Every previous filing on ChatGPT advertising in this newsletter dealt with OpenAI selling its own space: the self-serve manager, the country expansions, the bidding defaults. What changed on September 10 is the identity of the seller. Amazon Ads and OpenAI opened a pilot that lets select US advertisers extend campaigns already running in Amazon DSP into ChatGPT, which means a buyer sitting in Amazon's interface can now reach a surface Amazon does not own, does not host and cannot instrument end to end. That is not a rollout. It is a change in who controls the demand path into the largest consumer AI product in the market.

Chris Conetta, director of omnichannel supply at Amazon DSP, framed the arrangement in terms of growth rather than mechanics. "Conversational ads represent the fastest growing engagement opportunity for brands to reach new and existing audiences," he said, adding that the point was meeting "users where they spend time today." Delta Vacations was named as the first advertiser testing the format. Its president, Katrin Koenig, described the appeal in terms of timing rather than reach: "Travel planning is becoming increasingly personalized, and travelers expect experiences that feel relevant."

Digiday reported the deal as the latest entry in an eighteen-month run of supply acquisitions. Netflix, Roku, Spotify, SiriusXM, Disney, Hulu and ESPN have all been wired into Amazon DSP over that period, and the pattern is consistent: Amazon does not build the audience, it buys the pipe to it and then undercuts on fees, frequently charging near zero on programmatic guaranteed deals against the 15 to 20% that The Trade Desk historically took. Spotify's global audio and video inventory arrived on the same logic, as did the preferred-partner integration with Microsoft Monetize. ChatGPT is the first conversational surface added to that list.

The commercial context is lopsided. Amazon's advertising segment passed $70bn on a trailing twelve-month basis in 2026. ChatGPT advertising crossed a $1bn annualised run rate in early September, with mobile ad density up 163% between April and August. Standalone chatbot advertising is projected to produce under $1bn in US revenue this year, rising to a little over $5bn by 2030. Against those numbers, OpenAI's reported internal target of $100bn in advertising revenue by 2030 would require compound annual growth above 200%. Nate Elliott, principal analyst at eMarketer, was blunt about the gap between ambition and machinery: "They're still trying to build out even many of the basics of a functional ad sales operation."

That is precisely why the Amazon deal matters more than its pilot status suggests. OpenAI does not need to build an ad sales operation if a company with $70bn of advertising revenue brings one. What neither party disclosed is the attribution infrastructure. The announcement contained no detail on how a conversational impression is measured, deduplicated against other Amazon DSP placements, or tied to a purchase. For advertisers whose Amazon spend is justified by retail conversion data, that omission is the whole question.

The second move of the week pointed in the same direction from the other end of the funnel. On September 8, Meta released Muse, a personal agent that completes purchases autonomously, reachable through WhatsApp and through dedicated apps on iOS, Android and muse.ai, US only for now. PPC Land documented the architecture: the agent runs on a dedicated cloud machine called Muse Secure VM with its own browser, file system and terminal, and a separate program named Sentinel approves outbound actions. Tarek Sheasha, a vice president at Meta Superintelligence Labs, authored the safety document; Mona Sarantakos and Christine Awad wrote the design document, which disclosed the first line of the system prompt: "Your purpose is to make your user's life better."

Checkout is where the engineering gets interesting. Muse generates a one-time-use payment card for each transaction through Stripe Link. Code analysis by Juozas Kaziukenas, published on LinkedIn three hours after launch, identified three distinct checkout modes: browser mode, which fills forms on any site; Shopify mode, which runs over APIs and most likely over the Universal Commerce Protocol Google introduced in January 2026; and Stripe Link mode for sites already integrated. Shop Pay and 1Password support are listed as coming.

