By Vladislav Mykhailenko, Founder of SellerUp

AI × E-commerce · Issue #4 · Wednesday, August 12, 2026

AI news #4:
A court gave AI agents Amazon back, and US model prices fell for the first time

Issue #4, July 29 – August 12. An appeals court lifted the ban on Perplexity's agentic browser, Sellerboard shipped an official MCP, and OpenAI and Anthropic are holding and cutting prices in answer to the Chinese price war. One deadline for the calendar: on August 17 OpenAI switches on advanced matching on ChatGPT Ads pixels itself.

Court
An appeals court lifted the ban — Perplexity's AI agent browses and buys on Amazon again.
Prices
OpenAI cut API prices fivefold, Anthropic cancelled the Sonnet increase — an answer to the Chinese models.
MCP
Sellerboard shipped an official MCP — an assistant now reads a seller's net profit directly.
Block 1

The big AI players

AmazonAugust 4

Ban lifted: Perplexity's AI agent browses and buys on Amazon again

The Ninth Circuit Court of Appeals reversed the injunction that had kept Comet off Amazon's storefront since March — Comet being Perplexity's agentic browser, where the AI walks the sites and clicks the buttons for the user. The court's logic is simple: it is the shopper who arrives at the site, and the AI is just a tool in their hands.

  • Case No. 26-1444 (No. 3:25-cv-09514-MMC below, Northern District of California). The decision came down on August 4, 2026, in an opinion by Judge Milan D. Smith Jr.
  • Amazon built its case on two statutes: the CFAA — the US law on unauthorised access to someone else's computers — and California's CDAFA, which covers the same ground. The court held Amazon was unlikely to win on either: the user is the one visiting the site, and the AI assistant is only a tool in their hands.
  • The remaining factors — likely harm, the balance of the parties' interests, the public interest — the panel also counted for Perplexity. The injunction is vacated and the case goes back to the district court.
  • Amazon filed in November 2025 and pleaded only two grounds, CFAA and CDAFA. The district court granted the injunction on March 9, 2026. The appeals court noted there is almost no case law on who answers for an AI agent's actions, and deliberately kept its ruling narrow.
What this means for a seller: revisit the first bullet on your top products: does it answer the shopper's main question in words — what it is made of, who it suits, what is in the box. The agent reads text, not infographics. You do not need to rewrite every bullet and description over this — the first one is enough.
Source: U.S. Court of Appeals for the Ninth Circuit, decision No. 26-1444 (opens in a new tab)
Live The first page of the Ninth Circuit's August 4, 2026 decision: Amazon.com Services v. Perplexity AI, case No. 26-1444, with the district-court number 3:25-cv-09514-MMC below it.
◆ Issue partner · SellerUp

Need turnkey AI automation, or training for your team — that is what we do

SellerUp does AI automation for e-commerce: Amazon, Allegro and other marketplaces. If you need turnkey AI automation, personalised training for your team on rolling out AI automation, or a consultation — get in touch.

  • Coaching and training. We train you and your team on rolling out AI automation, personally — so that you build and grow it yourselves.
  • Turnkey automation. We set up AI processes for e-commerce on Amazon, Allegro and other platforms — from analytics to routine operations.
  • Consulting. We go through your project and show where AI pays off today, and where it is still too early.

Every team that has been through the training builds automations freely, pulls data from three to five sources — Keepa, Sellerboard, Helium 10 — tests new hypotheses every day, and gets analytical insight that used to be out of reach.

Subscribe on Telegram (opens in a new tab) Message us directly at @retail_and_ecommerce — we will go through your AI rollout with you.
E-commercethis week

Sellerboard shipped an official MCP — an assistant now reads your net profit directly

Sellerboard is the profit-analytics service for Amazon sellers: it holds cost of goods, inbound logistics and net profit per SKU — exactly the data no Amazon report has. It now has its own MCP server, and an assistant pulls those numbers itself, with no exports and no copying spreadsheets.

  • The server address is mcp.sellerboard.com/mcp. We checked it live: the endpoint answers and accepts MCP client connections.
  • MCP (Model Context Protocol) is the open standard AI applications use to connect to external services. To add it in Claude: Settings → Connectors → Add custom connector → paste the address and authorise your Sellerboard account.
  • This closes the main hole in any analytics built on the Amazon API: the platform reports sales and fees to the cent, but has no idea what you paid for the goods. Profit could not be computed without a manual export — now the assistant takes it from Sellerboard itself.
  • The MCP wave across seller tools has reached profit analytics: Amazon started back in February, Helium 10 joined in July (MCP there is Diamond-only, with a quota of 1,000 calls a month), then Sellerise — and now Sellerboard.
  • There is no public documentation for the server yet — the announcement reached users before the docs did. Check the tool list and the limits inside your account after connecting.
What this means for a seller: add the connector and ask the assistant a control question: net profit on your three top SKUs over 30 days. If the numbers match the Sellerboard dashboard, move your weekly report onto this pairing: prompts about profit, margin and restocking start working without a single manual export.
Source: Sellerboard, official MCP endpoint (verified live) (opens in a new tab)
Live The Sellerboard dashboard: net profit, margin and ROI per SKU — data that is in no Amazon report and that an assistant can now read directly
AI giantsJuly 30 and August 11

US models are getting cheaper: OpenAI cut API prices fivefold, Anthropic cancelled the Sonnet increase

For the first time in a year, prices on US models moved down rather than up — and not one at a time, but across the board. The reason is simple: users are leaving for Chinese models that are free or nearly free, and the vendors had to answer with money.

