Which marketing decisions will you let software make without asking you first? Six marketing technology and AI releases carried September 22, 2026 datelines, and four of them hand that question straight to the buyer. AdRoll, the AI-powered multichannel advertising platform, reported in its Q3 2026 State of Digital Advertising Report that display prospecting cost per mille (CPM) fell 45% year over year between July 1 and September 8, 2026. Aurora Mobile (NASDAQ: JG) added a decision model to its GPTBots.ai agent platform that lets each workflow set its own confidence threshold for acting, sending work to review or escalating to a person. The Kroger Co. (NYSE: KR) described its AI program in three stages, assisted, augmented and autonomous, at Groceryshop 2026 in Las Vegas. PulsePoint, Fullspan Health and Semarchy shipped audience-building, AI search measurement and governed-data products the same day.
AdRoll‘s figures describe prices on AdRoll’s own platform. AdRoll reported that display retargeting CPMs fell 29.1% year over year in the same July 1 to September 8, 2026 window, that prospecting and retargeting CPMs fell 25.5% and 40.2% against the same window in Q2 2026, and that account-based marketing CPMs rose 4.4% year over year. The direction runs against Meta. Meta Platforms (NASDAQ: META) reported in its Form 10-Q filings that average price per ad rose 12% year over year in both Q1 and Q2 2026, which The Agile Brand Guide covered in its September 22, 2026 edition. A team that leaned on social in 2026 paid more per impression during the same months AdRoll’s display prices fell.
AdRoll chief executive officer Vibhor Kapoor advised marketers to compare the cost of reaching each audience with the engagement, conversions and revenue that follow. Put that advice on a hypothetical account. A home goods retailer spent $60,000 on display prospecting between July and early September 2025 at a $5.00 CPM and bought 12 million impressions. At a CPM 45% lower, $2.75, the same $60,000 buys about 21.8 million impressions in 2026. Cost per acquisition equals CPM divided by conversions per thousand impressions, so the retailer’s CPA improves only while conversions per impression fall by less than 45%. If the cheaper inventory converts at half the old rate, CPA rises about 10%, and the retailer buys 9.8 million more impressions and about 9% fewer conversions for the same $60,000. AdRoll published no conversion or inventory quality data next to its price data, so the retailer measures that ratio itself, weekly, before November traffic arrives.
Aurora Mobile’s September 22, 2026 release describes the mechanics behind the opening question. Aurora Mobile stated that GPTBots.ai now splits work between a reasoning layer, which uses large language models such as GPT and Claude, and a decision layer powered by Jev, a model TypeSafe AI released on September 15, 2026. Every Jev decision returns a probability score, and Aurora Mobile stated that enterprises set thresholds per workflow to decide which decisions execute automatically, which go to review and which escalate to a human agent. Aurora Mobile cited TypeSafe AI benchmarks of 70 to 500 milliseconds per decision against 3 to 329 seconds for frontier LLMs, and a price of $0.042 per million input tokens with output free.
Aurora Mobile’s price figures cover the model call. The larger cost sits with the threshold, because a person on your team picks it and another person staffs the review queue it creates. That queue is where human-in-the-loop (HITL) stops being a box on an architecture slide and becomes a headcount line. I’ve watched teams spend six months selecting an agent platform and one afternoon choosing its thresholds.
Run the arithmetic on a hypothetical B2B marketing operations team that routes 40,000 inbound form fills a month through an agent. The team sets automatic routing at 0.90 confidence, review between 0.70 and 0.90, and escalation below 0.70. If 70% of submissions clear 0.90, the agent routes 28,000 on its own, 8,000 land in review and 4,000 go to a person. At 90 seconds per review, the 8,000 reviews cost 200 staff hours a month, roughly one and a quarter full-time roles. Lower the automatic threshold to 0.80 and the review queue shrinks, while the misroute rate on the newly automated share rises by an amount the team learns only by sampling the output. The model’s price moves neither number. The threshold and the review staffing are marketing decisions, and they belong in the budget next to the license.
The cost-of-wrong for each decision sets the right threshold. A misrouted lead costs a follow-up email. A wrongly published audience in a regulated category costs a compliance review, and possibly more. Kroger described the same sorting at enterprise scale. Yael Cosset, Kroger’s executive vice president and chief digital officer, said at Groceryshop 2026 on September 22, 2026 that Kroger places each AI use case at the assisted, augmented or autonomous stage its risk, repeatability and need for human judgment support. Kroger stated that its AI Shopping Assistant already plans meals and builds carts around budget and dietary needs, and that it is making AI literacy training and tools available to more than 400,000 associates. Kroger published no outcome figures. A marketing leader can copy the sorting rule without Kroger’s loyalty data: write down, per workflow, which stage it sits in and why.
