👋 Hey,
Welcome to the latest edition of The AI Driven Marketer!
Every week, I cut through the noise to bring you the most practical marketing AI news, tools, and a battle-tested automation workflow you can plug directly into your business.
Let's get into the build.
⚡ The AI Pulse: What Marketers Need to Know
OpenAI Sets a Shutdown Date for Custom GPTs
OpenAI began sending admin notices on September 11 and published its retirement FAQ on September 12. Custom GPTs are being retired, with plugins named as the replacement. The published calendar covers Enterprise workspaces: new custom GPT creation ends September 25, and custom GPTs stop running on December 11. OpenAI says the transition affects all ChatGPT plans and that other plans may follow the same timeline, without committing to the same dates for everyone. Every date is marked subject to change.
So What? If you built a GPT your team or a client relies on, you have about 11 days before you can create new ones and roughly three months before the ones you have stop working. Custom actions do not carry over. Get your instructions and knowledge files out this week. The bonus below is the extraction kit.
Meta Launches Muse, a Personal Agent That Can Complete a Purchase
Announced September 8, Muse runs on a dedicated cloud machine Meta calls Muse Secure VM, with its own browser and a model called Muse Spark. It can open a browser, fill forms, send email, book travel and check out using Link by Stripe, which issues a one-time-use card so your real details stay hidden. A separate Sentinel agent sits on the same machine, and nothing Muse does reaches the internet without its approval. Muse asks before sensitive actions and shows an audit trail. It is rolling out in the US on iOS, Android and muse.ai, and you can talk to it inside WhatsApp. Meta says it is free for most of what people need, with paid plans it has not priced.
So What? An agent that completes checkout changes who your funnel is talking to. When a buyer sends software to compare options and buy, your product pages have to be legible to a machine as well as persuasive to a person. Clear specs, real prices and plain descriptions carry more weight than clever copy in that setting.
OpenAI Ships ChatGPT Images 2.5 With Sketch and Shareable Prompts
Released September 8 to every ChatGPT, ChatGPT Work and Codex tier on desktop, mobile and web. It holds a subject's likeness better across edits, changes only the part you asked about, and stays consistent over a long back-and-forth. New in the interface: Sketch, where you draw in chat as a reference by typing @Sketch, templates for formats like posters and merch, comments placed directly on an image, and the option to share the prompt along with the image. Developers get GPT-Image-2.5 Flare as the default and GPT-Image-2.5 Sunburst for tighter control. Images still carry C2PA metadata and invisible watermarking.
So What? The shareable prompt is the part with distribution in it. A product shot that took you 20 minutes to dial in becomes a template other people run on their own products, with your framing attached. Multi-turn consistency also means one approved hero image can carry a whole campaign instead of drifting with every edit. Note that the 50% latency reduction and the figure of over 3 billion images a week are both OpenAI's own numbers.
Adobe Puts Generative Video and Sound Inside the Premiere Timeline
Announced September 8 ahead of IBC. Premiere's new Generative Media Tool creates editable video and sound effects at the point in the timeline you pick, and you choose the model: Adobe Firefly, Google Veo, Kling, Runway or Luma. Generate Video samples reference frames from your own project so new footage matches the surrounding shots. Generate Sound Effects lets you guide rhythm and intensity with your voice. Generate Soundscape, in beta, reads up to 15 seconds of video and builds layered ambience timed to the cut. Generate Music, also in beta, produces commercially safe tracks with tempo and looping controls. After Effects gets an AI Assistant in public beta that works across an entire project. Frame.io Mounted Storage is now on all accounts.
So What? Filling a b-roll gap or building a sound bed no longer means a second subscription and a second export. For a solo content marketer that removes the two most common reasons a good cut sits unfinished. The beta label matters though, and Adobe did not publish credit costs, so keep Soundscape and the After Effects Assistant off client deadlines for now.
Google Ads Adds a Metric That Prices Your First-Party Data Setup
Posted September 10. Google is putting Data Manager directly inside Google Analytics and Display & Video 360, launching enhanced conversions in both, and making the Data Manager API universal on the IAB Tech Lab ECAPI standard, with built-in diagnostics that flag data problems before they hit performance. The new Data Strength Uplift Metric in Google Ads calculates the extra conversions your first-party data setup recovered. Meridian, Google's open-source marketing mix model, gains agentic data auditing, accepts brand signals such as Branded Google Query Volume, and GeoX is now generally available worldwide.
So What? Arguing for first-party data plumbing used to mean arguing from principle. The Uplift Metric puts a number on your own account, and that is the number to take into a budget conversation. The uplift figures Google quotes, 26% incremental ROAS, 11% more Search conversions, 14% conversion uplift and over 20% for Demand Gen, are all Google's own averages and are not independently audited, so read them as a sales pitch and your own metric as the evidence.
🛠️ E-Books & Guides
🛠️ The AI Toolkit
3 Tools specifically curated for digital marketers.
1. IntelCue
Best for: Solo marketers and founders who want competitor intel to arrive inside Claude or ChatGPT rather than in one more dashboard.