  • On July 30 OpenAI cut the price of a call to its small model by 80% and to the mid-tier by 20%; the flagship kept its price. Bulk work — translating listings, answering reviews, rewriting bullets across hundreds of SKUs — got five times cheaper the same day.
  • On August 11 Anthropic made the launch price of Sonnet 5 permanent: $2 per million input and $10 per million output tokens. The increase to $3/$15 planned for September 1 is cancelled.
  • Anthropic holds the same line at the top end: on benchmarks Opus 5 comes right up to the flagship Fable 5 while costing half as much — $5/$25 against $10/$50 (we covered it in issue #3).
  • The pressure comes from China: DeepSeek cut prices twice over the summer — its latest model serves a million output tokens for under a dollar — and in July six of the world's ten most-used models were Chinese (Fortune's data).
  • Users leaving is no longer a hypothesis: Coinbase said publicly it halved its AI spend by moving staff onto China's Kimi and GLM, and Microsoft is looking at DeepSeek as a cheap replacement for part of the model stack behind Copilot.
What this means for a seller: recalculate your monthly AI bill at the new prices: move bulk work onto the cheaper small models and keep the expensive flagship only where meaning genuinely has to be parsed. Prices are moving down right now — locking yourself to one vendor has stopped paying.
Source: OpenAI (July 30) and Anthropic (August 11), official announcements (opens in a new tab)
Live The price-cut announcement on OpenAI's blog: performance per dollar as the main argument against leaving for the cheap models
AI giantsAugust 7

On August 17 OpenAI switches advanced matching on across every ChatGPT Ads pixel itself

Before: advanced matching only ran on new pixels. Now: on August 17 it switches on for every pixel already installed, and nobody will be asked. If you do not want it, you have to turn it off by hand, before that date. Now, in plain words, what this actually is.

  • A pixel is a piece of an ad platform's code on your site. It tells OpenAI that a person who came from an ad in ChatGPT got somewhere: bought, left an enquiry. Without a pixel the platform has no idea whether the advertising pays at all.
  • Why matching is needed: part of your sales gets “lost” — browsers cut cookies, the shopper clicks on a phone and buys that evening on a laptop. The purchase happened, but it never attached to the ad, and the campaign looks worse in the report than it works in reality.
  • Advanced matching fixes that loss: the pixel takes the email or phone the shopper entered into a form on your site, hashes it — turns it into an irreversible string — and sends it to OpenAI. OpenAI matches the hash against ChatGPT accounts and glues the lost sales back onto the ad impressions.
  • The trade-off: reports get more accurate and the algorithm learns better — but hashed data on your customers starts flowing to OpenAI. That is exactly why switching it on unasked caused a stir.
  • To turn it off by hand: Tools → Conversions → Data Source → Edit pixel. To stop matching switching itself on, do it before August 17.
  • In the same update: a beta conversion-optimized CPC for product-feed campaigns (you pay per click, but the auction factors in predicted conversion), link macros, Triple Whale and Hightouch integrations, and ad launches in Brazil and Mexico.
What this means for a seller: this only concerns anyone running ChatGPT Ads to their own site — it does not touch selling inside Amazon. If you have the pixel installed, decide before August 17: keep matching for accurate attribution, or switch it off so customer data does not flow to OpenAI.
Source: OpenAI, ChatGPT Ads product email (text published by PPC Land) (opens in a new tab)
Live A frame from OpenAI's announcement page: an ad card marked Sponsored right inside a ChatGPT answer. From August 17 the pixel starts gluing conversions on exactly this format back together from form data.
AI giantsAugust 10

AI agents are moving into ad consoles everywhere: Ask Advisor at Google, its own assistant at Allegro

On August 10 Google rolled an agent layer straight onto the main screens of Google Ads and Google Analytics. Ask Advisor is a built-in agent on Gemini: you ask in your own words and get an analysis and a list of actions, not a summary. Beta for now, and English-language accounts only.

