What’s new
What changed in growth marketing between early 2025 and October 2026, and where it touches this workshop
This page lists what changed recently and leaves out what the workshop already teaches. Each item says what changed, when, where we read it (Verified (date) means we opened that page on that date; Not read means we could not and says what we relied on instead) and which module it touches. Figures a company reports about itself are marked as the company’s own. Last reviewed 2026-10-16; the references page explains how we check sources.
Measurement
Google’s Meridian is open to everyone. Google released Meridian, its open-source Bayesian marketing mix model, to all advertisers on 2025-01-29, after a limited-access period. It works on geo-level data, models reach and frequency rather than impressions alone, and takes the results of incrementality experiments as priors, which is the calibration idea of Module 9. Releases have continued; version 2.1.0 reached PyPI on 2026-09-25 (Apache-2.0). Verified (2026-10-16): Google’s announcement (“Meridian accounts for reach and frequency, not just impressions”; “easily integrates incrementality experiment results as priors”) and the repository. In this workshop: Modules 8 to 10 and 17 use PyMC-Marketing; Meridian is the second open-source Bayesian MMM a team can compare it with. It is not used in a lab because it conflicts with PyMC-Marketing’s pins and prefers a GPU (decisions).
Chrome keeps third-party cookies, and most of the Privacy Sandbox is gone. On 2025-04-22 Google dropped its plan for a standalone cookie prompt in Chrome and kept the existing settings-based choice. On 2025-10-17 it retired most Privacy Sandbox technologies, citing low adoption, among them Topics, Protected Audience and the Attribution Reporting API; CHIPS, FedCM and Private State Tokens stay. Cookie-based tracking in Chrome therefore lasts longer than teams planned for in 2024, and the sandbox’s privacy-preserving attribution path is no longer coming. Verified (2026-10-16): Next steps for Privacy Sandbox (“will not be rolling out a new standalone prompt for third-party cookies”) and Update on plans for Privacy Sandbox technologies (“in light of their low levels of adoption, we’ve decided to retire the following Privacy Sandbox technologies”). In this workshop: Modules 6, 7 and 16: user-level paths remain available in Chrome for now, and Module 16 shows why they still do not measure incrementality.
The platforms now sell modeled incrementality. Meta’s “Incremental Attribution” setting steers delivery toward conversions its models judge incremental, trained on its own conversion-lift studies; on 2026-01-28 Meta said its Q4 2025 model drove 24% more incremental conversions than standard attribution and that the product reached a multi-billion-dollar annual run rate seven months after launch (Meta’s own figures). Google’s user-based Conversion Lift help page now lists Search and Performance Max campaigns alongside Display, Video, Demand Gen and App, with a floor of 1,000 observed conversions and a $5,000 minimum budget, and still says to contact a Google representative. Verified (2026-10-16): Meta newsroom, 2026-01-28 (“a 24% increase in incremental conversions compared to our standard attribution model”); Google Ads help: set up Conversion Lift based on users (“at least 1,000 observed conversions”; “To use Conversion Lift, contact your Google account representative”). Not read: Meta’s help page on Incremental Attribution (the page did not open), so the launch month is not verified. In this workshop: Modules 6, 7 and 9. A platform’s lift number is an experiment the platform ran on its own objective; the geo tests of Module 7 are the independent check.
Marketers say they will spend more on MMM. A TransUnion and EMARKETER survey of 196 US marketing professionals (fielded July 2025, published 2025-10-21) found 47% planning to increase spend on marketing mix modeling in the next year and confidence in measurement flat. It is a vendor survey with a paywalled methods chart; take it as directional. Verified (2026-10-16): TransUnion press release. In this workshop: Day 4.