The friction is on the merchant side, and it is not theoretical. Brent W Peterson put the problem plainly: an agent typing card details into a checkout form "looks like the spam bots they have been fighting for years." Merchants have no interface-level way to tell a legitimate buying agent from a scraper, which is why Kaziukenas concluded that browser-mode agents will simply be blocked and that WebMCP or backend interfaces become mandatory. Manil Uppal noted that UCP access on Shopify requires a double opt-in, Shopify approval plus merchant approval, which in practice reserves it for large operators. Originality.ai's May 2026 scan found only 26 public websites carrying UCP files out of three million checked, and that was after Amazon, Meta, Microsoft, Salesforce and Stripe joined the UCP Tech Council. The standard has backers. It does not yet have sites.

Meta's stated position on data carries the most commercial weight. The company says conversations and VM contents remain outside its advertising systems, which walls off purchase-intent signal from the business that generated $59.36bn in advertising revenue in the second quarter of 2026. The announcement also landed during the W3C and GS1 joint workshop on e-commerce for AI agents, held in Zurich on September 8 and 9, where standards bodies were still arguing about what an agent is permitted to do on a merchant's behalf. Two of the largest advertising businesses in the world shipped commercial agents into a specification that has not settled.

Forecasters raised their numbers, and credited the machines

The Interactive Advertising Bureau published a revised 2026 outlook on September 10 and lifted US ad spend growth to 12.3%, 2.8 points above the 9.5% it forecast in January. The revision drew on responses from more than 200 brand and agency decision-makers. Channel detail followed the same upward pattern: social media to 16.5%, a gain of 1.9 points; connected TV to 15.6%, up 1.8; commerce media to 13.6%, up 1.5.

IAB chief executive David Cohen attributed the improvement to the sporting calendar and to platform tooling, pointing at "strong spending around the Winter Olympics and soccer's World Cup" and at advertising platforms holding "increasingly powerful tools in their arsenal to find and engage customers." He also declined to call it a boom. "The first half was strong," he said, "but there are no easy wins."

Digiday paired the IAB revision with Madison and Wall's global read, which puts worldwide growth at 11%, or 9.8% excluding US political advertising, against a market past $1.3tn and a second quarter that grew 12.9%. The more consequential number in that piece is not the growth rate but the share. Roughly 12% of US ad spend is now directed by AI tooling, against 2% in 2023, and the projection runs to 27% of the market by 2030. Luke Stillman, managing director at Madison and Wall, described the movement as "a share shift" driven by convenience rather than by any demonstrated performance edge, and convenience is doing a great deal of work across advertisers of every size.

The supporting figures are specific. Google's Performance Max and AI Max together account for about 30% of search spend. Meta's Advantage+ is on track for $75bn in 2026, up from $60bn in 2025. Tinuiti's client data shows Performance Max campaigns absorbing 60 to 70% of retail client spending since the fourth quarter of 2025. These are not pilots. They are the default state of two of the three largest advertising businesses in the world.

Measurement is where the optimism stalls. The IAB survey found 86% of buyers expect to change how they measure media performance within twelve months because of conversational AI tools, and the most common intended change is measuring brand visibility inside AI platforms, a capability only 16% of brands currently hold in any systematic form. AdExchanger's Friday roundup surfaced the companion figure: 45% of advertisers name "comparing AI-driven customer journeys to traditional ones" as their single biggest measurement obstacle. There is also a counter-current worth noting. While 76% of buyers say they prioritise optimising content for AI discovery, only 69% report using generative AI in campaigns, down nine points since January.

Google's answer to the measurement gap arrived the day before. Meridian GeoX exited beta on September 9 during a livestream on the Google Analytics YouTube channel, alongside version 2.0.0 of the open-source mix modelling library released on September 2. Google claims budget savings of more than 31% for large advertisers against other open-source geo-experiment tooling, 94% faster experiment design, and runtime gains of two times speed and four times memory efficiency after switching the computation backend from TensorFlow, now deprecated, to JAX. A new automated calibration module applies four adjustments to translate experiment output into mix model inputs. A full-funnel framework uses branded Google query volume as a proxy for brand equity. An agentic skills repository provides model-building guidance inside the Antigravity CLI, and a chatbot interface now answers questions about model construction, data quality audits and setup errors.

