rolling out AI risks the same result if it skips the staff who run the process.
- Is fraud prevention worth the cost? Put hard rules where wrong is costly, and AI where judgement is needed
A growing share of fraud professionals rank cutting cost above avoiding loss. I think the hard limits belong in rules the system enforces, AI can flag odd cases for a person to judge, and the owner should compare prevention cost with actual loss.
- The same B2B keywords now cost about 21% more. Build distribution that compounds instead of renting clicks.
One agency's data on 14 B2B advertisers says the same non-brand Google Ads keywords cost about 21% more per click. It is a small sample, but a bootstrapped business should read it as a reason to build visibility in AI answers.
- AI DJs are ruining playlists, and voice works for input, not for output nobody asked for
YouTube Music and Spotify want synthetic hosts talking over your playlists. Voice helps when you dictate instructions. On the output side, it adds little unless a customer asked for it.
- Google admits your analytics can't see all its search traffic. Check whether the answer names you instead.
A new Google help page says some Search visits arrive with no referrer and that analytics counts won't match Search Console. Clicks are a blurry, lagging number, so check by hand and with dated records whether the AI answer names you.
- ElevenLabs' new app does dictation and meeting notes. Dictation is the part of voice that works.
ElevenLabs' ElevenAI bundles meeting notes with dictation. Dictation works because talking instructions to your AI is input, so judge any notetaker by whether its notes turn into work.
- The AI graveyard is full of launches nobody needed. Ask what problem it solves before you adopt it.
TechCrunch's list of AI projects that shut down or missed expectations has a common thread. Many were built around what the tech could do. Before you adopt a new AI product, name the problem it solves that your current tool doesn't.
- Lovable apps leaked passports and 200K messages. Vibe code your own tools, not your customers' data.
Researchers found Lovable-built apps exposing passport details and over 200,000 private messages. Vibe coding is fine for tools that hold only your own data. If you build for customers, a named person must review how it protects their data before launch.
- 72% of workers say their AI skills are untested. Ask for the record of what they overrode and checked.
A survey says most workers' AI skills have never been formally evaluated. The proof of AI skill is a dated record of where someone overrode the tool, what they checked and what happened. Operators should ask hires and agencies for that record, not a list of tools.
- ChatGPT now shows businesses their mention and click counts. Watch the mentions, date every reading.
OpenAI's Business Hub gives businesses the first vendor-supplied count of how often ChatGPT mentions them. The mention count is the number to watch, the click count is the small side, and both need a date.
- If AI solved coding, why is your software still the same? Because code was never the hard part
Vibe coding gets you from idea to working software fast. It does nothing for testing with real users, getting context right or getting customers, and those decide whether software succeeds.
- Most AI agent traffic is reading product pages, not checking out. Look at your logs, then fix the pages they read.
HUMAN Security found that 85% of AI agent requests in October hit product and search pages, and 1.9% hit checkout. A small business should find out which agents visit its site and which pages they read, then make those pages precise and different from competitors'.
- Google keeps moving the AI search buttons. Don't build your plan on one of them.
Google is testing the removal of AI Mode from the search tab bar and the Ask anything box from AI Overviews. A small business should get named clearly wherever answers are assembled and keep a dated record of what each engine says.
- Shield's agents do the legacy portal work. Automate the process your staff retype, not the chat
Shield uses OpenAI so its Forge agents work inside legacy portals and technicians keep the issues that need judgement. That is the useful kind of automation, and any operator can start by listing the portal tasks their staff repeat.
- Amazon wants to index every product online. Make sure the record of yours is one you wrote.
Amazon is reportedly building a data vault on every product sold online, including products it doesn't sell. For a small seller, the practical answer is to publish precise, measurable facts on your own pages.
- A block list can keep you out of AI answers. Decide which bots you want reading your facts.
Blocking an AI training crawler and blocking a retrieval bot have different consequences. A blanket block list can remove a small business from the answers it wants to be named in, so set a deliberate bot policy instead.
