Retail media was built on receipts. A shopper bought a product, the retailer recorded it, and an advertiser paid to reach people who had bought that product before or had bought something adjacent to it. The entire proposition rested on a record of something that had already happened, which is why retailers with loyalty programmes and membership cards commanded a premium over publishers selling inferred audiences.

On September 17 a warehouse club began selling the opposite. Sam's Club Connect launched a set of targeting products whose central claim is not what a member has purchased but what a member will purchase, in some cases a full year before the transaction. One of them identifies households that have never bought a product at all. Another watches for purchases of soundbars and gaming consoles in order to find people who are about to buy a television. The evidence offered for all of it is a single pilot with one consumer goods manufacturer.

The following morning, Walmart Data Ventures told suppliers at its Inspire conference that Scintilla would open marketplace data to first-party suppliers next year and that its embedded assistant would move from describing what the numbers say to instructing suppliers on what to do about them. Axios, meanwhile, described a product line built for customers that do not yet exist in commercial volume: three content feeds sold to AI models, corporate systems and, eventually, personal software agents. Snap priced a pair of augmented reality glasses at $2,195 for a market that has not been demonstrated. And Google spent the week rearranging what appears on its search results pages, removing one surface entirely, without announcing any of it.

The common structure is a bet placed on a described future and priced today. Some of those bets carry disclosed evidence. Most do not.

A membership file becomes a forecast

The four products Sam's Club Connect introduced on September 17 share one design principle: each converts a longitudinal record of member behaviour into a claim about behaviour that has not occurred yet.

Predictive Precision Targeting is the most aggressive of them. It builds segments of households with no prior purchase history in a product category, which inverts the logic that made retail media attractive in the first place. Procter and Gamble ran the pilot with Cascade Platinum dishwashing detergent pods. Sam's Club says the model reached a 94 per cent accuracy rate across roughly 800,000 identified households, and that spending on Cascade Platinum Plus rose 71 per cent against what the company describes as typical campaigns.

Those two figures deserve separate treatment. A 94 per cent accuracy rate is a meaningless number without the denominator and the time window that produced it: accuracy against what prediction, measured over what period, and against what base rate for a mass-market dishwasher detergent sold in a warehouse club. A category that a large share of members buys anyway will produce a high hit rate from almost any model. The 71 per cent spending lift is the more interesting figure, and it is anchored to a comparison, typical campaigns, that the announcement does not define. No media spend, no campaign duration, no control group design and no measurement partner were disclosed.

Harvey Ma, vice president and general manager of Sam's Club Connect, framed the underlying claim in terms of time rather than precision. "What we're trying to demonstrate is that much more than historical purchase behavior, this machine learning model is an indication of what might come 12 months into a member's life cycle," he said. The 800,000 households, in his description, were pre-identified for upgrade propensity.

The second product addresses attrition rather than acquisition. It identifies what the company calls lapsed or staggered buyers, members who once purchased a brand and then stopped or slowed, and combines historical shopping records with predictive modelling to reach them before a competitor does. This is the least speculative of the four, since the underlying signal is an observed gap in a purchase sequence rather than an inference about a category a member has never entered.

The third is the one most likely to be copied. Sam's Club is segmenting households predicted to buy expensive items with long replacement cycles, televisions and refrigerators among them, using adjacent purchases as leading indicators. A holiday 2026 test tracks soundbar, gaming console and streaming device purchases in order to anticipate television buying. The same logic is applied elsewhere: a pattern of laundry product purchases is treated as a signal that a washing machine may be nearing replacement. Buying a soundbar before buying a television is a sequence that a warehouse club sees and a publisher never will.

The fourth enables conquesting, targeting the customers of rival brands within a category, with what the company describes as guardrails preventing the tool from collapsing into simple retargeting. It runs across both self-serve and managed service campaigns.

What makes any of this saleable is the membership file. A warehouse club ties cash transactions to a member account, which closes the gap that defeats most retailers: the anonymous shopper who pays without identifying themselves. Longitudinal coverage of a household across years of grocery, apparel, electronics and fuel purchases is a materially different asset from a browsing history or a view-through conversion log. The models sitting on top of it use large language models and machine learning, though the announcement specifies neither the architecture nor how the predictions are validated after a campaign ends.