Every one of those efficiency claims comes from Google's own testing, with no published methodology on sample sizes or competitor benchmarks, and the tool measures the effectiveness of media largely sold by the company that wrote it. That tension is not new to mix modelling. A Zalando researcher found MMM overstating paid search return by a factor of 2.5, which is precisely the direction of error a media seller has no incentive to correct. What is new is the packaging: Data Manager integration into Google Analytics and DV360, adoption of the IAB Tech Lab ECAPI standard, and a conversational front end, all assembled in the pre-GML measurement push that has been building since spring. Meta, by contrast, was reported in July to have deprioritised promotion of Robyn, its own open-source mix model. One vendor is investing in the measurement layer while its nearest rival steps back from it.

Nobody can agree what answer engine optimisation actually is

AdExchanger surveyed the field on September 10 and produced something unusual for a category this young: a list of tactics that demonstrably stopped working. Reddit manipulation is the clearest case. Citation share for the platform inside ChatGPT fell from 3.83% in the window of July 18 to August 7 down to 0.5% by mid-August 2026, a collapse steep enough to suggest the model detected the gaming rather than the market abandoning the tactic. A Reddit spokesperson expressed open scepticism about vendors promising guaranteed LLM performance through Reddit presence, and pointed to a crackdown on stealth marketing.

Automated publisher outreach fared no better. Noble, whose chief executive Rahul Jain spoke on the record, ran email campaigns using fabricated sender names and recorded response rates of 4 to 10%, with reputational exposure attached to every send. Shafqat Islam, president of Optimizely, argued for the unglamorous version instead, doing the established work ethically, on the grounds that manipulation neither retains customers nor builds trust.

What appears to work is more specific and less transferable than the vendor pitch decks suggest. James Cadwallader, chief executive and co-founder of Profound, described citation sets as unstable on a timescale of weeks: "citations that were informing a category will sort of drift and be replaced by new citations, new sources." Freshness therefore behaves less like a ranking signal and more like a condition of remaining in the sample at all. Industry structure matters too. Apparel brands benefit from structured question-and-answer formats on their own domains, while consumer finance brands depend heavily on third-party affiliate articles and listicles, which means an identical tactic produces opposite results across two verticals. Leah Nurik, chief executive and co-founder of Brandi AI, made the case for human authorship optimised for machine reading rather than machine authorship: "human-driven, human-written content that is optimized for AI is absolutely the way to go." Aniket Deosthali of Envive and Gintare Rimolaityte of Trendos contributed the formatting argument, that schematic content such as brand comparisons and structured lists outperforms unstructured prose.

Set that against what the search platform itself is now saying about position. On September 10, John Mueller told a Reddit thread about Search Console impressions that "the old 'position 1 - 10' is hard to map" onto a results page carrying AI Overviews with multiple citations, featured snippets, People Also Ask blocks and assorted specialised treatments. Google currently tracks generative AI results as a block, the way it treats many search features, rather than assigning individual positions. Mueller acknowledged the approach lacks granularity and invited suggestions on what positional data would be genuinely useful. The admission is notable: the company that defined the ranking metric is asking the industry to redefine it.

Practitioner consensus has not caught up. Cyrus Shepard and Dawn Shepard published their expert ranking factors survey on Zyppy on September 10, drawing more than 100 respondents and 13,665 ranking-factor data points, and the top of the table looks much as it did five years ago: content relevance at 57.1%, backlinks at 54.8% and content quality at 47.6%, followed by trust and authority at 36.5%, behavioural click signals at 29.4% and brand presence at 27%. Technical SEO, topical authority and internal linking all sat below 20%. The survey measures belief rather than mechanism, and belief is lagging the interface. Earlier work on the citation question found that brands named inside a user's own query win mentions 33 times more often, a finding that sits awkwardly alongside a factor list headed by backlinks, and separate analysis has catalogued 23 factors associated with being cited by AI search engines. Commercial pressure is not waiting for the theory to settle: HubSpot launched an AEO product after organic traffic fell 27% across its customer base, and survey work indicates half of US shoppers fact-check AI answers on Reddit before buying, which is the behavioural loop the citation-share collapse just disrupted.

