AI Marketing News, 12 Oct 2026: ChatGPT Visual Ads, Claude Haiku 5.5 at $0.10, ZoomInfo Agent Teams
ChatGPT tests visual ads for 1.2B weekly users, Haiku 5.5 drops to $0.10 per million tokens, ZoomInfo Agent Teams ships, Kantar finds 23% trust AI picks.
OpenAI adds a visual ad format to ChatGPT and a wider measurement partner list, with WeightWatchers reporting 15.3% lower acquisition cost. OpenAI announced on October 5 that a visual format showing product inspiration and use will test during image generation for US Free and Go users later this month, with a small group of advertisers. ChatGPT reaches 1.2 billion people weekly. Hightouch, Tealium and LiveRamp can now pass conversion data into ChatGPT Ads, and attribution runs through partners including AppsFlyer, Triple Whale, Northbeam, Singular and Kochava. WeightWatchers reported an attributed cost per acquisition 15.3% below its blended paid-search benchmark, and Dose reported that 67% of incremental purchases came from net-new customers. Incrementality testing with Haus, Measured and WorkMagic is described as early, and DoubleVerify and Integral Ad Science are piloting brand-suitability checks.
Importance for marketers: Every result in the announcement is a consumer brand measured by the vendor’s own chosen partners, and the 15.3% figure compares attributed ChatGPT conversions with a paid-search blend, which is a different attribution basis. The useful part for B2B is the Hightouch and Tealium route, because it lets you push closed-won and opportunity events from your warehouse or CDP instead of relying on a pixel that sees form fills. Wire that feed before any test budget goes live, and judge the channel on sourced pipeline. The setup is in ChatGPT Ads as a B2B pipeline channel.
Digiday: seven months in, measurement gaps keep ChatGPT ad budgets at test level, with one agency’s clients spending under $100,000 a month against $10 million on Google. Digiday reported on October 1 that UK agency Accuracast saw ChatGPT report about 100 clicks where its own tracking recorded 20, and a lead-gen campaign showed form fills but zero platform conversions two weeks later. Web Guide Partner sees conversions arrive 24 to 36 hours late, against a few hours on Google. Legal teams at many advertisers refuse OpenAI’s conversion-tracking terms over liability and data concerns, and advertisers can now bid toward cost per acquisition. Accuracast keeps client test budgets between £10,000 and £20,000 until results are verifiable.
Importance for marketers: The blocker is contractual before it is technical, so a B2B team that wants to test ChatGPT Ads should get legal review of the tracking terms started now, in parallel with the creative. Agree the success metric before launch, log clicks from your own UTM-tagged landing pages next to the platform’s click count, and treat any gap above 20% as a stop signal. Bidding toward CPA on a delayed, unreconciled conversion feed will optimise against noise.
Anthropic releases Claude Haiku 5.5 at $0.10 per million input tokens and halves Sonnet 5.5 cache-read prices. Haiku 5.5 launched October 7 at $0.10 input and $0.50 output per million tokens for requests under 100,000 tokens, down from Haiku 4.5’s $1 and $5. Larger requests cost $0.50 and $2.50. Anthropic estimates about 75% average savings after an updated tokenizer that uses slightly more tokens per task. Sonnet 5.5 cache reads fall from $0.20 to $0.10 per million tokens, which Anthropic says makes most agentic tasks about 20% cheaper. Haiku 5.5 is the first Haiku with effort controls, scores 72.4% on the OSWorld 2.1 offline computer-use subset against 15.7% for Haiku 4.5, and is available on Claude Platform, AWS, Google Cloud and Azure.
Importance for marketers: At these prices, classification and extraction that you rationed to a sample can run across the full table. Enrichment cleanup, job-title normalisation, intent tagging on every inbound reply and call-note triage in Clay or n8n all become per-row costs you can ignore. Re-run your existing test set on Haiku 5.5 for the cheapest steps, and move long agent loops to Sonnet 5.5 so the cache-read cut applies. Cheaper tokens also remove the natural brake on runaway loops, so set a per-run token cap before you widen the volume, as laid out in Token budgets and audit logs for GTM agents.
ZoomInfo launches Agent Teams on GTM Studio after buying DoubleO.ai, billed as AI Credits against existing contracts. Agent Teams orchestrates specialised agents that run a full play from a trigger such as a buying signal through outreach and CRM updates. It ships with six templates, including renewal at risk (starting 90 days out), closed-won handoff, closed-lost win-back (monthly) and closed-won competitor outreach, where the agent finds 10 competitors of a new customer. Usage is metered in consumption-based AI Credits on the current ZoomInfo contract, with an estimate shown before a play activates, and no prices are published. ZoomInfo’s marketing says human approval is built in, but its FAQ says the Pause for Review step is still on the roadmap.
Importance for marketers: A vendor agent that can email prospects and write to the CRM without an approval gate is a production risk, so ask ZoomInfo for the current state of Pause for Review in writing before any play with send access goes live. The six templates are close to plays teams already build in Clay and n8n with their own routing and suppression logic, so the buy-or-build question is who owns the signal definitions and the exception handling. Price a pilot in credits against your current Clay and n8n run cost for the same play, using the estimate ZoomInfo shows before activation.
Snowflake previews AI Gateway for Advertising, connecting Cortex Agents to live Meta and TikTok campaign data through governed MCP servers. The gateway is built on Snowflake’s May 2026 acquisition of Natoma’s enterprise MCP platform, with PMG as an early customer. Agents combine live delivery, catalogue health and diagnostics with warehouse data on transactions, inventory and historical performance, then recommend actions that campaign managers approve. In Snowflake’s retail example, ROAS fell while pacing and click-through looked normal, and the agent traced it to a checkout payload change that degraded purchase events, catalogue warnings on promoted SKUs and low stock on a high-spend product group. It recommended a 20% budget cut instead of a pause, because the converting cohort had strong lifetime value. Every recommendation, its evidence and its approval land in one shared record.
Importance for marketers: The agent found its answer by joining ad-platform data to inventory and event quality, which most paid teams cannot do because those tables live in different systems. If your CRM and revenue data already sit in a warehouse, the join is the work to prioritise this quarter, whichever vendor’s agent reads it afterwards. The approval record is the other feature to copy: log each agent recommendation with its evidence and the human who accepted it, which gives you the audit trail any paid media agent needs before it touches budget.
Kantar: net 75% of marketers will raise AI assistant spend in 2027, while 23% of consumers trust assistant brand recommendations. Kantar’s Media Reactions 2026 report, released October 2 and based on over 800 senior marketers and 23,000 consumers, found that 62% of marketers expect generative AI to play a pivotal role in brand recommendations. Among consumers, 32% use AI assistants to research brands and products, and 23% trust what they recommend. Only 50% of marketers are confident in their brand and performance balance, down from 60% a year earlier, and 51% say they extract actionable intelligence from their data.
Importance for marketers: The gap between planned spend and consumer trust means the budget is chasing a channel most buyers treat as a first draft. Treat assistant visibility as a source of shortlist inclusion and measure it through branded search, direct visits and “how did you hear about us” fields, because the trust figure suggests buyers verify elsewhere before they act. Fund the fixes that feed both, such as accurate third-party profiles and clear pricing and integration pages, and hold back from big standalone AI-visibility tooling contracts until you can tie them to sourced pipeline.