PulsePoint, the health marketing technology division of Internet Brands, supplied the high end of that cost-of-wrong list on September 22, 2026. PulsePoint launched DTC Explorer, a self-service workspace where pharma marketers build consumer audiences from de-identified diagnosis, drug and treatment claims, filtered by clinical signals, prescribed drugs, insurance provider, geography or prescriber NPI list. PulsePoint stated that marketers will be able to reach the same functions through Hatch, its agentic platform, by describing an audience and letting Hatch construct, refine and publish it. The verb to watch in that sentence is publish. A pharma brand team should decide in writing whether an agent publishes a claims-based audience or drafts one for a person who signs off, and should make that decision before Hatch access arrives.
Semarchy, the Phoenix master data management company, addressed the input every threshold depends on. Semarchy announced on September 22, 2026 that every Intelligent Data Product in its platform automatically creates its own Model Context Protocol (MCP) server, so AI assistants and agents can query mastered data along with its business definitions, quality rules, sensitive-data controls, access rights, lineage and certification history. A confidence score on a bad record is a confident mistake. Semarchy sells a way to keep the governance attached to the record, and somebody still writes the quality rules. Decision rights over those rules, meaning who can change a customer definition that a dozen agents read, need a named owner before the MCP endpoint goes live.
Fullspan Health, the B2B healthcare platform from RVO Health, published the edition’s other headline number. Fullspan Health reported on September 22, 2026 that across an average of 200 prompts tested between July and September 2026, its own properties, including Healthline, Healthgrades and Medical News Today, appeared in up to 81% of AI citations for key condition questions and as much as 3.6 times more often than competitors across ChatGPT, Claude and Gemini. Fullspan Health sells the media that runs on those properties, measured the citations with AIQ, its own new tool, and defined category reach as the share of questions for which at least one tested model cited a Fullspan property. Under that definition, one citation from any one of five models counts a question as reached. A pharma brand evaluating Answer Engine Optimization (AEO) spend through AIQ can ask for the per-model citation rate on its own brand’s condition questions, using prompts the brand wrote.
Sort the six September 22, 2026 releases and two groups appear. AdRoll and Fullspan Health published numbers that feed a decision. Aurora Mobile, Kroger, PulsePoint and Semarchy published the machinery that increasingly makes the decision. Every number in the set came from the company selling the product it measured.
Check who measured each number before you act on it
AdRoll’s CPM declines describe prices paid on AdRoll’s platform between July 1 and September 8, 2026, and AdRoll disclosed no spend volume, advertiser count or inventory quality measure. Fullspan Health’s 81% and 3.6-times figures measure Fullspan-owned properties with Fullspan’s own tool, and both carry upper-bound wording (“up to” and “as much as”). Aurora Mobile’s latency and price figures are TypeSafe AI benchmarks relayed by Aurora Mobile. A decision model that always returns a valid answer format can still return the wrong choice with a high probability score, so any threshold needs calibration against sampled outcomes. Kroger, PulsePoint and Semarchy published no outcome figures, and PulsePoint described Hatch access to DTC Explorer as a future capability.
None of the six companies measured incremental effect. A holdout test your team runs is the method that shows whether cheaper impressions or an automated decision changed revenue.
Strategic priorities for marketing leaders
- Write a decision register before you switch on another agent. List each marketing decision software makes for you, its stage (assisted, augmented or autonomous), its confidence threshold and the named owner who reviews exceptions.
- Budget the review queue in the same line as the license. Model review hours from expected threshold pass rates, as in the 40,000-submission example, before procurement signs.
- Judge cheaper prospecting inventory by cost per acquisition. If display prospecting CPMs on your platforms fell anywhere near AdRoll’s reported 45%, move a test budget before Black Friday and track conversions per thousand impressions weekly against a holdout.
- Keep a human signature on regulated audience publication. Health, financial and other regulated audiences go to a person for approval at any probability score an agent reports.
- Ask every AI visibility vendor for per-model, per-brand citation rates on prompts you supply. A per-brand, per-model rate is the number your budget decision requires.