Verification moves upstream, and the exchange grades its own homework

Two announcements dated September 9 pushed ad verification into places it has not normally sat, and they pull in opposite directions on the question of independence.

BidSwitch opened free bidstream quality analysis from Adalytics to its supply-side partners. The service is opt-in and sits above the malware and invalid-traffic filters already running on the platform. Adalytics identifies suspicious trading patterns; BidSwitch then independently reviews and adjudicates those findings before passing recommendations to partners who have opted in. Barry Adams, general manager at BidSwitch, described it as "an opt-in layer of analysis that complements the malware and IVT protections already available on the BidSwitch platform, and we're proud to offer it at no cost." Krzysztof Franaszek of Adalytics framed the exchange of data and research as an effort to "strengthen transparency, accountability, and media quality across the digital advertising ecosystem."

The structural problem is visible from the sentence describing the workflow. BidSwitch, owned by Criteo, earns on volume routed through it, and is now the party that adjudicates quality findings about that same volume. This is the objection levelled at platform-graded brand safety for a decade, reproduced one layer down in the bidstream itself. The announcement disclosed no metrics, no thresholds, no consequences for supply that fails, no appeal process, no availability date and no indication whether demand-side platforms receive equivalent analysis. Adalytics carries genuine credibility from prior independent research, some of which preceded DoubleVerify's legal difficulties. Whether that credibility survives adjudication by the audited party is the open question.

The second announcement placed measurement after the money rather than before it. Inuvo, the Little Rock ad tech company listed on NYSE American as INUV, integrated FouAnalytics into IntentKey, its audience modelling platform. The mechanism is post-bid verification rather than pre-bid filtering: it measures which served impressions actually reached human audiences after purchase, then feeds the result back into inventory scoring. IntentKey models refresh every five minutes and activate through client DSPs or Inuvo's managed service. Eric Tilbury, Inuvo's vice president of programmatic operations and solutions engineering, stated the commercial driver: "Advertisers are under pressure to prove that media dollars are not just being spent, but are working." Augustine Fou was more pointed about the sequencing: "Independent post-bid measurement is the only way to see where ads actually ran."

The gaps are the familiar ones. The disclosure, issued August 26 via GLOBE NEWSWIRE, contained no baseline human-reach percentage, no campaign volumes, no client counts, no commercial terms, no statement on whether clients see the measurement results, and no definition of whether failing supply is excluded, downweighted or merely flagged. Timing supplies the context the announcement omits. It landed 20 days after Nielsen agreed to acquire DoubleVerify for $2.15bnon August 6, and hours after FouAnalytics announced unlimited verification at $2m a year. A market whose largest independent verifier is being absorbed by a measurement company is a market in which smaller buyers start shopping for alternatives, and vendors price accordingly.

PPC Land published five explainers on September 10 covering exactly this vocabulary, including post-bid verification, pre-bid filtering, bidstream and traffic shaping. The publication of a glossary is itself a signal about where the argument has moved.

Google changed the terms on both sides of its marketplace

On the buy side, David Melamed, a PPC consultant, posted on LinkedIn that Google Ads had issued promotional credits to client accounts, then invalidated them more than a month after the credited amounts had been spent. Search Engine Roundtable carried the account on September 10. Credits of $3,200 were involved in at least two accounts. Melamed said it had happened "twice, over a short period of time" and described the pattern as pushing "a really 'generous' credit to get advertisers to spend money... only to invalidate it after they spent thousands of dollars and waited more than a month." He called it "pure theft" and noted the asymmetry that makes it consequential: "you obviously can't crawl back your spend once it's gone." Ginny Marvin, the Google Ads liaison, replied that she had passed the report to the team. No explanation or resolution was published.

