September 2026 in AI Marketing: What Actually Shipped

Seven significant things happened in AI marketing in the last ten days of September. I read the primary document behind each one before writing this, which turned out to matter more than usual: three of the seven, as widely summarised, were wrong in ways that would change what a business does about them.

Taken together they describe one shift, and it is not the one the headlines are chasing.

what actually changed in AI marketing this month?Agents stopped waiting for a prompt. They now run in the background, on your accounts, spending your money — and in the same fortnight, regulators started asking who is accountable when they do.

The pattern: autonomy went up, and so did the paper trail

Until this month, an AI assistant was something you opened. You asked, it answered, the exchange ended. Every product announcement in the last ten days moves away from that, in the same direction, from four companies that did not coordinate.

And in exactly the same fortnight, two regulators and one platform moved on accountability — labelling, liability and reporting. That is not a coincidence so much as a consequence. When software begins acting on its own against real accounts and real budgets, the question of who answers for it stops being theoretical.

What shipped: agents that do not wait

OpenAI's Dots. Announced at DevDay on 29 September, Dots are always-on agents that work on your behalf between conversations rather than only when prompted. Available on Pro and Business Premium plans in eligible markets. The same day brought GPT-6.1 Sol, which OpenAI describes as near-Astra intelligence at a fifth of standard input and output token prices, and @ChatGPT in Slack and Microsoft Teams — mention it in a channel and it helps, with no individual licence per person.

Meta's Muse for Small Business. Announced 29 September. Meta calls it “a personal AI agent that works in the background to help your small business reach its goals”: you give it an objective, it does the tasks. It connects to Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom, plus Facebook and Instagram business accounts. It is free for most of what people need, with paid tiers above that, currently in the US and Canada.

One sentence in Meta's announcement is more important than the connector list: “nothing publishes, sends, or spends without your approval.” That is the vendor conceding, unprompted, that the control question is the real question. Note which products this month say it and which do not.

Google's Business Agent. September's Demand Gen Drop put a conversational agent inside YouTube video ads, so a viewer can ask about a product without leaving the video, plus one-tap paths from image ads on Shorts and Gmail. The ad stopped being a message that points somewhere else and became the place the conversation happens.

Perplexity's contextual embedding model. On 30 September Perplexity released pplx-embed-v2-context-9b-preview, a nine-billion-parameter model trained so each chunk of a document is encoded with the whole document in view, setting a new state of the art on the ConTEB benchmark. This one is technical and it is the most relevant to anyone who cares about being cited.

That last point deserves a sentence of its own, because it is the quiet one. Retrieval is moving from “find the passage that answers this” to “find the passage that answers this and the surrounding material that supports it.” A page that states a claim and nothing else becomes weaker. A page that states a claim, shows the working, and sits among related material that corroborates it becomes stronger. That is a change in what good content is, arriving through a model release rather than a guideline update.

What shipped: the accountability layer

The FTC opened a rulemaking on scam ads. On 24 September the Commission issued an advance notice of proposed rulemaking asking whether its Rule on Impersonation of Government and Businesses should be updated to address platforms' role in spreading impersonation scam ads. It seeks comment on the ad-optimization tools offered by social media, search engines and other digital marketplaces, and on whether platforms should be required to vet advertisers and monitor posted ads. Comments are open for 60 days after Federal Register publication. Consumers reported $3.5 billion in impersonation fraud losses in 2025.

The EU published icons for labelling AI-generated content. These form part of the Code of Practice on marking and labelling AI-generated content and support compliance with Article 50(4) of the AI Act.

Google made its automation legible. The AI Brief beta expanded to seven more languages, and AI Max campaigns gained unified reporting that links search queries, the creative assets matched to them, and the landing pages selected — in one view. After two years of advertisers being asked to trust automated decisions they could not inspect, this is the direction reversing slightly.

Three things the coverage got wrong

This is the part worth your attention, because each of these errors points the wrong way — toward doing something you do not need to do.

The EU icons are optional. Summaries said businesses running AI-generated marketing imagery in the EU must now implement these markers or face penalties. The Commission's own page says the opposite, plainly: “The use of these EU icons is optional, but the labelling requirements under Article 50 AI Act are not.” Two different things. Your disclosure obligations under the AI Act are real and unchanged. The icons are a provided convenience for meeting them, not a new requirement. Anyone selling you an urgent icon-compliance project this week is selling you the summary rather than the source.

The FTC notice is not about AI, and not aimed at advertisers. It was widely framed as the FTC moving to hold platforms liable for algorithmic impersonation, with advice that advertisers prepare for tighter bidding controls and scrutiny of automated creative. The actual notice concerns impersonation scam ads generally, not AI specifically, and it targets platforms rather than individual advertisers. The advice attached to it in circulation is not in the document at all. If you buy ads honestly, this is a thing to watch, not a thing to prepare for.