The corporate context matters here, because the same parent company approached the same problem from the supplier side one day later. At the Walmart Data Ventures Inspire conference, the division said Scintilla would add marketplace data for first-party suppliers next year, alongside customisable dashboards that connect multiple data sets, custom alerts, and tighter integration with Walmart's replenishment systems. Suppliers who sell both to Walmart directly and through its marketplace will be able to see both sides of their business in one place for the first time. Sellers who operate on the marketplace alone remain ineligible, with no announced plan to change that.

Mark Hardy, head of Walmart Data Ventures, described the objective as a single method of working across both channels. "We're having one consistent approach of working and collaborating with our suppliers, and eventually marketplace sellers as well," he said. The exclusion of marketplace-only sellers, set against the phrase eventually marketplace sellers, is the sort of gap that tends to persist for several product cycles.

The assistant embedded in Scintilla, named Marty, is where the year-ahead ambition sits. It currently explains metrics, flags changes between reports and summarises findings. Hardy described the next stage plainly: "What we're doing now is taking AI and moving from summarizing findings or highlighting key points to, now, [making] recommendations." The operational version of that shift, in his words, is that the tool "will allow you now to bring all those data points together and help you get to the conclusion of what you should do next."

That is a meaningful change in liability, if not in software. A summary that a supplier misreads is the supplier's error. A recommendation that a supplier follows is a different arrangement, particularly when the party issuing the recommendation also sells the media, sets the shelf placement and operates the replenishment system that the recommendation touches. No pricing for Scintilla was disclosed, and no user count.

Scintilla In-Store, launched in February 2026, extends the same data to supplier field representatives at store level. It is the current iteration of Volt, which Walmart acquired in 2022. Walmart has also said it plans to extend Scintilla to Sam's Club suppliers next year, which would put predictive targeting and supplier-facing analytics on the same membership data inside the same company. John David Rainey, Walmart's chief financial officer, has said the division continues to drive meaningful growth without attaching a revenue figure to it.

Amit Dodeja, chief marketing officer at Spreetail, gave the seller-side reading. "It's a wise move for Walmart to head in that direction. It only strengthens the confidence that a third-party seller has," he said. Confidence, in this context, means visibility into a channel that sellers have historically navigated by inference. It is worth noting how closely that description matches what advertisers say about retail media generally, and how rarely either group gets the underlying data rather than a report about it.

A publisher builds pipes for readers that are not people

Axios spent this week describing a customer base that barely exists. The company is preparing three feeds under the name Axios Direct, each aimed at a different kind of machine, and announced them on September 16 at the Digiday Publishing Summit in Key Biscayne, Florida. The distinction from the anonymous contract complaints that dominated that event is the point: this is a named publisher, on the record, describing a product with a stated pricing model rather than a licensing agreement it is forbidden to discuss.

The first feed is live and sells to investment firms. The reference model is explicit, and it is not a media business: a Bloomberg terminal, priced for institutions that will trade on market-moving information quickly enough for speed to carry a premium. Pricing runs on an annual fee scaled to company size and assets under management, with existing Axios Pro Deals subscribers as the primary upsell target. Contracts run two years.

Jacquelyn Cameron, chief revenue officer at Axios, explained the term length as a hedge against a product whose buyers are still forming. "We think that two years is really helpful because it allows us the opportunity to experiment and to fine-tune and to deliver on what could be a changing need from these clients within the contract terms," she said. That is an unusually candid description of selling into an undefined market. A two-year contract with a customer whose requirements may change is a research budget with revenue attached.

The second feed, still in development, targets companies running internal AI systems for employees. Distribution is direct to clients rather than through third-party marketplaces, which is the structural decision that separates this from conventional content licensing. A publisher selling into a marketplace accepts whatever terms and whatever measurement the marketplace provides. A publisher selling a connector directly to a corporate buyer keeps the relationship, the pricing and the usage data. No launch timeline has been set.

