Somewhere inside Meta's new shopping assistant sits a file of instructions, and one word in it says a great deal about where commerce is heading. Ranking well-known merchants first in the top five picks is "imperative," in the text Profound says it extracted on September 25. The assistant is to search at most three stores on the open web, return no more than ten products, discard listings priced far outside a category's normal range and offer in-chat checkout only for items in Meta's own catalog flagged for "agentic checkout." When a shopper names a brand, that brand's own store goes first, and direct sellers are favoured over resellers.
PPC Land reported the instruction set on October 11, drawing on an October 8 analysis of Profound's work. The numbers that came with it are stark. In a 50-prompt test run from a single account, a typical query pulled about 60 products from 30 merchants out of Meta's catalog, while the agent's browser visited three sites and found four products at two merchants. The final six picks usually split four from the browser and two from the catalog. The catalog supplied roughly 95% of candidates and won about a third of the slots. Jennifer Zou, an economist at Profound, put it plainly: "Much more weight is placed on candidates returned by the web search." Meta has not commented, the file may have changed since late September, and nobody outside the company knows how "well-known" is defined or how the three stores are chosen.
A second Profound report, covered by PPC Land the same day, adds a traffic proxy to the prompt test. Written by Allison Huang of Profound's research team and published on October 2, it tracked unverified Meta-WebIndexer requests across 280 sites between August 19 and September 27. Requests to the 115 US and Canadian sites rose 1,313% across two 20-day windows, against 827% for the 165 sites elsewhere, a ratio of about 1.6 times; daily North American requests reached about 6,700 by September 27. Profound calls the sample "small" and "exploratory," and the company sells bot analytics, so it has a stake in the question. Muse launched in the US on September 8 and runs in the US and Canada on iOS, Android, the web and WhatsApp. Third-party estimates put installs at more than 5 million in 22 days. Amazon has already blocked the agent from shopping on its site.
The concentration matters because Muse is no longer only a shopping assistant. On September 29, three weeks after launch, Meta extended Muse to small businesses in the US and Canada with 15 connectors, among them Shopify, Stripe, QuickBooks, Klaviyo, HighLevel, Canva, Figma, Notion and Slack, plus Facebook Pages, Instagram professional analytics and Meta ad accounts. Meta says nothing publishes, sends or spends without the owner's approval. What it has not said is whether access to an ad account is read-only or whether the agent can change budgets, bids, targeting and creative once approval is given. Tom Mulholland, who runs Mulholland Grocery in Malvern, Iowa, and works about 65 hours a week, said the assistant saved at least 15 minutes on payroll. "Muse is taking over a few of them," he said of his tasks.
The same company now operates an agent that decides which merchants a shopper sees and an agent that can sit inside the advertiser's account. Meta's second-quarter revenue was $60.80bn, of which $59.36bn came from advertising. That arithmetic explains why the line between the two agents will be watched closely, and why so much of the day's news turned on a single question: when software chooses, who gets to check the choice?
Luis Rijo
The agent in the buying seat
The buy side offered its own version of the arrangement. Olyzon, a Paris company with offices in New York and London, put its agentic orchestration platform into full commercial release on October 7, the opening day of Advertising Week New York. Its agents plan, activate, measure and move in-flight connected TV budgets across The Trade Desk, Amazon DSP, DV360 and Yahoo DSP. Olyzon sells no media and runs no DSP; buyers keep their own seats and direct deals, budgets move only within caps and approval thresholds the buyer sets, and every action is written to an audit log. An email agent drafts insertion order changes, but a human sends them. "The agent does the orchestration. The human keeps the decision," said Jules Minvielle, co-founder and chief executive. Olyzon has not disclosed which protocols it uses, its pricing, clients or performance, or exactly which actions require sign-off.
Publishers sound less hurried. Digiday spoke to three senior commercial executives at the same conference and reported on October 12 that most still treat agentic media buying as test-and-learn rather than proven revenue. Jamie Jouning, global head of advertising at The Economist, said: "It would be very hard for me to go and seek investment from our leadership team" for something whose value is unknown. The Economist will test in the background while it expands video from two shows a week to six from January. Newsweek's chief revenue officer, Danielle Varvaro, said the company is evaluating but has signed no agentic deal. Joy Robins, global chief advertising officer at The New York Times, was blunter still: "That's kind of the evaluation phase, and that's where we are right now." Amit Grover of KPI Village argued that publishers prioritise near-term revenue, while agentic buying mainly cuts operating costs.
