This piece pulls together three threads that ran through Shopify conversations on X in the second half of September 2026: Tobi Lütke's "slop grenades" warning, the arrival of agentic checkout with Muse, and the argument over whether AI search killed SEO. It answers one question for a Shopify merchant or app builder: what do these three stories have in common, and what should you actually change this week?

The month Shopify put AI on every surface at once

Within about a week, Shopify's leadership shipped or announced four things that each got their own news cycle.

The redesigned admin arrived with Sidekick, Shopify's AI assistant, placed front and center. The announcement from @katarinabatina (https://x.com/katarinabatina/status/2103161312901816394) drew 1,451 likes and more than 238,000 views. Shopify designer @MSchwaibold (https://x.com/MSchwaibold/status/2103227033891578047) described the work as two jobs at once: stripping layers and shadows out of the old interface while building AI-first interaction patterns through Sidekick.

Agentic checkout came next. @tobi (https://x.com/tobi/status/2102090718546198790) announced a deep partnership with Muse, Meta's AI assistant, to enable checkout with Shop Pay on all Shopify stores. Meta's @finkd (https://x.com/finkd/status/2102154890344849641) confirmed it and promised more partnerships. Muse's @alexandr_wang (https://x.com/alexandr_wang/status/2102092011021217911) framed the goal as giving Muse users access to a wide range of stores to find the perfect product.

Around the same time, Shopify President @harleyf (https://x.com/harleyf/status/2100300851705450965) announced ChatGPT Ads for Shopify merchants, with OpenAI pulling products straight from the Shopify Catalog. And @bradkowalk (https://x.com/bradkowalk/status/2099897698006737261) highlighted AI Autocomplete in store search, which maps a typed sentence to catalog attributes.

Read separately, these are an admin refresh, a checkout partnership, an ads product and a search feature. Read together, they are one move. Shopify is putting merchants' catalogs inside every AI surface a shopper might use, and letting the AI do the work that used to happen on the storefront.

The market noticed. @joecarlsonshow (https://x.com/joecarlsonshow/status/2102112430428197106) argued that if Amazon blocks agents, Meta will simply route shoppers to Walmart and Shopify, and "Amazon does not win in that situation." That post pulled 1,144 likes. @StockSavvyShay (https://x.com/StockSavvyShay/status/2101312580358930720) cited Oppenheimer's forecast of $28B in AI agent revenue for Muse by 2027.

The CEO's own warning about AI output

Here is the part that makes this month more interesting than a product roundup. While Shopify was pushing AI onto every surface, its CEO was publicly warning his own staff about AI output.

In a conversation on The Knowledge Project podcast, shared by @shaneparrish (https://x.com/shaneparrish/status/2099823553726070973), Tobi Lütke described employees who send unreviewed AI-generated emails and code to colleagues. He called that output "slop grenades." Fortune reported the remark on September 17, and a summary from @interesting_aIl (https://x.com/interesting_aIl/status/2102007936809841117) pulled 713 likes and more than 76,000 views.

The sharpest reading came from @tmiyatake1 (https://x.com/tmiyatake1/status/2101824478212669521), who noted that the definition of lazy work has flipped. It used to mean producing too little. In the AI era it means producing too much and making other people verify it. That post drew 380 likes and 50 reshares.

The phrase is sticking. @rauchg (https://x.com/rauchg/status/2102436383923339629) listed "slop grenades" on the agenda for his on-stage interview with Tobi at Vercel Ship on October 15, alongside agentic commerce and native rewrites.

Two things are worth holding together here. First, Shopify is not anti-AI. @mjackson (https://x.com/mjackson/status/2098073504931618946) noted that Shopify is migrating all its React Native code to Swift and Kotlin because AI made that lower risk than ever, a post that reached more than 51,000 views. Second, the same company insists that the person who generates AI output remains responsible for checking it. Those are not contradictory positions. They are the same position.

Former Shopify engineer @Sirupsen (https://x.com/Sirupsen/status/2098461778866798840), now a founder, put the leadership version plainly: "long-term it is much worse to ship things that are sub-par. The job of a leader is to pull pain forward." That post drew 1,042 likes.

For a merchant, the translation is direct. Every AI feature you turn on, whether Sidekick, a page builder's AI copy tool or a product description generator, produces output that someone must check before it reaches a customer. If the app does not make that checking cheap, the grenade lands on you.

What "reviewable AI" looks like, and who already does it

Sidekick itself is the best example of the pattern Tobi is asking for. @sh_sakamoto (https://x.com/sh_sakamoto/status/2101446357173289307) walked through a product page SEO review: Sidekick proposes changes to the title, meta description, image alt text and URL handle, marks each change, and waits for the merchant to approve before anything is applied.

Three things happen in that flow. The merchant sees the change. The merchant decides. Nothing goes live until they do. App developer @junaidkbr (https://x.com/junaidkbr/status/2102778882617463095) asked whether merchants will now expect Sidekick to work with every app. The more likely outcome is that merchants will expect every AI feature, from any app, to behave the way Sidekick behaves.