Back to the opening question. Kroger answered it with three stages, Aurora Mobile with a probability score, and PulsePoint, for now, with a person clicking publish. A Chief Marketing Officer (CMO) without a written answer inherits a vendor’s default setting. Open a document this week, list the ten marketing decisions software already makes on your behalf, and write a threshold and an owner next to each one.
Here’s The News for September 22, 2026
AdRoll Reports 45% Year-Over-Year Drop in Display Prospecting CPMs in Q3 2026 Report
AdRoll, an AI-powered multichannel advertising platform, released the Q3 2026 edition of its State of Digital Advertising Report on September 22, 2026, reporting that display prospecting CPMs fell 45% year over year from July 1 through September 8, 2026, and that display retargeting CPMs fell 29.1% over the same period. Against the same window in Q2 2026, AdRoll reported prospecting CPMs down 25.5% and retargeting CPMs down 40.2%, while account-based marketing (ABM) CPMs rose 4.4% year over year and fell 15.1% from Q2. AdRoll paired the pricing data with macroeconomic figures it described as not directly linked, including US inflation holding at 3.4% in August 2026 and the University of Michigan consumer sentiment index falling 13.2% year over year. AdRoll chief executive officer Vibhor Kapoor said lower CPMs give marketers room to test before holiday competition intensifies, and advised comparing the cost of reaching each audience with the engagement, conversions and revenue that follow.
Read the full press release (Published: September 22, 2026 | Source: GlobeNewswire)
Aurora Mobile’s GPTBots.ai Adds Jev Decision Model With Per-Workflow Confidence Thresholds
Aurora Mobile Limited (NASDAQ: JG), a provider of customer engagement and marketing technology services, announced on September 22, 2026 that its enterprise AI agent platform GPTBots.ai integrated Jev, a decision model TypeSafe AI released on September 15, 2026. Aurora Mobile stated that the integration creates a two-layer architecture in which Jev handles routing, filtering, classification and relevance scoring, and large language models such as GPT and Claude handle reasoning, text generation and open-ended dialogue. Aurora Mobile stated that every Jev decision carries a probability score and that enterprises can set thresholds per workflow that determine which decisions execute automatically, which require review and which escalate to human agents. Aurora Mobile cited TypeSafe AI benchmarks of 70 to 500 milliseconds per decision and a price of $0.042 per million input tokens with free output, and GPTBots.ai founder and CEO Chris Lo said the platform now pays a fraction of a cent per judgment where it previously paid for a full LLM call on each routing decision.
Read the full press release (Published: September 22, 2026 | Source: GlobeNewswire)
Kroger Outlines Three-Stage AI Strategy at Groceryshop 2026
The Kroger Co. (NYSE: KR) outlined its artificial intelligence strategy at Groceryshop 2026 in Las Vegas on September 22, 2026. Yael Cosset, Kroger’s executive vice president and chief digital officer, described three stages of AI use, assisted, augmented and autonomous, and said Kroger places each use case at the stage its risk, repeatability and need for human judgment support, with governance at the center. Kroger stated that its AI Shopping Assistant helps customers plan meals and build carts around budget and dietary needs, that it is working on agents that filter daily notifications and tasks for frontline teams, and that it is making AI literacy training and tools available to more than 400,000 associates. Cosset described a future in which customers share goals and preferences and AI agents handle planning, shopping and fulfillment as a managed service.
Read the full press release (Published: September 22, 2026 | Source: PR Newswire)
PulsePoint Launches DTC Explorer for Self-Service Pharma Audience Building
PulsePoint, a health marketing technology company and division of Internet Brands, launched DTC Explorer on September 22, 2026, a self-service workspace that lets pharma marketers build, optimize and publish clinically informed consumer Bespoke Audiences in hours. PulsePoint stated that Bespoke Audiences are built from de-identified diagnosis, drug and treatment claims, and that DTC Explorer users can filter by consumer attributes and demographics, clinical signals, prescribed drugs, insurance provider, geography or NPI list while reviewing real-time aggregate device and consumer counts. PulsePoint stated that marketers will be able to reach DTC Explorer’s functions through Hatch, its proprietary agentic platform, where a user describes the audience and Hatch constructs, refines and publishes it.