Yext presented a third model, with many people directing the same agents. Its multiplayer agent harness inside Scout, announced on September 30 alongside its fiscal second-quarter results, gives corporate staff, field teams, franchisees, local managers, partners and agencies a shared workspace in which to instruct agents from shared context. Scout monitors traditional and AI search, listings, reviews, web pages and social channels, and sends alerts down to a single location. The harness is live only for Yext's own account teams; customers can join a waitlist, and there is no general availability date. Yext's release opened with a line that could serve as the epigraph for most of this year's agent launches: "Brands bought AI agents expecting growth. Most are getting productivity instead." The company did not disclose its permissions model, approval workflow, underlying language models or pricing.
Microsoft, meanwhile, gave advertisers a way to tell its machines what not to say. PPC News Feed reported on October 11 that Microsoft Advertising has added messaging guidelines to text guardrails for AI Max campaigns, letting advertisers name themes, claims, products, offers or topics to emphasise or avoid in automatically generated assets. Guardrails now come in two parts: up to 25 term exclusions and up to 40 messaging guidelines. The feature is in a closed pilot, spotted by Hana Kobzová, and follows Microsoft's September rename from "text guidelines" to "text guardrails."
Before an agent buys anything, some agencies are now asking a simulated audience what it thinks. Sam Bradley's Media Buying Briefing for Digiday on October 12 described synthetic audiences, personas and virtual focus groups generated from survey data, moving from experiment to routine. Goodway Group has worked with Arima since August on a synthetic population of US, British and German consumers, running incrementality tests in simulation and confirming them with small live activations. "Against a synthetic audience, we can simulate it without actually having to run a test," said Paul Frampton-Calero, its chief executive. Vurvey Labs, which works with Magnum, Vans and Kenvue, said a strategy brief for Aveeno Body had gone from 30 days to three. Stef Hoffman of Code & Theory offered the caveat: "It informs judgment, but it doesn't replace judgment." The same briefing carried WARC's forecast that global advertising spend will grow 11.9% to $1.34tn this year, with Alphabet, Amazon and Meta taking 59.7%, about $659.6bn.
Who checks the machine
If agents are to buy, sell and write, someone has to verify what they do, and the verifiers are arriving with notable limits on what they can see. OpenAI describes its arrangement with Integral Ad Science and DoubleVerify as a "brand suitability evaluation pilot," in which the two companies assess how OpenAI's own safeguards are applied "in a controlled testing environment" and "without accessing private user conversations." PPC Land compared the two companies' October 5 posts and found they describe different things. IAS calls it an "integration" delivering measurement, never uses the word pilot and does not mention DoubleVerify. Its post argues that conversations need "continuous, dynamic contextual evaluation" rather than page-level categories, which is difficult to square with a test that excludes the conversations. What data the verifiers receive, when the work started, which markets it covers and whether any results will be published were all left out. ChatGPT ads now run in 63 countries, and OpenAI's David Dugan has said they passed a $1bn annual run rate.
IAS extended a tool on the same theme to YouTube on September 30. Low-Quality GenAI Avoidance now covers YouTube globally through YouTube Optimization and Total Media Quality, applying exclusions before bids are placed and refreshing them daily. The classifications are built from "contextual signals from the YouTube Data API," meaning titles, descriptions and tags, and are not reviewed by Google. IAS gave no accuracy or false-positive rates, did not say whether it analyses frames, audio or transcripts, and did not say whether Shorts and connected TV are covered. Earlier research gives a sense of scale: Kapwing found in December 2025 that 104 of a new account's first 500 Shorts, or 21%, were AI slop, and a TAG, ANA and Fiducia analysis in July put slop at 1.3% to 2.4% of open-web programmatic spend.