- Another AI agent for your life? Ask what problem it solves that your current tool doesn't
Claude Duet is another consumer agent aimed at people who don't write code. Before adding one, ask whether the tool you already use deeply can do the job.
- ChatGPT can now check out for shoppers. The rollout is tiny, so fix your product data first, not your checkout
ChatGPT Wallet lets agents buy for users, but only a limited group of US Pro subscribers can use it. A small merchant should spend its effort making price, stock, shipping and returns precise enough for an agent to buy without guessing.
- A clinic chain let AI answer the phone so staff could face the patient in front of them
American Family Care gave routine calls to voice AI and kept its front desk for patients in the clinic. That is the shape I expect: transactional work goes to AI, and people go where a human experience matters.
- The traffic pool is shrinking. Stop asking which channel refills it and fix the offer agents read.
Agents now finish the job a visit used to start, so a small business gains little by chasing the channel that refills the traffic pool. The work is making the offer precise and measurably different, wherever agents read it.
- AI didn't replace the hires a small team couldn't afford. It gave the owner more to direct, and more to check.
Stories of Asia uses AI to cover design, sales and research work it has no one hired for. The useful detail is the checking. Operators should pick one tool, name the job, and decide who verifies the output.
- Keep your own dated record of whether AI names you. Don't hand it to a tool
AI visibility tools are a young and unstable market, so a small business should own its prompts, its dated raw answers and a simple record of whether it was named. The tool is only a counter.
- AI agents 'discovered' a Maya site that was already on the map. Check the claim before you publish it.
A Google researcher said a team of AI agents found an unmapped Maya site. It had been published in an academic journal the year before. If agents produce your research, leads or reports, someone has to check each claim before it goes out, and you answer for it either way.
- A $300 embroidery machine wanted a subscription, so its owner built the app himself. Consumer utility software should worry.
A Show HN project replaced a machine's companion app with free software its owner built. Personal tools like this are where vibe coding is real, and consumer utility apps are the ones under pressure. B2B software is a different case.
- AI chatbots are recommending fake shops. If you don't state where to buy from you, someone else will.
A study found ChatGPT, Gemini and Perplexity can send shoppers to imposter sites. A small seller's defence is to publish its own facts where agents read, and to check what the chatbots say about it.
- Holiday Extras raised booking conversion with Codex. The customer expertise did the work, not the tool.
Holiday Extras credits Codex for better apps and higher booking conversion. I read the gain as coming from what the team already knew about its customers. Codex shortened the path from idea to working app.
- Getting named by AI is the easy part. Specific facts are what keep you on the shortlist.
A Search Engine Journal case shows a company named in 98.4% of AI answers while its pipeline shrank. Small businesses should publish specific, dated, measurable facts on their own pages and get other sites to repeat them.
- Google is moving AI Overview sources to the bottom. Get named in the answer instead.
Google is testing source icons at the bottom of AI Overviews, which makes a citation an even weaker signal. A small business should check whether the answer says its name.
- An AI agent can hand over working code and still lose the decision that made it right. Write the decision down.
A finished task can quietly undo the decision it was built to serve. Brief every agent with the decisions and the reasons, check the output against them, then fix the brief.
- 72% of Singapore firms can't scale their AI pilots. The pilot was a demo, the workflow was never named.
Kyndryl says 72% of Singapore firms struggle to scale AI pilots. I read that as a workflow problem. A pilot only scales once someone names the process, the owner and the number it should move.
- The most expensive AI mistakes read clean. Name who checks before the wiki goes live.
An AI-written wiki article that sounds confident is the error nobody thinks to check. Put a named person on anything AI turns into a source of truth, and put the most care where being wrong costs the most.
- Indie judges ban any AI use. The tool was never the problem, the missing owner is.