The third is the most speculative and the most revealing. It would serve individual AI agents acting for personal subscribers inside systems such as ChatGPT and Claude, functioning as the equivalent of an RSS feed for consumer software rather than for a reader. Exploration is planned for 2027. There is, at present, no established mechanism by which a consumer agent pays a publisher for access on a subscriber's behalf, and no standard by which either party verifies what was consumed.

The commercial results give the strategy its context. Axios reached its full-year revenue goal in September, a month earlier than in 2025, when the same milestone arrived in October. Growth against 2025 was described only as falling between 10 and 50 per cent, a band wide enough to convey almost nothing. The more useful disclosure concerns mix: between 52 and 54 per cent of this year's revenue will come directly through client relationships rather than intermediaries, which is the figure that explains why a publisher would build machine-readable feeds priced per institution instead of chasing distribution.

Cameron also raised the possibility of branded content or advertising aimed at AI agents inside the feeds, citing Time's approach. That is a second bet stacked on the first. The initial wager is that machines will pay for structured access to journalism. The second is that something recognisable as advertising can be sold inside the pipe that carries it, to a reader that does not have eyes.

A $2,195 device for a market that has not arrived

Snap priced its SPECS augmented reality glasses at $2,195 on September 17, with a $200 refundable deposit to hold a pre-order and shipping expected in the autumn across the United States, the United Kingdom and France. A cellular bundle costs $2,395 and includes a charging case carrying four additional charges. Verizon data plans run $10 a month for existing Verizon customers and $20 for everyone else. A demonstration space called SPECS First Look opens at Westfield Century City on October 1, 2026.

The hardware is specified in unusual detail for a consumer launch. The glasses weigh 132 grams in the 47mm size and 136 grams in the 52mm, sit on Swiss TR90 polymer frames, and project a 51-degree field of view through a waveguide the company describes as carrying billions of nanostructures rendering 16 million colours. Audio runs through custom stereo speakers and six MEMS microphones in an open-ear design. Two processors divide the work, with world understanding handled separately from the experiences running on top of it.

That processor split is not only an engineering decision. It is the mechanism behind the privacy architecture, because applications never automatically reach raw camera and sensor data. Snap has committed to no facial recognition, no continuous recording, an external LED that illuminates during capture, permission gates before applications reach sensitive data, and moderation of every Lens before publication. The assistant layer, SPECS Intelligence, is in preview on iOS for United States adults aged 18 and over, with Mac access by invitation, and offers anticipatory prompts drawn from connected accounts such as mail and calendar. "SPECS won't send a message, make a change, or take an action until you say yes," the company states.

One commitment carries direct commercial weight: personal content from connected accounts is excluded from model training and from personalised advertising. Snap is declining, at launch, to use the richest behavioural data the device produces for the purpose that funds the rest of its business. Meta has taken the opposite position on its own assistant data, having begun using AI chat interactions for ad personalisation in December 2025. Whether an advertising company can hold that line once the hardware needs to justify its cost is the question the next two years will answer.

The competitive framing is stark. Meta's Ray-Ban Display, shipping since September 2025, costs $799, and is a display attached to glasses rather than a spatial computer. Google's Android XR hardware is expected in 2027, Meta's full augmented reality device in late 2027, and Apple's entry in 2028 or later. Snap is therefore roughly a year ahead of every larger competitor, at nearly three times the price of the nearest available product, into a category with no demonstrated consumer demand at any price.

The developer and partner list is the hedge. Lens Studio counted more than 400,000 professional developers and teams on its desktop version as of June 2025, though only hundreds are described as building SPECS Lenses specifically. Content partners include YouTube, HBO Max, Spotify, the NBA, the WNBA, ALO Wellness Club, Tripadvisor and Shopify. Enterprise deployment is where the near-term revenue case sits: Salesforce Agentforce for field workers, Amazon Quick through AWS, an NVIDIA XR AI stack for spatial interpretation, and implementation partners including Trifork and Hololight.

Evan Spiegel, Snap's co-founder and chief executive, described the product in terms of shared experience rather than utility. SPECS "are built from the ground up to make computing more human," he said, and are meant to be "incredibly easy to use and bring new experiences into the world that you can share together with friends and family."