Trust also depends on where a conversation goes once it has been shared. DuckDuckGo launched share links for Duck.ai on October 7, and its design is the opposite of the indexing accidents that have exposed ChatGPT and Claude conversations in search results. The chat is encrypted on the sender's device; the URL carries a server identifier and a decryption key in the fragment after the hash, which browsers normally do not send to servers; the server stores only ciphertext. Links expire after one, two, seven or 30 days, with 48 hours as the default, and pages carry noindex in both the HTML and the HTTP header. The company describes the arrangement as "a locked box" whose key goes straight to the recipient. It has not named the encryption algorithm or key length, and no independent audit has been published. Anyone holding the URL holds the key.
Identity documents have a sharper version of the same problem. In September, journalists found a dark-web market selling scans of about 160 million US and Canadian driver's licences, reportedly taken from a compromised identity-verification service. On October 7 the W3C urged governments to adopt revocable verifiable credential barcodes, which pair a digital signature with a revocation check based on the Bitstring Status List standard. The issuer flips a bit in a published list, and verifiers download the list, so the issuer never learns when a card is used. The argument rests on one sentence from the release: "A digital signature alone does not prevent either attack." A copied image of a valid barcode still passes a signature check. California's DMV has issued licences with the barcode since late 2025, though it is not confirmed whether those cards include a status-list entry.
And sometimes the thing being sold is not what it claims to be. Zohar Pinhasi, 50, owner of the ransomware recovery firm MonsterCloud, was arraigned in the Eastern District of New York on October 7 on two counts of wire fraud and one of wire fraud conspiracy. Prosecutors say the company marketed a no-ransom service built on "proprietary tools" while secretly paying attackers for decryption keys. PPC Land's account of the indictment sets out the figures: clients were charged more than $19m, more than $8m went to ransoms, and in one August 2023 case about $8,200 paid to an attacker became a bill of about $150,000. "Instead, he turned the victim's crisis into his own profit center," said James C. Barnacle Jr., an FBI assistant director. The charges are allegations, and each count carries a maximum of 20 years.
Footnotes that do not match the headline
The day's measurement claims showed the same gap between what is offered and what can be checked. Snap's new advertiser campaign, Spend Smarter, published on September 30 and built on 1990s nostalgia, claims an incremental return on ad spend "approximately 19.3%" higher than brands' blended social spend, for brands in Measured's portfolio. The footnote's study title says the figure is 12.9%. The two numbers are not reconciled, and Snap published neither the study's method, sample nor date. "Spend Smarter isn't about telling marketers to spend more," the company said; the aim is to win share of existing budgets. Snap's second-quarter advertising revenue rose 9% to $1.28bn while cost per impression climbed about 10%.
Meta's own measurement paper, released on October 7, runs to 54 pages. Its headline is that campaigns with at least 30% of assets as partnership ads achieved a 45% lower median cost per incremental purchase and 35% higher incremental ROAS across 20 EMEA conversion lift studies run between January 1 and June 30, 2025. The confidence level is 92%, below the 95% convention, and page 24 says "over 30%" where elsewhere it says "at least 30%." The paper leans on IPA data showing creator activity with a short-term ROI index of 99 against an all-media average of 100, and a long-term multiplier of 3.35 against 3.27 for linear TV and 1.98 for paid social. Derya Matras, Meta's vice-president for EMEA, wrote in the foreword: "Measurement is usually the last thing our industry gets around to understanding." Meta, it should be said, is measuring inventory it sells.
NIQ has taken a different route to credibility by building on a model anyone can read. It has integrated Google's open-source marketing mix model, Meridian, into NIQ Cadence, its marketing effectiveness system, so that model building, scenario planning, optimisation and monitoring run in one workflow fed by NIQ data and advertiser inputs. "Marketing leaders have told us they don't need another standalone model," said Jason Tate, general manager for marketing effectiveness. NIQ's own survey found 74% of marketing leaders under more pressure to prove ROI and 9% with adequate real-time insight; no method was given. Pricing, markets, clients and the datasets behind the model were not disclosed, and Google offered no quote. Meridian needs 12 to 18 months of historical data.
The marketers themselves admit to large blind spots. A survey of 250 UK consumer goods leaders by Censuswide for Kantar and Binet Consulting, presented at the IPA Effectiveness Conference in London on October 7, found that 75% could not estimate what share of their price promotions were profitable over 12 months, even though promotions take 36% of marketing and trade spend against 17% for paid media. Some 76% repeat promotions without assessing profit, and 68% would run fewer if retailer support were not at risk. "Price promotions are like class A drugs: expensive, dangerous and addictive," said Les Binet. The illustrative model in the deck, which is not survey data, suggests only about 15% of promotional volume is genuinely new. PPC Land noted that the deck's figures do not always add up: 28% and 55% make 83%, not the stated 82%.