Some IGF jurors say they won't vote for any game that uses generative AI in any capacity. A blanket ban treats the tool as the risk. Operators should set who is accountable for the output.
- United won't let your AI agent book a flight. It also hasn't seen anyone try.
United is blocking personal AI agents from booking flights, yet says it has seen few people try. A giant can block without demand evidence. A small business should first check whether any customer is asking through an agent.
- The Philippines' BPO industry is rewriting its roadmap. Automate routine support and keep your people for the moments that matter
The Philippines' own BPO industry body has cut its 2028 workforce target. I read that as transactional support work moving to AI. Operators should automate that layer as far as it goes and put people on supervision, customer success and the moments where a human is the premium.
- Three CX giants bought AI vendors in a week. If AI is the only pitch, expect your tool to be next.
Salesforce, Dialpad and Five9 each bought an AI vendor in one week. The chat layer is being absorbed into bigger suites, so judge support tools on resolved conversations and actions in your systems, and keep contracts short.
- Claude can now run your meeting. Ask whether the plan becomes work, or just a tidier summary.
Claude Meetings drafts plans and pulls from connected apps during a call. I would judge it by what got done after the call, and by whether one assistant you already use well beats a separate meeting tool.
- Mars is worried AI agents will kill the impulse buy. An agent has no shelf to be tempted by, only facts to read.
Mars is preparing for AI grocery agents by chasing visibility in LLMs. A small food brand can't copy its immersive-store spending, and doesn't need to. It needs detailed, differentiated product facts that an agent can read.
- Google is testing AI Overview local packs without the call button. Your name and facts in the answer now do that job.
Google is testing a local pack inside AI Overviews with no action buttons, photos or reviews. That leaves the text of the answer as your storefront, so your name, services, hours and prices need to be precise enough to be quoted.
- When AI agents review every payment, subscriptions live on proof of value, not customer inertia
If a customer's agent checks every recurring charge, a subscription survives only on facts it can read and verify. A small subscription business should publish precise, measurable data about what the customer gets and uses.
- Airbnb cut support cost per booking 17%. Measure your AI against a number your business already tracks.
Airbnb judges its AI support assistant by cost per booking, a number the business already lives by. Operators should pick a metric like that instead of counting tickets deflected.
- ChatGPT's Business Agents will answer for your brand. Ask whether they can act in your systems.
ChatGPT will let shoppers talk to verified Business Agents that represent a brand. An agent that only answers comparison questions is still a chatbot, so operators should ask whether it can act and whether its facts are precise enough to win.
- eBay shut the door on AI agents. If you sell there, keep your facts where agents can still walk in
eBay appears to have blocked third-party AI agents, which is the split between open and closed platforms I expected. The giants can wait it out. A small seller can't, so the product and brand facts need to live on pages you control.
- Google admits AI Mode pays less. The discovery stage moved, so check whether the answer names you
Google's own executives now say AI Mode earns less for ads and that queries are moving up the funnel. Before reshuffling PPC budgets, check whether AI answers name your brand at the discovery stage.
- AI made the footage cheap. Your costs moved to the directing and the checking.
One AI production worker made about two hours of drama in a week and the client accepted it. Tools took most of the direct cost, but the real economics sit in the directing and checking.
- Every CMO says AI changed marketing. Ask what the agent did on its own last week.
BCG's CMO survey shows a gap between saying AI has changed marketing and naming what an agent did unprompted. My test is to name one workflow the agent runs end to end in your systems. If you can't, you've bought a label.
- CodeRabbit ships 20x more with the same team. Ask what the extra 19 were for.
CodeRabbit says Codex lets it ship 20 times more software with the same engineering team. Output is easy to count, but an operator should ask what customers got and whether checking and deciding kept pace.
- Another Claude model launched. Ask what your business can do now that it couldn't last month.
Anthropic's Fable 5.2 and Mythos 5.2 are said to take on daunting technical challenges. I'd skip the benchmarks and ask what job last month's model couldn't do that this one can. If there isn't one, keep going deeper with the tool you have.