The mechanics matter. A promotional credit changes spending behaviour by design; that is its purpose. An advertiser spends against a budget that includes the credit, the impressions are served, the auction clears, and the money moves to Google. Reversing the credit afterwards leaves the advertiser paying full price for volume they would not have bought at full price. Google's billing surface has drawn this kind of scrutiny before, including mass arbitration alleging billions in overcharges and the belated documentation of the Invalid Activity Credit Report. What distinguishes this case is that the credit was promotional rather than remedial, which places it outside any invalid-traffic reconciliation process.

On the sell side, the change is automatic and dated. On September 9 Google notified AdSense publishers of an account-level control named "Maximize message coverage" that activates on or after October 9, 2026. It replaces the existing "Create a consent message for your sites" setting and does two things: it deploys Google's own consent management platform as a backstop, serving a banner whenever an ad request from a regulated region arrives without a TC string, and it automatically generates and publishes consent messages for any new site added to the account after activation. The notice put the rationale in revenue terms, saying the control is "Enabled in 30 days to help you reduce traffic ineligible for personalized ads."

Coverage is broad. The feature touches every property using the adsbygoogle tag, Google Publisher Tag or Google Mobile Ads SDK across AdSense, Ad Manager and AdMob. Auto-enrolment captures publishers who previously selected a Google CMP message option and publishers who never made a selection at all; only accounts that specifically chose an alternative CMP certified by Google are excluded. During the 30-day window the toggle displays as on without affecting delivery. Opting out requires navigating to Privacy and messaging, then European regulations settings, then switching off Maximize message coverage, and that can be done before, during or after the window.

The liability allocation is stated in Google's own documentation: "You are responsible for ensuring the consent banner works for your intended purposes." A publisher who does nothing receives a banner they did not configure, on a schedule they did not set, and remains the controller accountable under GDPR for whether it is lawful. That arrangement follows a long sequence of consent mandates flowing downstream, including the TCF v2.3 migration deadline of February 2026. Taken together with the credit reversals, the week produced two instances of the same shape: the platform adjusted terms unilaterally, and the counterparty absorbed the consequence.

One smaller change ran the other way. Google began a beta in the experiments area of Google Ads letting advertisers test different budget, target CPA or target ROAS amounts on Search campaigns to find an optimal spending level, spotted by Thomas Eccel on LinkedIn and reported on September 10. The rollout also includes AI Max experiments with brand and location controls, and Performance Planner recommendations that can be applied directly to campaigns. It is a genuine addition of testable ground in a product that has been removing it.

Also noted

  • September 10: Garrett McGrath, Prebid's former chairman, becomes only its second president, starting September 15 and overlapping with outgoing president Mike Racic through October 15; he had stepped down as chairman in June after leaving Magnite, and faces a 90-day listening period before major decisions, saying that "if advertising pays for the internet, then Prebid is critical infrastructure".
  • September 10: Daily Mail US launches New Media, a creator-led social video division built on its UK model, which now employs more than 100 staff and averages around 300,000 views per video, aiming to lift social video from roughly 20% of direct advertising revenue to more than 33% by 2027.
  • September 11: Molson Coors reports engagement with creator content quadrupling after a March 2026 overhaul with Movers+Shakers' Shake Squad, rolled out across 230 marketers and more than 100 brands in the US and Canada, sorting legal review into fast-track, needs-discussion and hard-no lanes.
  • September 9: Apple vice president of AI product Lilian Rincon told the company's keynote that users make "over 2.5 billion requests every single day" of Siri, a figure disclosed while introducing an updated version of the assistant.
  • September 10: Scammers are uploading AI-generated photos containing substituted phone numbers to Google Business Profile listings through the Maps contributor program, with one account operating as "Jack Edward" posting more than 150 images across unrelated US businesses, all carrying the same number.