  • Ask Advisor works across Google's advertising and analytics services: a question in ordinary words instead of an export into a spreadsheet.
  • The Google Ads home screen now has insight cards and an input field: you can ask how competitors affected your campaigns, or what shifted in demand.
  • A dashboard is assembled from a text request: the system builds the charts and labels what moved the result itself.
  • In Google Analytics the home screen carries an AI Overviews block: what changed since your last visit, one-click hand-off into Ask Advisor, notifications by email and phone.
  • Ask Advisor opens up benchmarking — your campaigns against anonymised data from similar businesses. Ordinary reports do not have this.
  • This is not one platform experimenting — AI summaries are appearing across every ad system at once. Allegro already runs its own AI assistant in the Sales Center: per the platform's official announcement, it has real-time access to account data — account health, ad campaign results, complaints.
  • In the Allegro ad console the AI gives a summary of what happened to your campaigns: how far CTR (click-through rate — the share of people who clicked an ad after seeing it) rose or fell, and other metrics.
  • Our own observation from our Allegro Ads console: a summary like that happily highlights the good and rarely mentions the bad. Its job is to show the positive — read it as platform marketing, not as an audit.
What this means for a seller: open your console and put your first Ask Advisor question against your worst campaign, and pull the benchmark against similar businesses while you are there — it is the one place Google shows whether you are behind the market or the market fell with you. And the general rule for AI summaries in any console: they will show you the positive by themselves, so go looking for the bad in the numbers.
Source: Google (Ads & Commerce blog) and Allegro (media.allegro.pl press release) (opens in a new tab)
Live Google's illustration for the announcement: the Ask Advisor panel answers “how do I look against similar businesses” and returns a table of metrics with an Outperforming or Underperforming status.
E-commerceJuly 29

Facebook advertising started understanding what ads mean — conversions rose 15.7%

Meta put a language model into ad selection. A language model is a program of the same kind that runs inside ChatGPT and Claude: it understands what text and pictures mean rather than just comparing numbers. Nothing is required of the advertiser — it is already running.

  • How it was: Facebook's system decided who to show an ad to from click statistics — whoever clicked something similar before gets it. What the ad actually said, the system did not really understand.
  • How it is now: the model reads the ad the way a person does — what the product is, who it is for, what makes it different — and matches that against what it knows about a particular user's interests. The decision to serve is made on the meaning of the ad, not only on past clicks.
  • Meta named the result itself in its second-quarter report: conversions on Facebook rose 15.7%. A conversion is when a person does not just look at the ad but does the thing it ran for: buys, or leaves an enquiry.
  • Nothing to switch on: the rebuild happened on Meta's side for every advertiser at once.
What this means for a seller: now that the system reads the ad rather than only counting clicks, rebuild your creative set so each one explains the product from a different angle: use case, materials, comparison, review. Five recoloured versions of one image are, to a “reading” system, one ad and not five.
Source: Meta, second-quarter 2026 report and investor call (opens in a new tab)
Live A diagram from Meta's engineering blog: an offline model parses a person's behaviour, an online model ranks and predicts a Facebook click, an Instagram click and a conversion.
AI giantsAugust 3

Alibaba shipped Qwen3.8-Max: a million tokens of context and $2 per million input tokens

The Qwen team put out its largest model. For a seller this is not an abstract release: Qwen is a model people already genuinely buy through. Since May 2026 Alibaba has opened the whole Taobao and Tmall catalogue to it, and agentic shopping there runs in production, not in a demo.

  • 2.4 trillion parameters in an MoE architecture, 95 billion active per token, 1M tokens of context.
  • Price in QwenCloud is $2 per million input tokens and $6 per million output, cached input $0.25. The API is compatible with the OpenAI and Anthropic formats.
  • Confirmed use in e-commerce: since May 11, 2026 the Qwen app works as an agentic storefront for Taobao and Tmall — search across a 4-billion-product catalogue, comparison by seller, virtual try-on, 30-day price watching and order placement (TechNode, The Next Web).
  • Qwen has roughly 300 million monthly active users across the Taobao, Tmall and Alipay surfaces; over the New Year campaign the platform counted about 140 million first-ever AI purchases.
  • Qwen is also built into AliExpress for cross-border trade — seller listings there already pass through this model.
  • And the meme of the week for beginners: if you paid for a course on “how to find a supplier in China” to get your first test batch — you can throw it out. Alibaba's AI agent finds and filters suppliers against any criteria you set and negotiates around the clock (that is Accio, which we covered in issue #3), and Qwen is a model from the same ecosystem. There are no statistics yet on how well this works on real purchasing — but the trend is clear: the barrier to sourcing from China is falling.
What this means for a seller: if you source from China or sell through AliExpress, a Qwen agent already reads your products — look at how your listing appears through its eyes. And check the API economics on your own real task and your own data before moving anything: there are no confirmed cases from Western sellers on Qwen yet.
Source: Qwen (Alibaba), official blog (opens in a new tab)
Live A chart from Alibaba's newsroom: in the Text Arena ranking Qwen3.8-Max sits fifth with 1496 points, between two Claude Opus models.
AI giantsAugust 10

Meta opened Muse Glimmer: a 30-billion-parameter agentic model that runs on a home graphics card

Meta released an open-weights model built specifically for agents. Open weights means the model itself can be downloaded as a file and run on your own computer: it works without internet, for free, and your data goes nowhere. With closed models — ChatGPT or Claude, say — access is only through the developer's server, you pay per request and you send data out.

