AI Marketing News, 23 Sep 2026: ChatGPT Sponsored Agents, HubSpot Breeze Assistant, ChatGPT Share at 50%
OpenAI pilots Sponsored Agents and wires ChatGPT Ads into HubSpot, Salesforce ships Salesforce in Claude, Comscore puts ChatGPT at 50% of prompts.
OpenAI pilots Sponsored Agents in ChatGPT and connects ChatGPT Ads to HubSpot and Shopify. Announced September 16, Sponsored Agents let a user who sees an ad open a separate, labelled conversation with an agent run by the advertiser, ask about features, sizing or compatibility, then click through to the advertiser’s site. OpenAI says the sponsored thread sits apart from ChatGPT’s own answers and from the user’s original conversation, and that advertisers never see private chats. The pilot is limited to select US advertisers; press reports name Angi and Wayfair among the early participants. Ads Manager gains natural-language campaign creation and reporting through a plugin in ChatGPT, AI-generated copy and imagery built from a landing page, and opt-in headline customisation that adapts to the conversation and translates into the user’s language. HubSpot is the first CRM partner and says a connected ChatGPT Ads account sends traffic to HubSpot landing pages, enrols each contact in workflows and attributes it alongside every other channel. Shopify merchants get a ChatGPT Ads app, US first and international from September 23.
Importance for marketers: The ad unit is now an agent that qualifies the buyer inside ChatGPT before the click, so the agent’s knowledge base is the creative: product facts, pricing rules, integration answers and objection handling. Build that corpus once from the same source of truth your sales team uses, because the agent will be asked the questions your SDRs get on a first call. The HubSpot connection removes the attribution objection that made last week’s Amazon DSP route a reach-only play. Route the pilot through HubSpot and hold it to the pipeline test in ChatGPT Ads as a B2B pipeline channel.
HubSpot ships Breeze Assistant, a touchless Smart CRM and native ChatGPT Ads and Microsoft Advertising channels. Announced September 16 at Unbound in Boston, the Fall ‘26 Spotlight centres on Growth Context, a layer that scores how complete a company’s customer context is. Breeze Assistant takes a goal, assigns agents and returns proposals, reports and campaign plans. The Smart CRM now captures calls, emails and meetings without manual entry, Mobile Notetaker and Deal Progression push transcript updates into records with one-click approval, and the Prospecting Agent monitors more than 40 buying signals. Marketing Studio surfaces an AEO visibility score, underperforming segments and unfollowed leads, then routes work to content and campaign agents. HubSpot claims early Marketing Studio users create 81% more campaigns and Breeze Assistant users generate 2.2x more leads. Microsoft Advertising joins as a channel with targeting on verified LinkedIn accounts across Bing, Copilot, MSN and Outlook. No pricing changes were announced; agents still bill through HubSpot Credits at the outcome rates set in April, $1 per lead for the Prospecting Agent and $0.50 per resolution for the Customer Agent.
Importance for marketers: Everything HubSpot launched runs off the completeness score, so a CRM with stale lifecycle stages and half-filled company records gets a worse Breeze Assistant than a clean one, and you will see that in the Context Home number before you see it in pipeline. Run the deduplication and enrichment pass first, then turn the agents on. The Microsoft Advertising integration is the quieter win: verified LinkedIn account targeting on Bing and Copilot gives you LinkedIn’s audience at search CPCs, bought from inside the CRM that reports the pipeline. Budget the Prospecting Agent as media spend rather than software: at $1 per lead the credit line scales with volume the way paid social does, and finance should see it in the same column.
Salesforce launches AIforce at Dreamforce, with Salesforce in Claude, Slackforce and Agentforce Coworker. Unveiled September 15, AIforce is an interface layer that carries Salesforce business logic, permissions and governance into whichever AI surface a person is working in. Salesforce in Claude, in beta, ships 37 prebuilt sales skills covering prospecting through pipeline hygiene, plus a Claude Code plugin with 40 skills; marketing, service and Tableau skills are due later. Slackforce adds Slack CRM, so reps open accounts, log call notes and update records from Slack. Agentforce Coworker is an AI teammate inside Salesforce that calls deployed Agentforce agents. AIforce sits on the Headless 360 MCP servers and APIs released in March. Marc Benioff described it as an interface revolution built with zero data retention. Pricing for Coworker and Salesforce in Claude was not published.
Importance for marketers: Your reps will update Salesforce from Claude and Slack within months, so the CRM stops being where people work and becomes the system of record behind three interfaces. That removes the adoption excuse for data hygiene and creates a permissions problem: an agent skill that can edit pipeline stages needs the same field-level governance you apply to a human, and someone in RevOps has to own that mapping. Write the skill allowlist now, while the beta is small, and meter token spend per skill from day one.
Anthropic merges Cowork into Claude and adds Docs and Slides. Starting September 16, Claude Cowork and chat became a single product, rolling out to Pro and Max plans over several weeks, with Team and Free to follow and Enterprise admins given at least 30 days’ notice. Claude Docs and Claude Slides launch in beta on paid plans, Claude Design now works inside conversations, and outputs download as PowerPoint or PDF. Anthropic’s launch example is a scheduled weekly pipeline report: ask what moved in the pipeline last week, have Claude write the report and five leadership slides from the same CRM-connected conversation, then run it every Monday. Claude asks before taking actions by default, with a check-in-only mode available. Existing projects, artifacts, connectors and skills carry over. A September 17 follow-up rebuilt Projects around conversation rather than folders.