The case for that standard is not that AI is bad. It is that AI is wrong often enough to plan for. @TRPage_dev (https://x.com/TRPage_dev/status/2099492186857275745) asked Sidekick to recommend an app for a gift card migration and found that two of the three suggestions did not meet the ask. That ratio is harmless when the output is a suggestion the merchant reads. It is a disaster when the output is a change the merchant never saw.

Several practitioners described what good looks like in practice:

  • @jamiegrove (https://x.com/jamiegrove/status/2102810220648976526) had AI proofread 682 product pages. The first run burned about $185 and never finished. The second run, same prompt and model, cost $12.60. "The model wasn't the big lever. The harness was."
  • @jay_neyer (https://x.com/jay_neyer/status/2102178562375610443) migrated more than 1,000 products in under eight hours, with AI running QA against the original catalog and flagging mismatches for a human. "The hours that disappeared were the repetitive ones, and the checking stayed."
  • @chesny (https://x.com/chesny/status/2102778825591689255) described an agent that ran a dropshipping store for eight weeks while the human kept two jobs: approving spend and deciding what looked good. Nothing went live until the agent's own checks passed and the human signed off.
  • @ridark_eth (https://x.com/ridark_eth/status/2098891678480310680) shared a solo operator's prompt set in which any prompt lacking evidence must return UNKNOWN rather than guess. That post drew 116 likes.

There is also a lesson about restraint. @junaidkbr (https://x.com/junaidkbr/status/2103215196387119176) retired a sidebar toggle from their app Alfred because Shopify's new admin shipped a better one, and sounded happy about it. If the platform already does something well with a mark-and-approve flow, an app offering a worse version adds review burden rather than value.

A short test for any AI feature you are evaluating:

QuestionPass looks like
Can I see the diff?Old and new side by side, or changes marked
Is approval real?Accept some changes and reject others, not "Apply 40 changes?"
Can I undo in one action?Not "restore from a theme backup"
Does it admit uncertainty?Flags "could not find material for this SKU" instead of inventing one
How much per request?Small batches with a fast approve loop

When the buyer is an agent, the catalog is the storefront

The second thread connects to the first through one uncomfortable fact: every one of Shopify's new AI surfaces reads the Shopify Catalog, not your storefront. ChatGPT Ads pulls from the catalog. AI Autocomplete maps queries to catalog attributes. Muse searches the catalog and checks out with Shop Pay.

@yannis1kiefer (https://x.com/yannis1kiefer/status/2102654725883658684) described pulling their catalog the morning a ChatGPT Ads window opened: "half the titles still say best seller with no size or material. agent has nothing to match a real shopper question against." That is a slop grenade of a different kind. The lazy output was written years ago, and the person it lands on is a machine that will simply choose someone else.

@MIKS_ae (https://x.com/MIKS_ae/status/2099525865172283642) listed what an AI needs to understand before it recommends you: what you sell, who it is for, why it is better, what customers say, and whether the information around the product can be trusted. @WooCommerce (https://x.com/WooCommerce/status/2100559244424516083) compressed it to one question: agents are reading structured product data to decide what to recommend and buy. Can they read yours?

The counter-evidence deserves a fair hearing. @Yellow (https://x.com/Yellow/status/2097601654917062913) pointed out that the first summit dedicated to agentic commerce is about trust, identity, fraud and liability, not discovery. Juniper projects $1.5T in agentic spend by 2030 while reported current settlement is around $28,000 a day. "Nobody has solved who is accountable when an agent transacts." @sytaylor (https://x.com/sytaylor/status/2097651235272978894) cited the number that should anchor expectations: only 9% of consumers say they would let an agent complete a purchase for them.

Retail incentives also cut against the story. @TrungTPhan (https://x.com/TrungTPhan/status/2102196110848933957) noted that Amazon has blocked ChatGPT and Muse from crawling its platform. @eric_seufert (https://x.com/eric_seufert/status/2102066171658801416) argued that Amazon and Walmart have precise financial reasons to resist handing the customer relationship to third-party agents, and that a new interaction paradigm does not override those incentives by default.

Then there is the consolidation risk. @AiEvolutio58513 (https://x.com/AiEvolutio58513/status/2101332773525143678) relayed Cloudflare CEO Matthew Prince's warning that agents may favor established companies with more online data, because new brands have fewer reviews, transactions and verifiable signals for an agent to evaluate.

So the honest read is this. The rails are ready, the distribution is enormous, the consumer habit barely exists, and the biggest retailers are fighting it. That combination usually means early movers take a disproportionate share of a small but fast-growing channel. It does not mean your Meta ads stop working next quarter. It does mean that a catalog full of "best seller" titles is now a liability on four surfaces instead of one.

AI did not kill SEO. It killed the copied page

The third thread is the one most DTC brands feel first: organic traffic sliding while everyone sells them an "AEO" package. Answer Engine Optimization, the practice of getting AI assistants to cite and recommend you, is real. The hacks sold under that name mostly are not.