- Kuken found a repair drawing in 4.5 minutes instead of an hour. The 70 years of drawings were the asset.
Kuken's faster drawing search came from records the business already owned, not from the model. Operators should look for the records their staff keep digging through and point AI at that search first.
- Automattic's Spacefast makes deploying a vibe-coded app easy. Getting anyone to use it is still the hard part.
Spacefast gives vibe-coded apps a place to run, with a database, login and email built in. Hosting was never what decided whether these apps succeed, so treat it as a faster way to ship a prototype and spend your effort on distribution.
- Google is adding booking links to profiles without asking. Check where yours sends customers.
Google appears to be adding Book Online buttons to some business profiles, and some of them lead to a 500 error. Check where yours goes, then make sure customers can also book in your own chat.
- 1.54 million daily users says people showed up. Ask what the agent finished.
DeepSeek's agent platform reports 1.54 million daily users. Daily users count how many people showed up, not what the agent completed. An operator should judge an agent by the processes it finishes and checks in their own systems.
- Google's AI opt-out is a publisher's dilemma. For an operator, being named in the answer is the goal.
Publishers are weighing Google's new AI opt-out while Google makes AI Overviews more prominent. If your growth depends on being named in answers, opting out is a cost. Spend the effort on making your brand quotable and on measuring whether it is named.
- The numbers said stop four months ago. Set the kill line before you launch.
A founder kept a failing product line going for four months with the numbers already clear. I would set the stop threshold before launch, let a dashboard do the counting, and keep the decision to stop for myself.
- NYT Cooking's AI saved the chicken and nearly ruined the cake. Check hardest where wrong is costly.
A cooking assistant that rescues a weeknight chicken can still wreck a cake, because savoury cooking forgives swaps and baking chemistry doesn't. Put AI on the tasks where a wrong answer is cheap, and enforce rules or a human check where one wrong substitution, price or promise costs you.
- Agent optimization is the new SEO, but the hints are the cheap part. Precise facts are what get you named.
Cloudflare's CEO calls helping sites attract AI agents a new version of SEO. A small business should switch on the cheap, automatable layer and spend its real effort on precise, current, differentiated facts.
- Tripadvisor says search is still hurting growth. Page one was always rented traffic.
Tripadvisor blames search changes for a $369.5 million writedown and now spends nearly half its revenue on marketing. A small business should stop treating page one as the goal and check whether AI answers name it.
- Who owns an AI agent's mistakes? You do, so set what it's allowed to commit to before it runs
The business owns what its AI agent commits to, so the useful question is what the agent was allowed to commit to. Cap refunds, prices, promises and dates in the system itself, so the AI's judgement never has to carry the risk.
- Delivery apps now recommend as well as deliver. Your item data decides whether they recommend you.
DoorDash's $17 billion Instacart deal shows delivery apps becoming places where shoppers discover products. For a small merchant, the item data on each app is what gets read and recommended, so it needs to be precise.
- Claude Code leaked passwords into 1,122 public GitHub files. Vibe coding is fast, but the security check is still yours.
Apiiro reports that Claude Code's "Don't Ask Again" option left passwords in 1,122 public GitHub files. Vibe coding makes building fast, but nothing in a prompt tells the tool which files are secrets, so a person and a system rule have to check what gets published.
- Google Maps now has a chat button for your business. One more inbox is a cost unless the same automation answers it.
Google added a Message button to Maps listings on Android and iOS. It only helps a business if those messages are resolved in the thread by the same automation as every other channel.
- Google wants AI Mode to know each searcher personally. One check of whether it names you no longer tells you much.
Google is pushing AI Mode users to switch on Personal Intelligence, so two customers asking the same question can get different answers. One check of whether AI names you is now a single sample, and the fix is to publish more precise, varied facts and log dated results.