  • Who this is for: anyone who wants to play with local models — and anyone with a business process they would rather not, or cannot, push through a cloud API: sensitive data, large volumes, daily routine.
  • Apache 2.0 licence — commercial use needs no separate permission and no payment to Meta.
  • A 30-billion-parameter model; after quantisation (compressing the weights) it takes under 20 GB and fits on a single 24 GB gaming card — that is, literally on a home computer.
  • Claimed: precise function calling, multi-step agentic scenarios with recovery after failure, text and image input, and more than 100 languages.
  • Speculative decoding — the trick where a small model guesses the continuation and a large one checks it — speeds generation up by 1.5–3.1× depending on the hardware.
What this means for a seller: stand it up locally through Ollama — a free program that runs the model right on your computer — and hand it the routine you would rather not push through a paid API: bulk review analysis, cleaning up exports, draft bullets.
Source: Meta AI Research, blog (opens in a new tab)
Live The Muse Glimmer announcement on the Meta AI Research blog
AI giantsAugust 7

xAI shipped Imagine Image 2.0 — with ready-made templates for product photos

xAI's new image generation and editing model did not arrive bare but with a set of presets, half of them built directly for a product listing.

  • The model is available as Quality Mode on grok.com/imagine and in the Grok apps for iOS and Android.
  • Templates built for sellers: product photography for a listing, changing the product colour, a magazine-style poster, a “customer-shot” photo, background removal and replacement, merch.
  • Multi-ref editing is when several source images go into one generation; the app accepts up to 5. Smart resize rebuilds one frame for 9 aspect ratios: 1:2, 9:16, 2:3, 3:4, 1:1, 4:3, 3:2, 16:9, 2:1.
  • There is no API access yet: the announcement says “API access is coming soon” — meaning you cannot wire it into your scripts, only work by hand.
What this means for a seller: run two or three of your main SKUs through the E-Commerce Photos template and cut one frame into square, vertical and banner instead of three separate shoots.
Source: xAI, official announcement (opens in a new tab)
Live The E-Commerce Photos template in Imagine Image 2.0: a product on a white background, assembled by the model
Block 2

Amazon and the platforms

AmazonJuly 30

Amazon's advertising AI agent reached 11 more countries — and cuts acquisition cost by 6%

In the same quarterly report Amazon disclosed that Ads Agent — the AI tool that builds and runs ad campaigns instead of a person — has been rolled out to 11 more countries. And put numbers on the effect for the first time.

  • Over this year Ads Agent was rolled out in 11 more countries — European and Asian accounts that did not have it will now see it in the console.
  • Advertisers working through Ads Agent see an 8% lower cost per impression and a 6% lower cost to acquire a customer. Those are Amazon's own numbers.
  • The tool takes on planning, launching and running a campaign: what takes hours of setup and target picking by hand is assembled in minutes.
  • Amazon presents this not as a beta but as part of an advertising business that grew 26% year over year in the quarter.
What this means for a seller: run Ads Agent on one small ASIN (an Amazon product identifier) alongside your own manual campaign and compare acquisition cost after two weeks — it is the only honest way to work out whether to give it budget.
Source: Amazon, second-quarter 2026 report (opens in a new tab)
Live An Amazon Ads scene with the line “Full-funnel advertising. At scale. For everyone.” — the illustration to the newsroom piece on the quarter's advertising results.
AmazonAugust 7

Item highlights work in search on equal footing with the title — Amazon confirmed it in writing

Before: the product title is what works for search. Now: Item name (the title) and Item highlights (the short points shown under the title) are equal inputs, with neither field outranking the other. The unified display switched on August 10.

  • Amazon wrote it in plain text in a FAQ on the Seller Central forum, published July 30, 2026: “Item name and Item highlights are both inputs for search, and neither field takes priority over the other”.
  • Item highlights are 125 characters on top of the 75-character title, the same 200 in total. From August 10, 2026 they are shown under the title identically on desktop and mobile.
  • The rule also applies to Amazon's own listings — a moderator confirmed that in the comments to the post. Brand owners get 14 days to review and correct AI changes in Review Listing Changes.
What this means for a seller: do not leave Item highlights empty and do not copy the title into it — fill those 125 characters with the keywords that did not fit into the 75.
Source: Amazon Seller News FAQ (EcommerceBytes analysis) (opens in a new tab)
Live A SoldScope breakdown: before August 10 Item highlights ran together with the title on one line, after it they sit in a separate block underneath. On the right you can see how the platform counts the fields: title 73 of 75 characters, highlights 105 of 125.
AmazonAugust 4

Amazon is translating the text on your product photos itself, in fourteen countries

Infographics, size charts, packaging copy — Amazon runs all of it through AI translation and shows it to the shopper in their own language. There is no opting out.