Search budgets show their own drift. Brainlabs data from 31 Google Ads accounts between January 2024 and August 2026 shows that Search campaigns' share of clicks fell from 62% to 41%, while their share of spend fell only from 86% to 71%. A separate Brainlabs study of 54 clients found organic sessions down 10.5% after the arrival of AI Overviews, from 140,132,402 to 125,439,245, while AI referral sessions grew 163.5% to 1,323,650, replacing about 5% of the lost traffic. "The click is now one of the last steps in the decision," Alessio Messaggi wrote on October 7. The click mix reflects Brainlabs' own portfolio choices, and the post says ChatGPT ads have no conversion optimisation, which OpenAI's June and August releases contradict.
Even the data Google gives site owners arrived late. Barry Schwartz at Search Engine Roundtable reported on October 11 that the Search Console performance report was more than 64 hours behind at about 7am ET, which he called "incredibly unusual for this report"; by 11am the lag had fallen to 44 hours. The delays followed the September spam update, and Google had not commented when the story was published. Schwartz doubted that the lag signals a core update, though he expects one soon.
Apple's block and the open web's bill
The bluntest act of gatekeeping in the window came from a browser. Digiday's Ronan Shields, Jessica Davies and Sara Guaglione tallied the early cost of Apple's latest Safari tracking block on October 12. About a week earlier, Apple's update blocked code that lets ad tech companies recognise the same reader across sites, using a blocklist that is not public. Aditude, which helps publishers sell inventory, found that Safari on iPhones and iPads accounts for 11.7% of ads shown and 11.4% of revenue for its customers, and that eCPMs fell 10.5% among readers on the update. With only about 8% of users upgraded, the overall revenue hit so far is about 1%. The Trade Desk's ad server was cut off without warning; Apple has removed it from the blocklist in a test version of iOS, but users on the current version stay blocked and a fix is not expected before the end of the month at the earliest. "I would say it's going to get worse," said the consultant Davide Rosamilia. Andrew Casale, chief executive of Index Exchange, called the update an "invasive overreach" and asked: "Is that a restoration, or is that an admission of a pretty massive error?" The timing, at the start of the fourth quarter, is the worst possible for publishers.
The shelf gets wider, the data gets closer
While browsers narrow what can be tracked on the open web, retailers and platforms keep building first-party targeting closer to the purchase. Amazon has added pre-built brand penetration indices to Amazon DSP in nine countries: the US, Canada, Germany, Spain, France, Italy, the UK, Australia and Japan. They compare a brand's penetration with its category's, using Amazon's retail data and the advertiser's tracked ASINs, show the result on a map and refresh weekly without any Amazon Marketing Cloud setup. Advertisers choose whether to chase low- or high-penetration markets at ad group level. Amazon says the indices cut the time from insight to activation "from days to minutes," without giving a baseline. The granularity of the maps, the size of the bid adjustments and the definition of penetration were not disclosed.
Retail media inventory is also becoming easier to buy outside the retailer's walls. Teads and Koddi announced on September 29 that Koddi-powered onsite inventory from Wolt Ads, Gopuff UK and Hopper is now buyable in Teads Ad Manager over OpenRTB, covering sponsored product ads and display in the US and Europe. Nicholas Ward, Koddi's president, called it "a significant milestone in making retail media truly programmatic on a global scale." Fees, revenue share, a go-live date and the full list of networks were not disclosed. Teads' second-quarter revenue fell 17% to $284.6m, and it suspended guidance on August 6.
DoorDash wants the customer record itself. Its new layer, DashOS, announced on September 30, links a restaurant's customer profiles across the DoorDash marketplace, direct online ordering, SevenRooms reservations, in-store rewards, loyalty and voice AI. From 2027 it will extend to more than 80,000 restaurants using third-party point-of-sale systems, including Clover, Qu, Checkmate, Chowly, Deliverect, Stream and Urban Piper. DoorDash says loyalty members order 1.8 times as often; that compares self-selected members with non-members and proves no causation. "Our guests don't think in channels," said Marco Montevago, a Bay Area restaurant manager. Consent, opt-out, pricing and whether the data feeds DoorDash Ads were not explained.