- A bot for every chore makes the chores lighter. The checking is now the job.
Beardbrand's founder runs Grok Bot for email, reports, Klaviyo flows and SEO. The chores got lighter, and the owner picked up a new job: briefing each bot and checking everything that goes out under their name.
- An AI visibility score is a vendor's number. Test it against two of yours: were you named, and did a customer say so
Search Engine Journal says an AI visibility score has to be checked against customer behaviour before it earns a budget. I agree. The two checks worth running are whether the answer names you, dated, and whether customers arrive saying the AI sent them.
- Gemini will write your Google Business Profile. It will write your competitor's too. Only your facts tell them apart.
Google is bringing Gemini's Business Profile tools to Workspace accounts. When every listing in your area comes out of the same model, the adjectives match. The hours, prices, services and exclusions you feed it are what still separate you.
- Cava made hospitality a contest and two-thirds of stores moved the score. Put your people where the human touch is the product, and measure it.
Cava ran a summer contest among its general managers to raise hospitality at the counter, and about two-thirds of locations saw guest metrics improve. That is where people belong once routine work is automated: the in-person moment, run like any other operating number.
- McDonald's is moving reviews into its app. Your customers already told you in the chat. Read it.
McDonald's is right to collect feedback where its customers already are. Most operators already hold richer feedback than any survey, in their support chats. The human job is reading those conversations, not building another form.
- Reformation's AI shows up in its in-stock rate. If your AI can't move a number you already track, it's a demo.
Reformation put AI on one operating decision, what to stock, and tied it to numbers the business already watched. That is the test for any AI project: name the number first, then see whether it moves.
- Brave's index now feeds several AI engines. One precise page serves all of them.
Brave sells one search index to Cohere, Mistral and Perplexity. The page you wrote for Google is now read by several engines at once, so being in the index gets you in the door and the precision of the page gets you named.
- AI use is up and AI spend is flat. Buyers are working out how much tool they actually need.
Ramp's data shows companies using more AI while paying the same for it. I read that as prices falling and buyers getting disciplined, and the right measure of your own AI budget is how many workflows run end to end on it, not how much you spend.
- Cult.fit's teams are building their own tools. Build the back office yourself, buy what touches the customer.
Cult.fit staff building internal tools with Codex is the business version of the personal-tool shift. A tool that replaces a spreadsheet is now cheap to build yourself. Anything that touches customers, payments or delivery still gets bought and integrated.
- Amazon's agent takes orders by text and sells ads in the replies. Inside the wall, the agent works for the landlord.
Amazon blocked Meta's shopping agent, then shipped its own with Sponsored Products in the replies. That is the closed half of agentic commerce, and it tells a seller exactly where the work has to go.
- ChatGPT will let you claim your business. Claiming gets you a profile. Precise facts get you named.
OpenAI is building a Search Console-style tool where businesses claim their listing in ChatGPT and watch mention counts. The claim is free. What you fill it with decides whether ChatGPT names you when a buyer asks.
- Akai and Moog are letting customers vibe-code their own mods. Give away the toolbox, keep what can't be built.
inMusic has stopped fighting modders and is handing them the inner workings of its instruments. That is the right posture. The small utilities around a product are now cheap for customers to build themselves, so an operator should keep the hardware, the platform and the customer relationship, and let the edges go.
- Databricks says the AI ROI debate favours data companies. Your return lives in what you've written down.
Databricks is right that the question of AI's return has become a question about data. For a shop, a clinic or a marketing team, the data that decides the return is your own prices, policies, products and workflows, written down clearly. No platform supplies that for you.
- Your outsourcer says AI plus humans is working. Ask what share now runs with no human.
TaskUs says its blend of AI and human agents is working. An outsourcer paid per seat has every reason to keep a person in every conversation, so the number a buyer needs is the share of enquiries that now close with nobody on the line.
- Muse can talk now. Your customers' agents won't be listening.