  • Amazon announced on August 4 in Seller Central that it is launching the feature in 14 stores: the US, Canada, Mexico, Brazil, the UK and nine European sites.
  • The translation is done by AI; Amazon says there is a quality check before publication.
  • In the US, Canada, Germany, Spain and Belgium the version of the image is picked to match the individual shopper's language settings.
  • There is no opting out, and the main image is left alone: the translated version is added rather than replacing the original, and it is visible in Image Manager — the listing images section in Seller Central.
  • The only way to remove the machine version is to upload your own localised image; it automatically displaces the generated one.
What this means for a seller: open Image Manager on your top ASINs in Europe and Mexico and check whether the AI translated your brand name, trademarks and the numbers in the size chart.
Source: Amazon Seller News (Seller Central forum) (opens in a new tab)
AmazonAugust 5

Seller Assistant now explains why you are losing the Buy Box

The Buy Box (Featured Offer in Amazon's documentation) is the main offer on a listing with the Add to Cart button: many sellers are on the listing, one gets the button. Amazon has built a per-ASIN breakdown into its assistant: it shows exactly where your offer loses to the winner.

  • It runs right inside Seller Central: open Seller Assistant and ask “Why am I not the Featured Offer on [your ASIN]?”.
  • The assistant walks four blocks: whether the offer is buyable at all (out of stock, listing suppressed, incomplete product data), the price gap against the current winner, delivery speed, and regional coverage.
  • You can check up to 10 ASINs in one conversation.
  • The feature arrived as Amazon drops its separate eligibility threshold for the Featured Offer: from July 6 the filter that screened out sellers below a metrics bar is being removed, from July 20 in the EU and the UK, with the global rollout running to the end of 2026. Every seller's offer now enters the ranking.
What this means for a seller: run the ten ASINs with your worst Buy Box share for July and write down what the assistant names as the cause — most often it is delivery speed or coverage, not price.
Source: Amazon Seller News (EcomCrew analysis with a screenshot of the dialogue) (opens in a new tab)
Live The new Seller Central interface: the Seller Assistant button in the top bar and live cards on shipments and claims — exactly the signals the assistant uses to explain a lost Featured Offer.
AmazonJuly 30

Platform assistants can check “Made in USA”, but go quiet when asked directly

The Center for Law and the Economy — a research centre co-founded by former FTC chair Lina Khan — published a report on how platform assistants handle country of origin. The conclusion is unpleasant for honest sellers: the assistant has the data, but will not hand it over when asked directly.

  • Asked to find products “made in the USA”, Alexa for Shopping replied that it had no access to that information.
  • Rephrasing it to “in the USA” was enough — the assistant produced a working list of American products.
  • Asked for “Made in China”, the lists came straight back, with no caveats.
  • Walmart's assistant first refused to assess the credibility of origin claims for items in the cart, and after a rephrasing ranked them from “high confidence” to “likely not U.S. made”, explaining its refusal like this: “If I've said in a past conversation that I can't do that, that was overly cautious”.
  • The researchers recorded listings that carried both a domestic and an imported origin marking at the same time.
What this means for a seller: state the country of origin in plain text — in the attribute, in a bullet and on the infographic — and remove the contradictions between listing fields: the assistant answers from what is written, and on conflicting data it simply will not name you.
Source: PPC Land (on the Center for Law and the Economy report) (opens in a new tab)
Live The Alexa for Shopping panel over amazon.com and in the mobile app — the assistant that dodged the Made in USA question in the test.
E-commerceAugust 5

Shopify: orders from AI search tripled, and three quarters of them fall outside the top-100 categories

The first big report from a Western platform where the AI channel is shown in numbers rather than promises. And the key figure there is not about money but about who gets that traffic.

  • For the quarter: platform GMV $115,567M (+32% year over year), revenue $3,583M (+34%), free cash flow $654M at an 18% margin.
  • AI traffic to storefronts and orders from AI search tripled year over year, and new customers arrive from AI channels almost twice as often as from the rest.
  • 75% of AI-attributed orders went to products outside the top-100 categories.
  • AI search across Shopify's structured catalogue converts twice as well as AI search over data a bot scraped off the page itself, and roughly 80% better than ordinary organic search. The catalogue holds more than 1 billion products.
  • Half of all AI sessions land straight on a product page — 2.5 times more often than with ordinary search.
What this means for a seller: give the platform a structured feed with every variant, size and material instead of hoping the agent parses your page itself, and look at your long tail — that is where the volume goes.
Source: Shopify, second-quarter 2026 report (opens in a new tab)
Live A screenshot of the Shopify admin: the Agentic storefronts block, where the shop's products are automatically visible to AI agents and the merchant decides with toggles where checkout happens — ChatGPT, Google AI Mode and Gemini, Microsoft Copilot.
E-commerceAugust 5

The Shopify admin has an Agentic section: you can see what people buy through ChatGPT, Gemini and Copilot

Along with the quarterly report Shopify showed a redesigned admin with a separate Agentic section. A merchant used to see agentic sales as anonymous “other traffic” — now it is its own screen with channels and metrics.