Meta made two announcements, not one. The Enterprise Platform on 28 September and Muse for Small Business on 29 September got merged in summaries into a single item, producing a description of an enterprise platform with Shopify and Klaviyo connectors. The connectors belong to the small-business product, which is free and available now. The Enterprise Platform is a division and a hire — CJ Desai as Chief Enterprise Platform Officer — with no stated pricing, availability or eligibility. If you are a small business, the thing you can actually use today got described as something you cannot.

3 of 7items materially misstated, in summaries that each linked to a correct primary source. Every error was visible in the first paragraph of the original document.

I will be direct about where I found these. The automated news feed on this site produced versions of all three before I checked them against the sources, which is why that page tells you to read the original rather than my summary of it, and why I am now building a verification step into it. An automated summariser reaching for the more dramatic reading is not a bug peculiar to my setup. It is what these systems do, and it is why the four-minute check in the previous piece exists.

Curious how AI assistants describe your business right now? The free audit asks a live assistant real customer questions and shows which sources it used to answer.

Run the free audit

What I would actually do

Decide your approval rule before you connect anything. Muse connects to Stripe, QuickBooks, Shopify and your ad accounts. Dots run while you are not watching. Both are useful. Before either touches a live account, write down which actions require a human and which do not — publishing, sending, spending, changing a price, replying to a customer. Meta already drew that line for you in its own announcement. Most businesses will not have drawn it anywhere.

Try the zero-cost one first. If your team is in Slack or Teams, @ChatGPT is the lowest-friction adoption path announced this month, because nobody changes where they work and nobody buys a seat. You will learn more about where AI actually helps your operation in a week of that than in a quarter of planning.

Re-price the project you shelved. If you costed an AI build and the token price killed it, run the numbers again against GPT-6.1 Sol. A fifth of the cost moves the line on what clears.

Write to be retrieved with your evidence. The Perplexity release is a signal about where retrieval is going. Pages that assert things are being outcompeted by pages that assert, support and sit in context. Practically: keep the claim and the proof on the same page, and link the related material rather than stranding each page on its own.

And check the number before you act on the feeling. Three of seven, this month alone. The error rate in AI coverage right now is high enough that reading the source is not diligence, it is basic competence.

Common questions

Do EU businesses have to use the new AI labelling icons?
No. The European Commission states that the use of the EU icons is optional, while the labelling requirements under Article 50 of the AI Act are not. The icons form part of the Code of Practice on marking and labelling AI-generated content and exist to support compliance with Article 50(4). Your underlying disclosure obligations are unchanged; the icons are a provided means of meeting them rather than a new requirement.

Does the FTC rulemaking mean advertisers face new rules on AI ads?
No. The advance notice of proposed rulemaking issued on 24 September 2026 asks whether the Rule on Impersonation of Government and Businesses should be updated to address the role of platforms in spreading impersonation scam ads. It concerns impersonation scam ads generally rather than AI specifically, and it is directed at platforms rather than individual advertisers. Comments are open for 60 days following Federal Register publication.

What is Meta's Muse for Small Business and what does it cost?
Announced on 29 September 2026, it is a personal AI agent that works in the background on business objectives, connecting to Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom, plus Facebook and Instagram business accounts. Meta states it is free for most of what people need, with paid subscription tiers above that, and that nothing publishes, sends or spends without the owner's approval. It is available in the US and Canada. It is a separate product from the Meta Enterprise Platform announced on 28 September.

What did OpenAI announce at DevDay 2026?
Dots, always-on agents that work between conversations, available on Pro and Business Premium plans in eligible markets; GPT-6.1 Sol, described as offering near-Astra intelligence at a fifth of standard input and output token prices; @ChatGPT in Slack and Microsoft Teams, usable in a channel without individual licences; ChatGPT Space for shared team workspaces; and the ability to add plugins to Sites.

Why does Perplexity's embedding model matter for marketing?
Because it changes what retrieval rewards. The model encodes each chunk of a document with the whole document in view, improving the retrieval of supporting evidence alongside direct answers. The practical implication is that a page which states a claim in isolation becomes weaker relative to one that states the claim, shows the supporting detail, and sits among corroborating material. It is a change in what counts as good content, arriving through a model release rather than a published guideline.

Sources

About the author. Siamak Kalhor is an AI marketing and LLMO consultant in Los Angeles with more than forty years of experience in marketing strategy. He advises professionals and e-commerce leaders on becoming the answer AI assistants give. Get in touch.

The month the agents stopped waiting to be asked.