Nextdoor is selling a finer map. Its Neighborhood Intelligence product, unveiled on October 7, lets advertisers target 350,000 drawn neighbourhood boundaries rather than ZIP codes, combining recommendations and interests, mover and household data, and local business activity, weather and sentiment. A covered ZIP averages 12 neighbourhoods. Nextdoor's own pilot figures show brand awareness 1.28 times higher, ad recall 1.16 times and favourability 1.28 times, all upper-funnel survey measures, with the comparison group described differently in the press release and the blog. Michael Kiernan, chief revenue officer, said: "We kept hearing the same thing from customers across every vertical." It is a managed service for US advertisers, available on request.
Pinterest is trying to stretch the season rather than the map. A Pinterest Business UK guide dated September 29 claims that one in two monthly European users keep shopping after Christmas, urging advertisers to refresh creative for "Q5," the week from December 25 to January 1. Its only named advertiser, TUI, reached 41% more people with Premiere Spotlight than its annual average. All nine footnotes were collapsed on the page and could not be read, and the 2.7% ROAS figure for ad groups with five to ten ads is correlational. Pinterest's average revenue per user in Europe was $1.35 in the second quarter, against $8.30 in the US and Canada.
Screens, speakers and the attention between them
The living room and the earbud saw expansions of their own. Amazon's Fire TV will show live programming from more than 40 apps, including DirecTV, Peacock, Sling, Tubi and YouTube TV, on its home screen and Guides tab, with Alexa+ search that draws on subscriptions, watch history, availability and "world context." The updates reach US customers in November; the new Fire TV Stick 4K and other countries follow in early 2027. Amazon claims customers find something to watch "tens of millions more times per month," with no baseline. Whether the new hub carries sponsored placements, and whether payment affects how live channels are ranked, went unanswered.
Spotify announced at the Frankfurt Book Fair on October 7 that it is taking audiobooks from 22 markets to 181 by the end of 2026, with more than 350,000 titles in over 120 languages. Audiobooks+ has more than one million subscribers and more than $100m in annual recurring revenue, and Premium subscribers get 12 hours a month, down from the 15 hours offered at the 2023 English-language launch. None of the access models includes advertising, and prices in the new markets were not disclosed. Spotify's earlier disclosures on listening growth have shifted from 37% to 60% to "more than 30%," which makes the trend hard to read.
SiriusXM Media said at its Built With Audio event in New York on October 6 that adding YouTube audio inventory to beta campaigns lifted total unique reach by up to 91%, raising its unduplicated US monthly audience to about 255 million from 170 million a year earlier. No campaign count, baseline, method or measurement provider was given. SiriusXM is the exclusive US reseller of YouTube and YouTube Music audio-first inventory under an April 22 deal, but its chief financial officer expects meaningful revenue only in the second half of 2027. "YouTube brings millions of incremental listeners," said Scott Walker, chief ad revenue officer.
Disney supplied a reminder that broadcast still marks its milestones. Abbott Elementary's sixth season premiered on ABC on October 7 and streams on Hulu from the following day, with the series passing 100 episodes during the season. Disney's feature stressed families watching together and quoted Quinta Brunson's aim to "make a show that people could watch with their family members," but contained no ratings, streaming hours or demographic data, and named no sponsors. Disney's streaming advertising revenue rose 12% to $821m in its fiscal second quarter.
Google is turning prompts into play. Playground, an experimental browser platform announced on October 7, lets users create, play and share games by typing instructions, with a public gallery ranked by player ratings and play activity. Creation is limited to US adults aged 18 and over and is being rolled out by Google AI subscription tier, with plans priced from $4.99 to $199.99 a month. A Unity Spark tie-in offering 3D and the Unity runtime is in testing, with a closed beta "coming soon." Google did not say which model builds the games, how age is verified, or whether creators can earn money or carry ads.