Meta gave Muse a voice and an expressive face. Voice helps when you give instructions. It does much less on the output side, because the conversations that turn into orders and bookings are moving to agent to agent, and agents trade data, not speech.
- Facebook's six million AI pages are what volume without a frame looks like
Google has indexed over six million AI-generated Facebook Answers pages, and the SEO world is asking whether that counts as scaled content abuse. Whatever Google rules, the lesson for an operator is about the claim behind each page, and that part is yours to control.
- Perplexity is selling AI teammates. You don't need more agents on the roster, you need one you know deeply.
Perplexity's always-on Teammates are a pitch to add another agent to your stack. Before you do, ask which workflow it would run end to end that your current tool cannot. If the answer is none, go deeper on what you already have.
- AI customer service only has a cost crisis if you pay per resolution
Gartner says AI vendors are about to stop subsidising their prices. The bill only turns into a crisis when you pay per resolution, because then the better the automation works, the more you owe. Price it flat, like infrastructure, and judge it against what the same conversations would cost in people.
- Box has an AI workflow builder. The test is whether you describe the process or still have to draw it
Box has expanded its Nvidia partnership and unveiled an AI workflow builder. A workflow builder only changes an operator's day if you can describe the process in your own words and it then acts in your systems. If it is still a canvas of branches with an AI step added, it is the old flow builder under a new label.
- Your agency's 2027 report needs one number first: how often the answer names you
Search Engine Journal's inclusion, framing and endorsement ladder is the right direction. For an operator it collapses to one dated number at the top of the report: how often the AI answer names your brand. Citations lead it. Rankings are a relic.
- Nine in ten Singaporeans use AI and only a third are comfortable. The gap is mindset, not tooling.
Digital 2027: Singapore shows AI habit running far ahead of AI confidence. That is what happens when people pick up the toolset before the mindset. The fix for a business is a working frame, not another tool or a course on where the buttons are.
- AT&T is bringing its contact centre home. That only pays if AI has already taken the routine calls.
Reshoring makes every human conversation cost about three times more. It only works when AI already handles the routine 80 to 90% of enquiries and the people coming home are kept for the moments that set the business apart.
- Booking Holdings runs one AI from search to service because it's one conversation. Stop buying two bots.
Booking Holdings put AI across the whole trip, from choosing a hotel to cancelling a booking, because the traveller asks both in the same chat. An operator who buys a sales bot and a support bot separately has split a conversation the customer never split.
- Google's AI Mode now books the plumber. You get that booking only if your prices and availability are data.
Google has turned a home-services answer in AI Mode into a Book Online button, fed by Jobber, Housecall Pro and Angi. The business that gets the job is the one whose services, prices and open slots a machine can read, so the work for an operator is making that data precise now, not waiting to see if agents catch on.
- Only 4% of service teams got their AI payback. The rest bought an answer machine and measured deflection.
Gartner says only 4% of customer service teams have the AI return they expected. Most of the rest installed a chatbot that answers questions and counted tickets avoided. The return only shows up when the AI finishes the job.
- Put AI on the checking before the writing
Checking thousands of product pages for wrong prices, dead images and missing details needs no judgement, and every error is a page you pay ads to send people to. That makes checking the first job to hand to AI, ahead of the writing.
- Banks asking who is liable for AI shopping is the second gate closing. Plan for the split, not the agent.
Amazon shut the first gate on shopping agents. Now the banks want to know who pays when an agent buys the wrong thing. Agentic commerce is still 50/50, and the money side is deciding it. A smaller business should plan for a split, not bet on agents paying at checkout any time soon.
- Google won't sell Search without AI, even to schools. There is no AI-free Google left to rank in.
Google will switch off Search, Gemini and Gemini in Classroom together for K-12 student accounts. If the AI can't be separated from Search even for schools, a page-one ranking with no place in the answer is a result nobody sees.