  • Inside the section you can see which AI channels the shop is currently live on: ChatGPT, Google AI Mode and Gemini, Microsoft Copilot, the Shop app — with real-time metrics for each.
  • There is a search over real shopper queries: you can look at where your products rank in the assistant's answer for a specific phrasing.
  • Sidekick — the AI helper inside the Shopify admin — suggests what to fix when the product is found but not bought.
  • The quarter's numbers: 34 million conversations with Sidekick, its daily merchant audience up 3.6× year over year, and merchants built 36,000 of their own apps through Sidekick against 12,000 in the first quarter.
  • What the section does not show: purchases through Claude and Perplexity. The named channels are only ChatGPT, Google AI Mode with Gemini and Microsoft Copilot; everything else falls into an anonymous “Other channels” line. Why Claude and Perplexity were not invited to the party, Shopify does not explain; the section cannot yet be read as the whole picture of agentic traffic.
What this means for a seller: open the Agentic section and run your three main phrasings through the query search: if the product is found in the answers but there are no orders, what needs fixing is the listing fields and the price, not the ad budget.
Source: Shopify, second-quarter 2026 investor presentation (opens in a new tab)
Live The Agentic section in the Shopify admin: earnings and the number of agentic sessions over 30 days, and the sales split by AI channel — ChatGPT, Google AI Mode and Gemini, Copilot, Shop.
E-commerceAugust 5

Shopify switched on WebMCP: a browser agent searches, adds to cart and goes to checkout on its own

WebMCP is a proposed standard by which a page declares to an agent what can be done on it. Shopify storefronts became machine-readable by default: there is nothing to install and nothing to configure, and the tools for the agent are already live.

  • Support is on for every storefront on Liquid (Shopify's standard theme engine) and in a preview version of Hydrogen — the React storefront framework.
  • The page registers ten tools itself: for the catalogue — search_catalog, browse_store, get_product, show_variant; for the cart — get_cart, update_cart, cancel_cart; for checkout and orders — proceed_to_checkout, manage_orders; plus a separate search_shop_policies_and_faqs across the policy and FAQ pages.
  • An agent used to have to read the code of every page and imitate clicks — Shopify explicitly calls that approach slow and error-prone.
  • Agent support so far exists only in Chromium browsers and only in origin trial mode (a feature trialled on a limited set of sites). Everything the agent does happens inside the shopper's own live session.
  • The cart tools pull the same standard storefront actions that apps do: if your theme opens a side cart, it will open for the agent too.
What this means for a seller: clean up exactly the data the agent pulls: product variants, attributes, and the policy and FAQ pages — Shopify has a separate tool just for those.
Source: Shopify Developer Changelog (opens in a new tab)
Live A Shopify diagram from the agent documentation: the four stages of an agentic purchase — finding the product, cart, checkout, orders. That is precisely the action set WebMCP hands a browser agent.
Block 3

Rules and tools

E-commerceAugust 2

California's AI transparency law is in force: a generator must mark every image

Before: the origin of an AI image was recorded nowhere. Now: a service must embed a machine-readable mark in the file and run a free public detector. The fine counts per day of violation.

  • The law is SB 942 as amended by AB 853; the main part took effect on August 2, 2026.
  • It covers every generative service with more than 1 million monthly users that is available in California. Games, streaming and film are carved out.
  • What the service must do: a hidden machine-readable mark to an industry standard, a visible label at the user's option, and a free public detector that accepts a file, a link and an API request.
  • The detector does not disclose personal data about the file's origin, and the provider may not keep the uploaded content longer than necessary.
  • The fine is $5,000 per violation, and each day counts as a separate violation. Text is out of scope — only images, audio and video.
What this means for a seller: run your recent AI creative through your provider's public detector and decide whether to stay on a generator that does not add the mark: proving a file's origin to a platform is going to be asked of you soon.
Source: California Legislative Information, AB 853 (opens in a new tab)
Live What a public detector looks like in practice: OpenAI's image-checking page answers “Generated with OpenAI tools” after finding the watermark and the Content Credentials manifest in the file.
E-commerceAugust 6

Cloudflare shows you whether Claude and ChatGPT recommend you or not

The first way to see your visibility inside assistant answers as a number rather than a guess. Half the set is already free for everyone.