Platforms automate the campaign
Meta’s ad delivery runs on retrieval and ranking models that learn from everything. Meta rebuilt ad retrieval (Andromeda, announced 2024-12-02) and ranking around one large foundation model, GEM, built “on an LLM-inspired paradigm”. On its Q3 2025 earnings call Meta said combining retrieval and early-stage ranking raised ads quality 14% on Facebook surfaces, that revenue through its end-to-end automated Advantage+ products reached a $60 billion annual run rate, and that advertisers using its video-generation tools rose 20% quarter on quarter; its 2026-10-13 engineering post claims an 18.9% conversion improvement from GEM since. All figures are Meta’s own. The practical change for an advertiser is where the levers are: targeting is automated, so the inputs that remain are creative variety, the conversion signal fed back, and independent measurement. Verified (2026-10-16): Andromeda, GEM, GEM in 2026 and the Q3 2025 earnings call transcript. No research paper on GEM was found; do not cite one. In this workshop: Modules 8, 9 and 11: when the platform optimizes delivery, the ROAS it reports is a model of its own targeting, and a lift test or an MMM calibrated by one is what tells you what the spend caused.
New channels: agents and answer engines
AI agents can now shop and pay. Three open protocols appeared within four months. Google’s Agent Payments Protocol (AP2, 2025-09-16) carries cryptographically signed “mandates” that record what a user authorized. OpenAI and Stripe’s Agentic Commerce Protocol (ACP, 2025-09-29, Apache-2.0) powered Instant Checkout in ChatGPT, first with Etsy sellers; on 2026-03-24 OpenAI pulled Instant Checkout back as a standalone feature, letting merchants use their own checkout while ChatGPT focuses on discovery, and the protocol continues. Google’s Universal Commerce Protocol (UCP, 2026-01-11), built with Shopify, Etsy, Wayfair, Target and Walmart, powers checkout in AI Mode and Gemini. Shopify’s Agentic Storefronts (2025-12-10) sell through ChatGPT, Perplexity and Microsoft Copilot and report orders with the originating AI channel; Mastercard (Agent Pay, 2025-04-29) and Visa (Intelligent Commerce, 2025-04-30) issue agent-specific tokenized credentials with user-set limits. Verified (2026-10-16): Stripe newsroom, Stripe on ACP, Google Cloud on AP2, Google on UCP, Shopify, 2025-12-10 and 2026-03-24, Mastercard, Visa. Not read: OpenAI’s own announcement pages (they refused the request); we relied on Stripe as co-announcer and on TechCrunch, 2026-03-24 quoting OpenAI for the pull-back. In this workshop: Modules 1 to 5 and 16. An agent-referred order is a new acquisition channel with its own repeat behavior, and a “visit” that is an agent call breaks the journey data attribution relies on. Treat it as a channel to measure with Modules 2 to 5, not as a touch.
Answer engines change what “organic” means. Google rolled AI Mode out to all US users on 2025-05-20, put ads in AI Overviews on desktop and in AI Mode from 2025-05-21, and by 2026-06-03 reported AI Overviews above 2.5 billion monthly users and AI Mode past 1 billion. Pew’s panel study of March 2025 browsing (900 US adults) found users clicked a traditional result on 8% of visits that showed an AI summary against 15% without one, clicked a source inside the summary on 1% of visits, and ended the session more often (26% against 16%); Pew reports the association, not a cause. Adobe’s own analytics put AI-referred traffic to US retail sites up 693% year on year over the 2025 holiday season (vendor data). Verified (2026-10-16): Google, 2025-05-20, Google Marketing Live 2025, Alphabet investor presentation, June 2026, Pew Research, 2025-07-22, Adobe, 2026-01-07. In this workshop: Modules 8 and 17. The organic baseline of a marketing mix model fitted across 2025 has a structural break in it; Module 17’s residual check is the tool that finds it.
Retail media keeps compounding. Amazon’s advertising services revenue was $17.7 billion in Q3 2025, up 24% on the year, from its own filing. Retail media networks sell closed-loop measurement on first-party purchase data: convenient, but attribution inside a walled garden, not incrementality. Verified (2026-10-16): Amazon Q3 2025 results, SEC exhibit 99.1. In this workshop: Modules 8, 10 and 16.