@semrush (https://x.com/semrush/status/2103416758598005136) put the frame in one line on September 25: "AI didn't kill SEO. It exposed every page that copied the SERP and added nothing new for readers."

The data behind that line is less dramatic than the headlines. @semrush (https://x.com/semrush/status/2103124321250033725) summarised its State of Search Q2 2026 report across 10 million desktop users: traditional search is still going strong as AI use grows, ChatGPT leads but Gemini and Claude are catching up, and people are typing longer, more detailed questions. @green_but_red (https://x.com/green_but_red/status/2103054590216401126) added a number that should reset expectations: 83% of ChatGPT carousel product sources are pulled from Google Shopping, according to research they cited. If your Google Shopping feed is weak, your ChatGPT presence is weak too.

@SkomorNick (https://x.com/SkomorNick/status/2101204112444383504) explained why the AI layer sits on top of the old work rather than replacing it: "A model cannot cite a page it never retrieved." They cited the C-SEO Bench study finding that getting retrieved matters roughly seven times more than any rewording of the paragraph. Their warning for anyone who cut SEO budget to fund AEO: "this is the part that stings."

The strongest practitioner evidence came from @SEOKeval (https://x.com/SEOKeval/status/2101312354029879604), who works with an ecommerce founder whose AI-referred sales are now outperforming traditional Google search. The method: "I haven't even done any AEO/GEO-specific strategies. No listicles. No PR. I just optimized their product pages. Published topically relevant blog content. And built a ton of backlinks." The product pages are now being cited.

What the hacks look like, and why to avoid them:

  • @jakezward (https://x.com/jakezward/status/2099482586560995756) listed the usual pitches, "just add an llms.txt file," "add some FAQ schema," "get mentioned on Reddit," against the roughly thirty things winning AI search actually involves.
  • @green_but_red noted that self-promotional listicles help short-term but are already targeted for demotion by Google, and that hidden prompt injections in "Summarize with AI" buttons are now classified as security threats by Microsoft.
  • @neilpatel (https://x.com/neilpatel/status/2103182749083312193) used three sneaker searches to show SEO, AEO and GEO producing different result types from the same playbook, not three separate jobs.

There is a real consumer behaviour shift underneath this, and @alexgroberman documented it across three posts. Citing a Semrush Exploding Topics survey of 2,338 US consumers, they reported that 57.5% of AI users said a chatbot had caused them to decide not to buy something (https://x.com/alexgroberman/status/2103119727468179521). "You can win a customer on Google and still lose them when they ask ChatGPT whether they should buy." A second post (https://x.com/alexgroberman/status/2103476987780341867) pointed to Surfer's analysis of 289,105 cited URLs across ChatGPT, Google AI Mode and Perplexity. A third (https://x.com/alexgroberman/status/2103845216415605080) argued most visibility work can be done without hiring anyone: strong product pages, useful content, backlinks, brand mentions and clear site structure.

For the page itself, @chris_nectiv (https://x.com/chris_nectiv/status/2102025492840739175) offered the most shared template, at 335 likes: optimized title and H1, key takeaways, structured content formats, freshness updates and FAQs.

Note that the numbers above are relayed through posts. Before repeating 57.5%, 83% or the seven-times retrieval figure in your own materials, open the original Semrush, Surfer and C-SEO Bench reports.

One conclusion, and five things to do this week

Put the three threads side by side and they say the same thing.

Tobi's slop grenade is AI output that pushes the checking onto someone else. A catalog full of vague titles is the same failure aimed at a shopping agent. A product page that copied a competitor's spec sheet is the same failure aimed at a search model. In every case, the fix is not more AI and not less AI. It is making the facts on your pages true, specific and readable, and keeping a human in the loop where the machine is likely to be wrong.

That leads to a short list:

  1. Run the Sidekick test on every AI feature you use. Diff visible, approval real, undo in one click, uncertainty admitted. If an app fails three of five, it is costing you more in cleanup than it saves in generation.
  2. Audit your top five product titles the way @yannis1kiefer did. Read each one aloud with no image. "Everyday Tee, Black, Organic Cotton, Unisex" passes. "The Classic" and "Best Seller" do not.
  3. Answer the five questions from @MIKS_ae on the page, in text. What it is, who it is for, why it is better, what customers say, and why the information can be trusted. Review counts and certifications inside images are invisible to every machine reading you.
  4. Do not cut SEO to fund AEO. Follow @SEOKeval's path: product pages, topical content, backlinks. Adopt @chris_nectiv's structure for the pages that matter. Skip the llms.txt and self-ranked listicle pitches.
  5. Remove every step between "yes" and paid. The Muse flow ends in Shop Pay. If your own checkout is still three pages with Shop Pay off, the agent path to your store is faster than your own. @alexpagepilot (https://x.com/alexpagepilot/status/2101714252776280457) described the two-click version; treat the lift claim as a hypothesis to test, but do the two clicks.

None of this requires rebuilding your store for agents. It requires the same discipline Shopify is asking of its own staff: generate as much as you like, but check it before it lands on someone else. The someone else is now your customer, a shopping agent and a search model at the same time.

Sources