  • The AEO Visibility Dashboard works out the site's industry and category itself, runs plausible shopper queries — for recommendations, product comparisons and advice — through Anthropic and OpenAI models, and parses the answers. AEO is optimisation for assistant answers, what SEO was for search.
  • The scoring is hybrid: one model reads each answer and rates how your citations and mentions look in it, with exact text analysis layered on top so the model is not grading itself.
  • The metrics: Citation Rate (how often you are cited), Prominence (how visibly), Mention Rate (how often you are mentioned), Share of Voice (your share against competitors) and Industry Fit.
  • Alongside it runs an AI Operator Activity panel — real AI-crawler traffic and referrals by operator, including 403 and 404 errors.
  • The second half of the set, Agent Readiness, is already available in the Cloudflare dashboard: it checks robots.txt, the sitemap, markdown versions of pages and the rules for AI crawlers.
What this means for a seller: run your site through Agent Readiness and close the quick gaps, and apply for the AEO dashboard — it is still in early access, and the application goes in from the Overview tab.
Source: Cloudflare, blog (opens in a new tab)
Live The AEO Visibility panel in the Cloudflare dashboard: citation rate and share of voice against competitors
Briefs

In brief — what else happened

  • Anthropic put a spending ceiling on an agent session On August 7 a session budget for managed agents appeared on the Claude Platform: on hitting the limit the session stops with the reason budget_reached and burns no more money, and after the budget is changed it continues from the same place.https://platform.claude.com/docs/en/release-notes/api (opens in a new tab)
  • OpenAI introduced a $125 seat in ChatGPT Business On August 10 OpenAI announced a Premium seat in ChatGPT Business: $125 per user per month billed monthly ($100 billed annually) against $25 ($20 annually) for a regular seat. Premium gives five times the volume and removes the five-hour usage limit. The first 10,000 workspaces to add Premium before August 20, 2026 get $100 in credits for each such seat, for no more than five seats.https://cryptobriefing.com/openai-chatgpt-business-premium-seats-pricing/ (opens in a new tab)
  • Meta shipped Muse Spark 1.2 and a terminal agent, Muse Code August 5: 1M tokens of context, priced at $1.25 per million input and $4.25 per million output tokens, cached input $0.15. A separate plan counts the limit in tokens over a five-hour window rather than in number of requests.https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2 (opens in a new tab)
  • Google named its next flagship Gemini 4 On August 5, in a letter to staff, Demis Hassabis mentioned “our new models including Gemini 4” — the first confirmation of the name from the primary source. He himself is stepping back from operational management, with Koray Kavukcuoglu taking it over.https://blog.google/company-news/inside-google/message-ceo/next-chapter-ai-momentum/ (opens in a new tab)
  • ByteDance opened a public API for Seedance 2.5 On July 31 Seedance 2.5 opened inside ByteDance's creative products, and on August 7 the company's cloud arm officially launched a public API for the model. Up to 30 seconds of video in a single pass with no stitching, and a clip can be extended twice. Up to 50 multimodal references per generation, lip sync in more than ten languages and noticeably better-preserved audio.https://www.chinanews.com.cn/cul/2026/08-07/10673736.shtml (opens in a new tab)
  • MiniMax released the H3 weights openly On August 3 MiniMax published the H3 weights on Hugging Face: 33 billion parameters in an omni-modal transformer. Clips of 4–15 seconds, up to 2K (the base model gives 768p), 32 kHz stereo audio, up to 9 reference images, and “first frame” and “last frame” modes. The licence carves the US, the EU, the UK and South Korea out of the “applicable territory”: you cannot deploy the weights yourself there without a separate application by form.https://huggingface.co/MiniMaxAI/MiniMax-H3 (opens in a new tab)
  • Alibaba opened the Wan 3.0 beta On August 6 the wan3.0-video model appeared in Alibaba Cloud Model Studio: up to 30 seconds per generation, 5 seconds by default. Three resolutions — 480P, 720P and 1080P. Text, images, audio and video mix on the input, and the model generates the sound itself. List price is $0.05, $0.10 and $0.20 per second respectively; a thirty-second clip in 1080p comes to about $6. Status: closed beta, access by application, weights not published.https://help.aliyun.com/zh/model-studio/wan3-video-generation-api-reference (opens in a new tab)
  • Microsoft Clarity split AI citations into branded and non-branded On August 3 a Branded label appeared in the query list in Clarity's free analytics, along with a breakdown of the Share of Authority metric. The point is to separate the people already searching for you by name from the ones finding you as one option in a category.https://clarity.microsoft.com/blog/branded-non-branded-queries/ (opens in a new tab)
  • Shopify replaced someone else's model behind Sidekick with its own On August 5 Shopify's engineering blog took apart how the Sidekick agent works, answering merchant questions by composing its own Admin API queries. The external frontier model was replaced with an in-house fine-tuned one: by Shopify's calculation the same traffic on a frontier model would cost on the order of $27M a year, where the fine-tuned one costs about $1M. In a load test, time to first token fell by roughly 19% and end-to-end latency by roughly 38%.https://shopify.engineering/sidekicks-continual-learning-loop (opens in a new tab)
  • Campaign Autopilot learned to buy impressions in Copilot On August 4 Shopify added Microsoft Advertising as a channel in Campaign Autopilot, its advertising agent. The autopilot will open a Microsoft Advertising account for you or connect an existing one and run Performance Max campaigns there alongside the other channels, splitting the budget between channels itself, by performance. Early access for now.https://changelog.shopify.com/posts/microsoft-advertising-now-available-in-campaign-autopilot (opens in a new tab)
  • Alexa+ reached Australia From August 6 Amazon is rolling Alexa+ out in Australia on an early-access programme — the assistant's first launch in the Asia-Pacific region and the tenth country after the US, Brazil, Canada, Mexico, the UK, Germany, Italy, Spain and France. Free in early access, then included with Prime, and AU$29.99 a month without Prime. Amazon promises more than ten further countries in 2027.https://www.aboutamazon.com/news/devices/alexa-plus-international-launch (opens in a new tab)
  • Klaviyo bought the AI startup Agency On August 5 Klaviyo announced the purchase of Agency, a customer-experience company founded by Elias Torres. On closing, Torres becomes Klaviyo's chief product officer and takes over the whole agent line-up, including Composer (which analyses performance and optimises campaigns) and Customer Agent (which talks to shoppers directly). The close is expected in the third quarter of 2026.https://www.klaviyo.com/newsroom/agency-elias-torres (opens in a new tab)
  • Profasee is joining Titan On August 10 Profasee founder Chad Rubin announced that the AI pricing service for Amazon is moving under Titan Network. Nothing changes for existing customers at the start: same product, same prices, same team. Titan brings more than 1 billion data points on real store performance — the promise is to fine-tune the Profasee agents on them. Deal terms were not disclosed.https://www.profasee.com/blog/profasee-is-joining-titan/ (opens in a new tab)
  • Gravity raised $30.5M for advertising inside chats August 6: the company has a full advertising stack for AI assistants, and its ads already run in ChatGPT. The next product targets bots — an advertiser hands over a catalogue and a shopping agent gets a detailed supplier list.https://thenextweb.com/news/gravity-ai-ads-30-5m-series-a-agent-to-agent (opens in a new tab)
  • AI lengthens the purchase decision rather than shortening it RTB House surveyed 1,840 people in the US, the UK, France and Japan: 42% of American shoppers say AI has made the path to purchase longer, against 32% in the other countries. 59% of Americans note that an assistant leads them to unfamiliar brands (52% in the other countries) — meaning it is easier for a newcomer to get into consideration.https://ppc.land/ai-lengthens-purchase-decisions-for-42-of-us-shoppers-rtb-house-finds/ (opens in a new tab)
  • The Prompts report only shows the questions that got a click On August 3 ZonGuru took apart the Prompts report in the Amazon Ads console: it is pulled through Reports → Create report → the Sponsored Products category, holds 12 columns and is US-only for now. Only queries that got a click make it into the report — impressions without a click leave no trace. By ZonGuru's own measurement across 5,000+ listings the median listing AI-readiness score is around 64 out of 100, weakest precisely on covering shopper questions — around 54.https://www.zonguru.com/blog/amazon-prompts-report (opens in a new tab)
  • An “AI-made” label does not save an unproven claim On August 3 Brian Heidelberger of Loeb & Loeb took apart advertising with AI images: disclosing AI and the truth of a claim are two different questions. An “AI-generated” label does not earn you the right to show an unrealistic product result, and it does not defuse a challenge to a number or a promise on the same infographic.https://www.loeb.com/en/insights/passle/2026/08/aigenerated-advertising-when-is-disclosure-required-and-when-is-it-not-enough (opens in a new tab)
Wrap-up