- Google won't tell you whether AI names your brand. Count it yourself, and date every result.
Google has no AI click data to add to Search Console yet, with a UK deadline weeks away. Even when the column arrives it will count clicks, not whether the answer named you. Run your own checks and date them.
- An AI content factory loses the human touch at the brief, long before anyone reviews the output
Search Engine Journal says you can scale AI content across markets without losing the human touch. I agree, with one correction. The human touch goes in at the front, as approved facts, local evidence and an opinion. A factory without that frame makes slop no matter how many people check the output.
- Zalando's AI shopping agent only helps its margins if it cuts returns
Zalando has launched an AI shopping agent while its margins are under pressure. An agent that recommends products touches no cost line. One that gets size and fit right, and changes an order before it ships, does.
- Claude's new anti-cloning rules won't protect your software. Your customers will.
A model maker can write rules against copying, but anyone with a prompt and a screenshot can rebuild your product. The part of a business a clone can't generate is its customers, its distribution and the relationships behind them, so that is where the defence has to be built.
- ChatGPT ads brought most advertisers no sales. You can buy the slot beside the answer, not the name inside it.
A study of 102 US advertisers found ChatGPT ad visits earned a fraction of what paid search visits earned, and most advertisers made no sales from them at all. The shopper asked for a recommendation and got one. The work is to be the brand that answer names.
- AI doom and AI hype are the same sales pitch. Judge the tool by a normal Tuesday in your business.
The founders warning that AI could escape human control are the same founders selling it as the biggest thing ever built. An operator should treat both stories as marketing, judge AI by what it does in the business on an ordinary day, and set the rules on what it can commit to themselves.
- A florist gets one in five orders from ChatGPT. The engine didn't find it. Its city pages did.
Persa Flores didn't win a fifth of its orders from ChatGPT by being big or by doing anything clever with the model. It had one plain page per city, each answering one delivery question. Your job is to count how many of your buyers' questions have a page of their own.
- Google says there is no AI SEO shortcut. Correct. The groundwork is the same, the finish line moved.
John Mueller is right that short-term AI SEO tactics switch off and the technical base still has to be there. But keeping the plumbing sound only gets you into the race. What you should measure is whether the AI names your business, and that comes from being the brand it can describe most precisely.
- Three AIs shopped one catalogue and picked three different winners. That spread is your to-do list.
A Search Engine Journal test ran three AI models over the same 20 Shopify products and got three different favourites. I read the spread as listings with nothing specific enough to win, and something a store can check for itself every week.
- Google search lost its free pass because the click is gone. Spend on being named in the answer.
Buyers are finally questioning the Google search line item. The reason is the click it used to buy. The question for an operator is what outcome the money should buy now.
- AI won't inherit your retiring expert's know-how. It only keeps what they put into words.
The Business Times says AI may be manufacturing's best chance to keep the black art of its retiring veterans. The need is real. But a model doesn't learn by watching. It keeps what the expert says out loud, so the hard job is the briefing, not the model.
- A reader cloned a tool with Claude. If you sell simple utilities, that reader is now your competitor.
One Hacker News reader had Claude build a site covering 16,500 books by steering it, not coding it. Personal tools are now cheap to make yourself, which is bad news for anyone selling a simple utility, and leaves the B2B gap exactly where it was.
- AI that summarises your meetings does not raise your ceiling. AI that runs the work does.
An e27 columnist found that months of AI-prepared meeting lists left their scale ceiling where it was. The lists only automate the preparation. The ceiling moves when the workflow itself runs end to end inside rules you set, and people handle the exceptions.
- Meta's Muse hasn't upended anything yet. Ask what need it meets before you pivot for it.
Wired says Meta's Muse could upend all kinds of businesses, and admits it hasn't. Before you plan around a new personal agent, ask what a customer can do with it on a normal day that they couldn't already do. Until that answer exists, it is one more tool in the pile.