Importance for marketers: The pipeline review, the campaign recap and the QBR deck are now one scheduled conversation connected to HubSpot or Salesforce, and the output is a document the CMO can edit rather than a chat reply someone pastes into slides. Most marketing ops teams still build that report by hand on a Monday morning. Set the schedule against your CRM connector this week and spend the freed hours on the data problems the report exposes.
Comscore: ChatGPT’s share of prompts fell from 70% to 50% in the first half of 2026. Comscore’s Q2 AI Intelligence Report, published September 22 from an opt-in panel of between 500,000 and one million users, shows ChatGPT’s share of prompt volume dropping from 70% in January to 50% in June, while Gemini nearly doubled from 17% to 30% and Claude rose from 2% to 11%. Total usage grew: 35% of desktop users and 29% of mobile users visited an AI assistant in June 2026, against 25% and 12% in January 2025. Gemini desktop conversations went from 8 million to 89 million in a year and Claude peaked at almost 40 million in April before settling at 22.3 million. Google desktop searches showing an AI Overview rose from 25.8% in July 2025 to 39.4% in June 2026. Comscore separated sourcing from citation: Tripadvisor is 61% of AI travel sourcing but 21% of citations.
Importance for marketers: A GEO programme that only tracks ChatGPT now covers half the prompt volume, and the half it misses skews toward Gemini, which sits inside Workspace, and Claude, which skews to knowledge workers, and B2B buying committees sit in both. Track visibility per assistant, and measure retrieval and referrals as well as citation, because the sourcing-versus-citation gap means an assistant can read your page every time and never name you.
Google Ads lets offline CRM uploads fill gaps in tag-based conversions. Documented September 21, the multi-source conversions beta lets an advertiser connect a CRM or other data source through Data Manager or its API, and Google matches offline records to tag-captured conversions by transaction ID. Where the tag missed user-provided data such as email or phone, the upload supplies it for Enhanced Conversions matching, with Google recommending uploads within 24 hours. Offline data can also overwrite conversion values with a final order total, upsell or refund. For matched transactions the GCLID in the upload is ignored, and a newly connected source enters a 14-day trial in which added conversions show in reporting but do not feed bidding.
Importance for marketers: B2B teams lose most of their conversion value to the gap between form fill and closed-won, and this closes it: the tag records the lead, the CRM upload rewrites its value to the qualified or won amount, and Smart Bidding optimises to that. Build the n8n job that emits a Data Manager upload on every HubSpot or Salesforce stage change, keyed on the transaction ID you already pass in the form, and start it now so the 14-day trial is over before Q4 budgets land.
Semrush: 81% of shoppers who use chatbots have decided against a purchase on the chatbot’s advice. A Semrush survey of 2,338 US adults, covered by MarTech on September 18, found 58% of AI users have bought something a chatbot recommended and the same share have rejected a purchase because of what a chatbot told them. Among people who use chatbots specifically when shopping, nearly 81% have decided against a purchase on AI guidance. 74% of all consumers would be less likely to buy if a chatbot surfaced mixed or negative reviews, rising above 85% among AI users. Only 29% of consumers use chatbots for product research, but that reaches 55% among weekly AI users, for whom chatbots now outrank friends, social media, retail sites and YouTube. On ads inside chatbots, 42% dislike them and 66% of that group say ads make them question the integrity of the answer. Semrush owns MarTech.
Importance for marketers: The assistant is a gatekeeper that reads your G2 page, your Reddit threads and your competitor’s comparison post before the buyer ever hits your site, and it will repeat the negative review verbatim. Accurate citation matters more than citation count. Audit what each assistant says about your product against your top three competitors, fix the review and comparison content it is drawing from, and put the same product facts you give sales into the pages the models retrieve. The weekly audit that catches this is built out in What AI assistants say about your product.
Agencies start billing AI tokens to clients as a separate line from labour. Digiday reported September 17 that agency AI costs now combine subscriptions, metered token usage and staff time. Ramp data shows about 31% of companies spend more than $10,000 a month on AI. Markacy pays for enterprise Claude across 25 or more staff, Crispin runs roughly 100 AI tools on flat annual fees, and Dept convenes a monthly AI Tool Committee while saying AI is not yet a double-digit share of budget. Forrester finds 74% of agencies use generative AI to summarise documents and 70% for research. Omnicom CFO Phil Angelastro told Goldman’s Communacopia conference the holdco is building a model to measure token costs and get reimbursed, moving toward outcome-based pricing, and one agency is already invoicing tokens to clients as materials.
Importance for marketers: Your 2027 agency contracts will arrive with a token pass-through line, and you cannot audit it without knowing what an agent run should cost. Meter your own agents first, per workflow and per output, so you have a benchmark to negotiate against, and ask the agency for the same audit log you would demand from an in-house build.
A three-layer attribution framework for AI search in B2B SaaS. Writing in Search Engine Land on September 21, Kristina Frunze of WebView SEO argues GA4’s AI Assistant channel is a floor rather than a count, because most assistant interactions end without a click, app referrers arrive as direct traffic and device switching drops UTMs. Her framework has three layers: direct attribution through GA4 plus an AI search lead source in HubSpot or Salesforce and an LLMs option on the how-did-you-hear form field; influenced attribution through direct and branded organic lift in Search Console, shorter sales cycles and competitive win rates; and a visibility layer tracking AI share of voice, citation rate and sentiment. She recommends six months of monthly tracking before drawing conclusions.
Importance for marketers: Add the LLMs option to the form field this week, because self-reported attribution is the only signal that survives zero-click, and add the lead source property in HubSpot so the number reaches the pipeline report rather than a slide. The board question is what share of qualified pipeline named an assistant, and that is answerable in a quarter. The rest of the gap is set out in The attribution gap nobody’s measuring.