What to do this week

1
Revisit the first bullet on three top products

A court gave Perplexity's agent Amazon back, and Shopify showed that orders from AI search tripled. The agent reads text, not infographics. Open your three top listings and check the first bullet: if it is a keyword list rather than an answer to “what it is made of, who it suits, what is in the box” — fix it. Leave the other bullets and the description alone; there is no need to redo everything over one piece of news.

2
Decide on the ChatGPT Ads pixel before August 17

On August 17 OpenAI switches advanced matching on itself, across every installed pixel. Either you keep it for accurate attribution, or you switch it off by hand in Tools → Conversions → Data Source → Edit pixel, so hashed customer data does not flow to OpenAI. After that date the decision is made for you.

3
Connect the Sellerboard MCP and recalculate your AI bill

Two checks for one evening. Connect mcp.sellerboard.com/mcp as a custom connector and ask the assistant a control question — net profit on your three top SKUs over 30 days: if it matches the dashboard, move your weekly report onto this pairing. And recalculate your monthly AI bill at OpenAI's and Anthropic's new prices — bulk work may now cost five times less.

More issues

Helium 10 just handed your data to Claude. And GPT-5.6 is catching Fable — at a third of the price. Helium 10 connected to Claude over MCP: you ask, it pulls your Amazon data. Plus GPT-5.6 against Fable 5, and Shopify's agent rules change 31 August. Read the issue Amazon's AI rewrites listing titles, Prime Day hits $26.4B the week's essentials for e-commerce From 27 July Amazon's AI rewrites listing titles longer than 75 characters. Plus a record $26.4B Prime Day, where ChatGPT traffic converted 40% better. Read the issue