What it does: Point it at your market once and it monitors 12+ source types continuously, including competitor newsletters, website pricing changes, Google Ads Transparency Center creatives, patents, SEC filings and newly issued SSL subdomains, then serves findings into Claude over MCP, into ChatGPT as an app, or as a weekly newspaper-style brief by Slack and email. Flat $8.99 a month with everything included and no tiers.
2. Kinetic Pricing
Best for: Founders with paying customers who are weighing a price increase and want evidence instead of a hunch.
What it does: Runs a price sensitivity study on your own customers, then models what a candidate price does to monthly revenue and names the retention rate at which the move breaks even. One study starts at $149, or $99 a month for unlimited studies with a 30-day trial. It does not recruit respondents, and it states plainly that survey answers are stated preference rather than a checkout prediction.
3. unquestion
Best for: Marketers running a churn, feedback or discovery round in bursts who do not want another monthly subscription.
What it does: Turns a static form into an AI-led conversation that asks its own follow-up questions, then returns answers as structured table rows instead of chat transcripts, exportable to CSV. One-time packs run from $29 for 250 conversations to $349 for 5,000, building and previewing are free, and credits last 12 months.
Bonus
The Custom GPT Migration Kit
OpenAI published the retirement calendar on September 12. For Enterprise workspaces, new custom GPT creation ends September 25 and custom GPTs stop running December 11. OpenAI says the change affects all ChatGPT plans and that other plans may follow the same timeline, and it marks every date subject to change. Either way, something you built now sits on a clock you do not control.
The replacement OpenAI names is plugins. A plugin bundles skills, which are reusable instructions living in a folder with a SKILL.md file, and optionally connectors to tools like Slack, GitHub or Google Drive. A GPT's instructions map cleanly onto a skill. Its custom actions do not.
Here is the extraction in order. Budget about 20 minutes for three GPTs.
Step 1: Inventory what you have
Open your GPTs and build one row each in a spreadsheet:
Name | Who created it | Who uses it | Custom actions? | Knowledge files? | Rebuild or retire? |
|---|
Most people discover half their GPTs were abandoned experiments. Mark those retire and stop thinking about them.
Step 2: Copy the instructions out word for word
Open each GPT you are keeping, click Edit, and copy the entire Instructions field into your doc. Do not summarise or tidy it. The exact wording is what produced the behaviour you liked.
Step 3: Download every knowledge file
In the same Edit screen, download each file under Knowledge. These are not stored anywhere else you control. When the GPT stops running, anything you did not download is gone.
Step 4: Flag the custom actions separately
If a GPT calls your own API through Actions, write down the endpoint, the auth method and the schema. This is the part that does not migrate. In plugin terms it becomes an MCP server, which is a build rather than a copy and paste. Decide now whether it is worth rebuilding.
Step 5: Publish any draft you want to keep
A draft GPT cannot migrate. If you have something half-built that matters, publish it before creation closes on September 25. Sharing it publicly is not required, but publishing is what makes it eligible.
Step 6: Rewrite the instructions as a skill
Create a folder named after the job, not the tool. Inside it, create SKILL.md in this shape:
---
name: [Job Name]
description: [One sentence naming the artifacts this applies to. Under 200
characters. This is the only thing the model reads before deciding whether
to load the skill, so name the deliverables explicitly.]
---
# [Job Name]
## When to use this
[The trigger. Name real requests: "a weekly report", "a client proposal",
"a product description".]
## Inputs I will give you
[What you paste or attach each time.]
## Steps
1. [One instruction per line. Your GPT's instructions go here, broken into
ordered steps.]
2.
3.
## Rules
- [Your non-negotiables, lifted from the GPT instructions.]
- [Anything the GPT used to get wrong, written as a prohibition.]
## What good looks like
[The finished artifact: length, format, tone.]
## Before you return anything
Check the output against the Rules above line by line. If any rule is
broken, fix it and name which one.The description field is where these migrations fail. A skill loads by description match, so "brand voice work" will sit there never firing, while "newsletter, LinkedIn post, landing page and email drafts" gets picked up.
Step 7: Test with four prompts, one of which should fail
Three that should load the skill, one that should not:
[The exact request your GPT was built for.]
[The same request phrased the way a colleague would say it.]
[An edge case the old GPT handled well.]
"Summarise this PDF." This one should NOT load the skill.
If 1 through 3 do not load it, the description is the problem, not the steps. Rewrite it to name deliverables and test again. If 4 does load it, your description is too broad and the skill will fire on everything.
The honest limits
Custom actions do not transfer. An API-backed GPT is a rebuild.
Drafts cannot migrate at all, so publish before September 25 or lose them.
Migration is rolling out per workspace, so a missing option is not necessarily a fault. On Enterprise you need plugins enabled, the GPT published, and to be its creator or a workspace admin. Permission to use someone else's GPT does not let you migrate it.
Skills do not sync across every surface automatically.
A skill enforces the rules you wrote. It will not tell you the rules were wrong.
💡 Do steps 1 through 3 today for your top three GPTs. That is the part with a deadline attached. The rebuild can wait until the migration tool reaches your account.
That's a wrap for this week in AI.
See you next week,
Rananjay
P.S. If you found this valuable, share it with a fellow marketer who's trying to make sense of AI. And hit reply to let me know—what's the one AI workflow you want to automate this quarter?