The corporate structure says more than the quote. SPECS Inc. operates as a wholly-owned Snap subsidiary, deliberately arranged to permit outside investment and partnerships. A company that expected the device to pay for itself from advertising would not need that structure. A company preparing to raise separate capital against a hardware bet would.

Google rearranges the page and says nothing

Four changes to Google's surfaces surfaced within two days, none of them announced by Google, all of them spotted by practitioners watching screens.

Sponsored Places advertisements in search results are being tested in an expandable format that opens on hover, revealing additional detail with overlay and click behaviour attached. Khushal Bherwani shared screenshots on September 15 and the test was documented on September 17. Google has tested hover expansion on shopping and image advertisements previously, so the mechanism is familiar; applying it to local Sponsored Places extends paid units into more of the page without a formal size increase, since the additional real estate is claimed only on interaction.

Free listings are moving in the other direction. Google is testing product grid results that link straight to a merchant's product page, removing the overlay panel that previously intercepted the click and presented several retailers at once. Brodie Clark documented the change on September 16, noting it extends a test run inside Google Shopping in June 2026. The overlay has been a point of friction for retailers, because it inserted a comparison step between a shopper's intent and a merchant's page, and its removal on free listings is a straightforward improvement for individual sellers. It also removes a surface on which Google controlled what else the shopper saw.

The most consequential change was a removal. Google Discover, which the company tested on its desktop homepage in 2023, announced for desktop in 2025, and had running across multiple countries as recently as June 2026, has been switched off globally on desktop. Damien Andell documented the disappearance by comparing countries between June and September. The desktop homepage has reverted to what the report describes as just a big search box and a lot of white space. Google has issued no explanation. For publishers who had begun to model Discover as a partial replacement for declining search referrals, a distribution surface appeared, ran for roughly a year, and vanished without notice on one of the two devices that matter.

The fourth change alters who describes a business. Google is rendering local knowledge panels as AI Overviews, placing an AI Overview label at the top of a Business Profile listing with a Show more control that expands into a chat interface. Ben Fisher identified it on LinkedIn and Barry Schwartz reproduced it independently on September 17. Schwartz found the generated description broadly accurate but slightly out of date, and fairly noted that his own site content was out of date too. The structural point survives the caveat: a Business Profile is a record a proprietor fills in and controls, while a generated summary of that profile is a description the proprietor can influence but not write.

Set those four alongside the week's other announcements and the pattern resolves. Sam's Club will sell access to households it has decided are twelve months from a purchase. Walmart will tell suppliers what to do next rather than what happened last week. Axios is pricing feeds for institutional customers whose successors have not been built yet. Snap has priced a spatial computer a year ahead of every rival and ring-fenced the data that would make it pay. Each of those is a claim about a future state, offered with varying amounts of evidence, and each is available for purchase now. The changes to Google's pages are the counterpoint: material, immediate, already live in front of shoppers and business owners, and accompanied by no statement at all. The predictions arrive with press releases. The facts arrive on somebody's screenshot.

Also noted

  • September 17: IAB Australia's Creator Connect 2026 found 61 per cent of Australian online shoppers discover brands through creators, rising to between 84 and 85 per cent among 18 to 39 year olds and falling to 21 per cent among those aged 60 to 70, while 44 per cent said a paid-partnership disclosure makes them less likely to act on a recommendation (PPC Land).
  • September 17: Nielsen moved streaming measurement from weekly to daily reporting, with its top 10 charts now arriving 11 days after a reporting period closes (AdExchanger).
  • September 17: Formula 1's shift to Apple TV cut United States household reach by 68 per cent at the Miami Grand Prix and 66 per cent at Monaco against 2025, while Monza rose 15 per cent with average viewing time up 78 per cent, as Ampere Analysis projects sponsorship revenue above $3 billion in 2026 (Digiday).
  • September 17: Todd Parsons, previously chief product officer at Criteo, becomes chief executive of Blackstone-backed Prodege on October 1, taking over a consumer panel of more than 125 million people behind Swagbucks, MyPoints and InboxDollars (Adweek).
  • September 17: Microsoft is testing new labels on Bing product carousels including Top Picks, Price Drop, Sale and view counts, spotted by Khushal Bherwani, with Search Engine Watch reporting the view count display (Search Engine Roundtable).