Creators, commerce and the classroom
YouTube's chief executive has made his bet explicit. In a post-show interview after Made on YouTube 2026, recorded on September 23, Neal Mohan named brand-creator commerce and Gemini-based AI as YouTube's two priorities, with the shopping affiliate programme due to reach 35 countries by the end of 2026. "Creators really are the most trusted voice," he said, "and brands really want in on that." The creators in the room were less certain about the new tools. YouTube now lets creators test up to three cuts of a video and up to three thumbnails served to different audience segments, but Marques Brownlee noted that "a small difference between thumbnails doesn't really help me learn very much." No release dates or eligibility rules for video testing were given.
The creator economy is also becoming a degree subject. Alyssa Mercante reported for Digiday on October 12 that Arizona State University offers a BA in content creation, Stanford Graduate School of Business runs a course taught by Matt Abrahams that drew a 110-person waitlist, and Syracuse University has opened a Center for the Creator Economy whose first report, in early October, found YouTube has overtaken television as Americans' top source of political news. Ryan Schram, the centre's executive director, estimated that about 50 million people take part in the creator economy worldwide. "The creator economy has crossed the line from an emerging marketing channel into a legitimate business discipline," said Cole Mancini Elboeck of Go Fish Digital.
The calendar of the vote
That finding about political news lands three weeks before a national election. Meta will block new US political, electoral and social issue ads from 12:01am PT on Tuesday October 27 to 11:59pm PT on Tuesday November 3, election day, under a notice dated October 6 and signed by Mani Pandya. Ads that delivered at least one impression before the window can keep running, with only bid, budget, end date and pause status open to change; creative, placement, targeting and optimisation edits are blocked. Meta recommends authorisation by October 13 and submission by October 20, though the binding requirement is the impression before October 27. The stated reason is that "there may not be enough time to contest new claims made in ads." The notice does not say which of Meta's apps are covered or how boosted posts are treated. AdImpact projects $11.6bn in midterm advertising, more than the $11.2bn spent in 2024.
Those who prefer to bet on the outcome may take some comfort from economics. NBER working paper 35846 by Eric Zitzewitz, highlighted in an a16z newsletter on October 9, found no partisan bias in prediction markets in data going back to 1880: a partisan coefficient of 0.03 percentage points, with a standard error of 2.66. Non-US elections were the only outlier, at minus 1.42, and the result there was not significant on a much smaller sample. The paper measures bias rather than accuracy, and its abstract notes a favourite-longshot bias and slight overpricing of left-leaning outcomes in polling-average markets, which the newsletter left out.
What the day added up to
The thread running through these stories is not that machines have taken over marketing decisions. It is that the decisions are moving into places where they are harder to inspect. A shopping agent ranks merchants by a rule written in a file that a vendor had to extract. Verifiers test a chatbot's ad safety without seeing the chats. A safety classifier for YouTube reads titles and tags, and nobody publishes its error rate. A browser blocklist that removes a tenth of some publishers' iPhone revenue is not published either. Measurement claims arrive with footnotes that give a different number, confidence levels below convention and survey totals that do not add up.
Against that, the most credible products of the day were those that published their constraints: Olyzon's audit log and buyer-set caps, Microsoft's 40 messaging guidelines, Duck.ai's 48-hour default and fragment-held key, the W3C's revocable status list, Meta's political ad calendar with its exact minutes. None of them proves that the systems behind them work as described. Each at least states what the system is allowed to do and where a human, or a published standard, can check it. Muse's instruction file may be the most revealing document of the week precisely because Meta did not publish it. The next one may not need to be extracted, if advertisers and merchants ask to see it first.
By the numbers
45%
Lower median cost per incremental purchase Meta reports when partnership ads made up at least 30% of assets.
Meta says partnership ads cut incremental purchase costs 45% in EMEA tests
350,000
Neighbourhood boundaries Nextdoor now offers advertisers as an alternative to ZIP-code targeting.
Nextdoor lets advertisers target 350,000 neighborhoods instead of ZIP codes
41%
Search campaigns' share of Google Ads clicks across 31 Brainlabs accounts, down from 62%.
Search campaigns drop to 41% of Google Ads clicks in Brainlabs data
More than $8m
Paid to ransomware attackers by MonsterCloud, prosecutors allege, while clients were billed over $19m.
MonsterCloud owner faces 3 fraud counts over $8M in secret ransom payments
October 27
First day Meta stops accepting new US political ads, eight days before the midterm vote.
Meta blocks new US political ads for 8 days before the midterm vote
Discussion