Read the full press release (Published: September 22, 2026 | Source: PR Newswire)
Fullspan Health Launches AIQ to Measure Health Brand Visibility in AI Search
Fullspan Health, the B2B healthcare platform from RVO Health, launched AIQ on September 22, 2026, an AI measurement and media optimization product that shows health brands how they appear, get cited and earn links in AI-generated answers. Fullspan Health reported that its content appeared in up to 81% of AI citations for key condition questions and as much as 3.6 times more often than competitors across ChatGPT, Claude and Gemini, based on an average of 200 prompts tested across conditions and therapeutic areas from July through September 2026. Fullspan Health defined category reach as the percentage of unique questions for which at least one tested model, including Google AI Overviews, Claude, ChatGPT, Gemini and Perplexity, cited a Fullspan Health property. AIQ is available across Fullspan Health advertising products, which run on properties including Healthline, Healthgrades, Medical News Today, Psych Central and Bezzy.
Read the full press release (Published: September 22, 2026 | Source: PR Newswire)
Semarchy Adds Automatic MCP Servers to Intelligent Data Products
Semarchy, a Phoenix-based master data management company, announced on September 22, 2026 expanded AI capabilities for Intelligent Data Products, part of the Semarchy Data Platform. Semarchy stated that every Intelligent Data Product automatically creates its own Model Context Protocol server, letting users and AI agents query mastered data along with its business purpose, entities, relationships, quality rules and sensitive-data controls, and that quality, access rights, lineage and certification history travel with each data product. Semarchy stated that its AI Data Engineering capabilities generate schemas, rules and models for human review, and chief technology officer Craig Gravina said that designing a data product amounts to building the context layer enterprises need for AI.
Read the full press release (Published: September 22, 2026 | Source: Business Wire)
Frequently asked questions
What marketing technology news was announced on September 22, 2026?
Six marketing technology and AI releases carried September 22, 2026 datelines. AdRoll reported a 45% year-over-year drop in display prospecting CPMs, Aurora Mobile added per-workflow confidence thresholds to its GPTBots.ai agent platform, Kroger outlined a three-stage AI strategy at Groceryshop 2026, PulsePoint launched DTC Explorer for pharma audience building, Fullspan Health launched the AIQ AI search measurement product, and Semarchy added automatic MCP servers to its Intelligent Data Products.
How much did display advertising CPMs fall in Q3 2026?
AdRoll reported in its Q3 2026 State of Digital Advertising Report, released September 22, 2026, that display prospecting CPMs on its platform fell 45% year over year from July 1 through September 8, 2026, and that display retargeting CPMs fell 29.1%. Against the same window in Q2 2026, AdRoll reported prospecting CPMs down 25.5% and retargeting CPMs down 40.2%. AdRoll published no conversion or inventory quality data alongside the price figures.
What did Aurora Mobile announce for GPTBots.ai on September 22, 2026?
Aurora Mobile (NASDAQ: JG) announced on September 22, 2026 that GPTBots.ai integrated Jev, a decision model from TypeSafe AI, to separate routing and classification decisions from large language model reasoning. Each Jev decision returns a probability score, and enterprises can set per-workflow thresholds for automatic execution, review or escalation to a human. Aurora Mobile cited TypeSafe AI benchmarks of 70 to 500 milliseconds per decision and $0.042 per million input tokens.
How is Kroger using AI, according to its Groceryshop 2026 presentation?
Kroger (NYSE: KR) stated on September 22, 2026 that it places AI use cases in assisted, augmented or autonomous stages based on risk, repeatability and the need for human judgment. Kroger’s AI Shopping Assistant plans meals and builds carts around budget and dietary needs, and Kroger is making AI literacy training available to more than 400,000 associates. Kroger published no outcome figures for these programs.
What does Fullspan Health’s AIQ measure?
Fullspan Health launched AIQ on September 22, 2026 to measure how health brands appear, get cited and earn links in AI-generated answers. Fullspan Health reported that its own properties appeared in up to 81% of AI citations for key condition questions, based on an average of 200 prompts tested from July through September 2026 across models including ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews. Fullspan Health counts a question as reached when at least one tested model cites a Fullspan Health property.
What should marketing leaders do with the September 22, 2026 releases?
Marketing leaders can list each marketing decision software makes for them, assign it an autonomy stage, a confidence threshold and a named reviewer, and budget review staffing alongside agent licenses, following the threshold model Aurora Mobile described on September 22, 2026. Teams can also test cheaper display prospecting inventory on cost per acquisition before the holiday season, since AdRoll reported prospecting CPMs down 45% year over year without publishing conversion data.
