ChatGPT ads expand, NVIDIA routes models: Aug. 12

ChatGPT ads now reach five more countries, NVIDIA has released a tool for cutting the cost of automated AI work, and Grok Bot can learn a browser routine by watching you. Give this dispatch 30 minutes. You'll leave with a visibility check, a model-cost rule, and one tightly controlled task to test.

ChatGPT ads expand into five more markets

Five new markets received ChatGPT ads on August 11: the United Kingdom, Mexico, Brazil, Japan, and South Korea. OpenAI's updated announcement says the placements appear for logged-in adults on the Free and Go plans. Paid Plus, Pro, Business, Enterprise, and Education accounts remain without ads.

This adds a paid placement beside the answers customers use while comparing purchases. It does not replace the work that earns a mention inside the answer. OpenAI says ads are labeled, separated from the answer, and selected using the conversation topic, past chats, and earlier ad interactions. Advertisers receive combined view and click totals, not a customer's chats or personal details.

Treat paid and unpaid ChatGPT visibility as separate lines in your reporting. First, run the five buyer questions from your AI recommendations visibility setup using a Free account in each market you serve. Record the answer, any sponsored placement, and the date in one sheet. Then add two rows to your weekly report: ChatGPT unpaid visits and ChatGPT ad visits. Do not combine them into one “AI” total.

NVIDIA releases a router for AI model costs

NVIDIA released Nemotron 3.5 Lightning and NeMo Switchyard on August 11. Lightning is a 30-billion-parameter model, meaning the AI has 30 billion adjustable values behind its answers. NVIDIA's launch post says Switchyard is an open-source router that sends each job to a suitable AI model based on requirements such as quality, response time, and cost.

The business consequence is simple. An automated routine does not need your most expensive AI for every step. Sorting form responses and labeling support tickets are narrow jobs. Writing a sensitive customer reply or deciding which exception needs a person deserves the stronger model. NVIDIA reports its own benchmark results, so use them as a reason to test, not as a promised saving.

If your business already pays for AI through an API, the connection that lets software call an AI model, export last week's usage by task. Mark each task routine, judgment, or customer-facing. Ask the person who maintains the automation to route only routine work through a smaller model for one week, while keeping the existing model as the fallback. Compare cost, errors, and processing time before widening the change. The same control rule belongs in any weekly AI marketing report.

Ignore Switchyard for now if staff use AI only by typing into ChatGPT or Claude. There is no routing bill to fix. If you do run several automated AI jobs, review the model choice before replacing your automation platform. This AI automation tool comparison covers the separate question of which system should move the work between apps.

Grok Bot learns a browser routine by watching you

Grok Bot's public product page went live on August 11 with scheduled routines and a teaching mode. The official page says a Bot can use websites and apps through its own computer, remember how work is done, and pause when approval is required. You can also run several Bots at once.

That makes browser-only work easier to automate, including systems with no ready-made connection. It also gives the Bot the same reach as the account you hand it. Start with a reversible task. A weekly report draft is suitable. Sending refunds, editing prices, deleting records, or publishing content is not.

Open Grok Bot and choose Teach a task. Demonstrate one routine using copied data, then stop before the final send or publish button. Give the Bot a separate user account with permission to view only the records it needs. Run the saved routine three times, compare every field with your source sheet, and require approval for the last action. Keep parallel Bots off until one routine produces the same result three times.

The part that breaks is account access. A Bot that signs in as an owner inherits too much power, and a learned routine can preserve a bad click. Limit the account first. The workflow comes second.

On the bench

• OpenAI's Daybreak cyber models are now available through Amazon Bedrock, but the access and security-review burden puts them outside a normal marketing stack.
• llama.cpp now pairs with a local coding agent, which means the files and requests can stay on your computer. Test it only when that privacy is worth the setup time.

For the next practical connection, use the n8n and Make decision sheet and choose the system that matches who will maintain it. Get the next build.