Notifications under the operating system’s control
The phone can rewrite or silence your push. iOS 18.1 (October 2024) introduced Apple Intelligence notification summaries, which condense a stack of notifications into one line. In iOS 18.3 (January 2025) Apple made summaries “temporarily unavailable” for News and Entertainment apps after misleading summaries of news alerts, set summarized notifications in italics, and added Lock Screen controls to turn summaries off per app; the press reported their return for news apps with a warning in the iOS 26 betas (2025-07-22). On Android 13 and later an app’s notifications are off until it wins the POST_NOTIFICATIONS runtime permission. The copy a growth team writes may be rewritten before the user sees it, and the user can mute the app in two taps. Verified (2026-10-16): About iOS 18 updates (“Notification summaries for News & Entertainment apps are temporarily unavailable”; “Easily manage settings for notification summaries from the Lock Screen”), Android notification runtime permission (“your app’s notifications are off by default”). Not read: an Apple page confirming the iOS 26 return; 9to5Mac, 2025-07-22 is the evidence. A widely reported Android “notification cooldown” is not on the Android 15 or 16 developer pages we read. In this workshop: Module 18: the opt-out is the binding constraint of the volume trade-off, and the operating system makes it cheaper every year.
Methods and tools
Sequence models for notification policy. LinkedIn researchers reframed the send decision as sequence modeling with a Decision Transformer instead of value-based offline reinforcement learning, with a multi-reward design for a task that never ends and quantile regression for the return target; in production it gave +0.72% sessions over a multi-objective conservative-Q-learning agent. A preprint, not peer reviewed. Verified (2026-10-16): arXiv 2509.02458 (submitted 2025-09-02). In this workshop: Module 18 fits the value-iteration policy that this line of work starts from, and its briefing lists the LinkedIn papers it builds on.
Digital twins of respondents. Toubia and colleagues at Columbia released Twin-2K-500, a public dataset of 2,058 US respondents answering more than 500 questions over four waves, to test whether large-language-model “digital twins” can stand in for people in pretests; the authors report that twins capture relative differences between people but predict individuals poorly. Verified (2026-10-16): arXiv 2505.17479 and the authors’ project page (the journal record in Marketing Science could not be opened). In this workshop: Module 6. A simulated respondent can order message variants before a test; it does not replace the test.
Open-source tooling moved under us. PyMC-Marketing reached 1.0.0 on 2026-08-07 (“After two years of 0.x releases, this is our stable API”), made the multidimensional MMM the default, and in 1.1.0 and 1.2.0 (2026-09-29, the version this workshop pins) restructured the budget optimizer and added a lift-test case study. CausalPy shipped Bayesian synthetic difference-in-differences in 0.9.0 (2026-07-28) and a 1.0.0 release candidate with geo-lift analysis for MMM calibration (2026-10-06). Meta’s Robyn has had no commit on its main branch since 2025-06-27; the repository is not archived and carries no deprecation notice, and reports that Meta is winding it down rest on anonymous sources. Verified (2026-10-16): PyMC-Marketing releases, CausalPy releases, Robyn. In this workshop: Modules 7 to 10 and 17; the package pins are in _variables.yml and the versions page.
Rules
The EU AI Act applies, with its high-risk rules delayed. Prohibitions and AI-literacy duties applied from 2025-02-02, obligations for general-purpose AI models from 2025-08-02, and the Act became generally applicable on 2026-08-02; the “Digital Omnibus on AI”, in force since 2026-07-27, moved the high-risk obligations for Annex III use cases to 2027-12-02 and for product-embedded systems to 2028-08-02. In the United States, 19 states had comprehensive privacy laws by 2025-10-27, several giving consumers a right to opt out of profiling for automated decisions. France’s competition authority fined Apple €150 million on 2025-03-31 over how App Tracking Transparency was implemented, without ordering changes to it. Verified (2026-10-16): European Commission, regulatory framework for AI (updated 2026-08-03), IAPP state privacy overview (“each of the 19 states that have passed comprehensive privacy laws to date”), Autorité de la concurrence. Not read: the Official Journal text of the omnibus regulation and the per-state effective dates (member-gated pages). In this workshop: Modules 11, 15 and 18, which make automated decisions about individuals: whom to discount, whom to credit, whom to notify. Keep the opt-out and the explanation of the rule as part of the deliverable.
What we checked and left out
Claims we could not trace to a primary or reputable source are not on this page: that Meta will offer fully AI-generated ads by the end of 2026; a Meta research paper on GEM; an Android “notification cooldown”; Shopify’s “11x” growth in AI-attributed orders; an incrementality-adoption percentage from the EMARKETER survey (paywalled); and that Meta is winding down Robyn. If you have a primary source for any of them, open an issue.