- Nobody is ordering dinner through an AI agent yet. Build for the data, not the agent.
DoorDash says outside AI agents have brought it only hundreds of orders. That is the 50/50 I expected. The one move that pays either way is making your menu, prices and availability precise enough for any agent or AI answer to read.
- Big brands lead AI search by default. Smaller sites get in by being more precise.
BrightEdge's travel study shows the giants winning most AI answers and losing the ones that need specifics. The AI names the brand it can describe most exactly, and a small site can out-describe a giant on purpose.
- Shopify graded its people on AI use and learned nothing. Grade the work, not the tool.
Shopify asked staff to rate each other on AI use, and its CEO now says the answers didn't matter. Counting who touched the tool measures the toolset. Measure whether the work got better and what judgement it still needed.
- Jos. A. Bank is spending its people on fittings, not on 'where's my order'. Every retailer should make that split.
Jos. A. Bank built its new store around consultations and fittings. That is the right place for people. The transactional traffic belongs to AI, so the people are free to do the work that changes a sale.
- Apple won't ship AI for AI's sake. Neither should your business.
Apple's new CEO says the company will only ship AI where it makes a product people love. That is the same test every business owner should run before adding an AI tool: what problem does it solve that I couldn't already solve, and is it worth adding?
- The new AI training controls are a two-minute job, not a strategy
OpenAI will honour a shared robots.txt signal for AI training preferences. Set it once, let it run, and don't confuse controlling training with being named in an AI answer. That second job still needs the real work.
- Starbucks got the order right: warm stores first, AI at the drive-thru second
Starbucks fixed service times and made its stores warmer before moving AI onto the app and the drive-thru. That is the sequence any operator should copy: fix the transactional layer, put people where the warmth counts, then let AI run the repeatable part.
- An AI just booked dinner for a group over text. Your restaurant was either in its data or it wasn't.
Zest's text agent picks a restaurant for a group and books the table without anyone browsing or calling. If your availability lives on a phone line or a PDF menu, the agent doing the choosing never sees you.
- The big AI labs just gave you an opt-out switch. If you sell anything, leave it off.
OpenAI, Google, Microsoft and Anthropic are backing one shared signal for opting out of AI scraping. A business that wants to be named in AI answers should stay open, treat the file as a chore to automate, and not mistake a training setting for a visibility strategy.
- Google's AI will now answer your Business Profile messages. Answering was never the hard part.
Google is putting AI on Business Profile messages. It helps with the easy questions, but it answers from your profile and hands the rest back to you. Treat it as a front door, and make sure the messages that need action reach a system that can act.
- Reddit just closed the door. If your GEO plan was 'be on Reddit', it has nothing left.
Reddit is moving behind a login as search referrals fall. Building AI visibility on one platform was always a bet on someone else's business model. The brands that keep getting named state what they are on their own pages, precisely enough to be quoted.
- Amazon's 400 million Alexa shoppers are a data story, and a warning for smaller shops
Amazon says Alexa for Shopping reached 400 million customers who spend 50% more. I read that as a selection effect and a reorder button, not proof that voice is where commerce is going. The part that matters for a smaller business is who Amazon lets in.
- The AI decides how to help. You decide what it can promise.
What building AeroChat taught me about AI customer service: let the AI resolve the conversation, and keep the commitments (money, dates, promises) in rules you control. That split is what lets you trust it with more.
- The AI Search Playbook, 2nd edition
Your buyers are asking ChatGPT and Google who to buy from. Six months of testing what gets a brand named in AI answers, rebuilt into a six-step method.
- The tool was never the skill
Prompt tricks and button locations have a shelf life measured in months. Ten mental models for working with AI that still hold when the tools change.
- Getting a new store cited in AI search in ten days
I built a one-product Shopify store with no backlinks, no ads and no reviews, and got Google's AI Overview and ChatGPT to name it first. What happened, why it worked, and where it came short.
Subscribe by RSS