# AI Marketing Agent: It Proposes, You Approve

> An AI marketing agent can read your analytics and Search Console, then propose the next page or fix. Here is how to set one up so nothing goes live without a person saying yes.

**Published:** November 17, 2026
**Updated:** November 11, 2026
**Author:** Matthew Clarkson, Osher Digital
**Category:** AI Agents
**Reading time:** 8 min read
**Source:** https://osher.com.au/blog/ai-marketing-agent/

---

An AI marketing agent can read your analytics, your Search Console data and your competitors' pages every week. It can then suggest the next page to write or the next fix to make. The safe setup lets the agent propose and a person approves each change before anything goes live.

We are Osher Digital, a [Brisbane AI consultancy](https://osher.com.au/ai-consultants-brisbane/). We build agents that read freely and write only with a human yes. This post shows how that pattern works for a marketing website, and where it breaks.

## What Jon Yongfook set up

On 3 November 2026, the founder [Jon Yongfook posted](https://x.com/yongfook/status/2117555432525099013) a short description of his "marketing loop":

> "It looks at my site analytics, search console data, ahrefs, and competitor activity. Then suggests things to add to the marketing site. I approve, it executes, and learns based on what I approve / reject."

He runs it daily across a portfolio of software products. That is the full description. He gives no traffic or revenue figures, so we cannot say if it works. It is one person's setup, not a case study.

We still think the design is right. It has four parts:

1.  The agent reads data from several sources.
    
2.  The agent proposes changes.
    
3.  A person approves or rejects each proposal.
    
4.  The agent does the approved work and records the decision.
    

## Why the approval step matters

An agent can write a page in seconds. It can also be wrong in seconds. The approval step catches the errors before your customers see them.

These are common errors when an agent writes marketing copy:

-   It states a fact about your service that is not true.
    
-   It writes a page that competes with a page you already have.
    
-   It makes a claim about results that you cannot support.
    
-   It names a client without permission.
    
-   It changes a URL and breaks the links to it.
    

Your business is responsible for what your website says. Under the Australian Consumer Law, a claim on your website must not mislead. An agent does not carry that responsibility - you do.

The approval step also gives the agent its best training data. Each rejection, with a short reason, tells the agent what you will not accept. Most agents do not change their model when you reject an idea. They read a log of past decisions before the next run. So keep the log, and write the reason in one plain sentence.

## What this post does not cover

This post is about one loop: read, propose, approve, execute. We do not repeat these topics here:

-   General marketing automation. Our guide to [automating your marketing with AI](/blog/how-to-automate-your-marketing-with-ai/) covers that.
    
-   Pages at scale from templates. See our post on [programmatic SEO](/blog/programmatic-seo-scaling-content-strategy/).
    
-   Which agent platform to use. Our [Grok Bot for Business](/blog/grok-bot-for-business/) post explains how we run our own bots and compares the options.
    

## A sensible loop for a small Australian firm

### Inputs worth connecting

Start with the data you already have. Give the agent read access only.

| Source | What the agent learns from it | Access to give |
| --- | --- | --- |
| Google Search Console | The queries that show your pages, and the pages that lose clicks | Restricted (read-only) user |
| Google Analytics 4 | The pages that bring enquiries, and the pages people leave fast | Viewer role |
| A backlink tool such as Ahrefs | The pages that earn links, for you and for competitors | Read-only API key |
| Competitor websites | New pages and new services on public pages | None. Public pages only |
| Your CMS | The pages you have, so it does not duplicate them | Read, plus create drafts |
| Your CRM | The pages that come before a real enquiry or sale | Read-only |

Our own Marketer bot works this way. It connects to Search Console, GA4, Google Ads, Bing Webmaster Tools and our CMS, and it does a weekly search check. It leaves every new ad campaign paused for a person to review.

The CRM input is the one most people skip. Without it, the agent chases traffic. With it, the agent can see which pages come before an enquiry.

### The weekly cycle

For most small companies, a weekly run is enough. Search Console data arrives a few days late, and a small site does not change much from day to day. A daily run on a small site mostly reports noise.

1.  Pull the last 7 days of data from each source.
    
2.  Compare it with the 4 weeks before.
    
3.  Write no more than five proposals. Give each one the evidence and the expected effect.
    
4.  Send the proposals to one person for a decision.
    
5.  Do the approved work as drafts in the CMS.
    
6.  Let a person review and publish each draft.
    
7.  Record each decision and its reason in the log.
    
8.  Check the result of each published change four weeks later.
    

The limit of five proposals is important. If the agent sends twenty, people stop reading them and approve all of them.

### What the agent may propose, and what a person decides

| The agent may propose | A person decides |
| --- | --- |
| A new page for a query you show for but do not answer | If the page goes live |
| A new title and meta description for a page that loses clicks | Any price, guarantee or result figure |
| Internal links between related pages | Any client name or testimonial |
| New FAQ questions from real search queries | Any statement on legal, medical or financial matters |
| A refresh for a post with old facts | Any change of URL, and the redirect for it |
| A fix for a broken link | Any spend on ads |

## Read freely, write with a yes

The permission rule is simple. We use the same rule in the [custom AI assistants](/services/custom-ai-assistant-development/) we build for clients:

-   Reads run freely.
    
-   Anything that changes data waits for a human yes.
    
-   Anything that deletes data asks every time.
    
-   Sending an email always waits for a person.
    

Apply the rule in the tools, not only in the prompt. A prompt is an instruction. A permission is a limit. Give the agent a CMS account that can create drafts but cannot publish. Give it a read-only analytics role. Then a bad prompt cannot publish a bad page.

## Where this kind of loop breaks

### Too little data

A site with 200 clicks a month gives the agent very little to read. It will find "trends" in random changes. On a small site, ask for fewer proposals and longer periods.

### Approval fatigue

The loop fails when the approver stops reading. Keep each proposal short. Show the evidence first.

### Pages that add nothing

An agent can produce many similar pages quickly. Google's spam policies name "scaled content abuse": many pages made mainly to rank, with little value for the reader. Approve a new page only when it answers a question better than your current pages do.

### A log nobody keeps

If the agent cannot read past decisions, it proposes the same rejected idea again next week.

### Too much access

Do not give the agent your admin login. Give each tool the smallest role that does the job, and remove the access when you stop the loop.

## Frequently asked questions

### What is an AI marketing agent?

An AI marketing agent is software that reads your marketing data and does marketing tasks. A safe one reads analytics, Search Console and your website, then proposes changes. A person approves each change before it goes live.

### Should an AI agent publish to my website on its own?

No. Let the agent create drafts. A person should review and publish each page. Your business is responsible for every claim on your website, and the approval step is where you check those claims.

### What data should an AI marketing agent read?

Start with Google Search Console, Google Analytics 4 and your CMS. Add a backlink tool and competitor pages when the basics work. Connect your CRM too, so the agent can see which pages come before an enquiry.

### How often should the agent run?

For most small firms, once a week. Search Console data arrives a few days late, and small sites change slowly. A daily run on a small site mostly reports noise.

### Does the agent learn from my decisions?

Usually it reads a log of your past decisions before each run. It does not change its model. So record each rejection with a one-line reason, and make sure the agent reads that log.

## Talk to us about a marketing agent

We build agents that read your systems and prepare the work, with publishing left to you. If you want a loop like this for your website, [book a 15-minute call](/book-a-call/). Our [AI agent development](/ai-agent-development/) page explains how we build and guard them.
