Vibe marketing with an agent: a setup that works

Vibe marketing is the habit of describing a marketing outcome to an AI agent and letting it do the legwork. "Find what customers complain about, draft three angles, show me the evidence." It carries the same risk as vibe coding: the output looks finished before anyone checked it.

This post is a plain version. It explains what vibe marketing is when you remove the hype, a four step workflow that holds up, the jobs where an agent with live data does well, and the checks that keep the result honest.

What is vibe marketing?

Vibe marketing means you state the goal in plain language and an AI agent chooses the tools and steps to get there. You review the output and steer. You do not write each step yourself.

That works for marketing because most of the work is research and first drafts, both of which an agent does quickly. It fails when the agent has no real data, because then it invents. The fix is to give it tools that return facts, and to require a source for every claim.

What does a working vibe marketing workflow look like?

It has four steps, and the order matters. Most bad results come from skipping the first.

1. Collect facts. The agent pulls real data: search results, keyword volume, competitor pages, customer posts. Every row carries its source and the time it was fetched.
2. Find the pattern. The agent reads the facts and proposes what they show. Which questions repeat, which pages rank, which complaint is loudest.
3. Make the asset. Only now does it draft the page, the post or the email, using the pattern and quoting the evidence.
4. Check it. You read the draft against the sources. The agent can help by listing each claim next to its source, so you spot the claims with none.

If you start at step 3, you get generic copy. A model with no data writes what most pages already say, which is exactly what you do not want.

Which marketing jobs work well with an agent?

The ones with a lot of reading and a clear output.

Each has a recipe page with the steps and the providers: the keyword gap recipe, Reddit brand listening and the landing page teardown are good places to start.

What does the data cost?

You pay per call for the data, and the model's own cost comes on top. A keyword research job is cheap because one call returns up to a thousand keywords with volume. The table reads the cheapest priced endpoints from the live catalog when the page loads.

What 1,000 lookups cost explains how to turn those into a job price. A useful default is to cap each job at a small amount and raise it only after you have read the output.

Where does vibe marketing go wrong?

Three places, and each has a cheap fix.

It invents numbers. An agent without a data tool will still give you a search volume. Require a source and a call id next to every number, and delete any number without one.

It sounds like everyone. The draft reads like the average of the web. Feed it your own material: customer quotes, your product's real limits, a founder's rough notes. Specifics are the part a model cannot guess.

It confuses activity with results. An agent can produce fifty pages in an afternoon. That is not a strategy. Pick one keyword, one page and one measure before you scale. The GEO guide is a good example of measuring what you can before you promise a result.

What does one pass look like on a real topic?

Suppose you run a small accounting tool for freelancers and want one blog post that can rank. You tell the agent the topic, "invoice reminders for freelancers", and the rule, "no number without a source".

In step one the agent asks the catalog for keyword volume on thirty related phrases and for the top ten results on the three biggest. It returns a table. You see that two phrases have decent volume, that the top results are all long guides, and that one phrase has volume but only forum threads rank for it.

In step two the agent proposes three angles: a template post, a how-to on polite wording, and a post that answers the forum questions directly. Each angle lists its evidence. You pick the third, because a forum-only result page is the one a short, direct post can win.

Only then does it write. You have spent a few cents on data, read one table and made one real decision. The draft will be better for it, because the agent had a reason for every choice.

How do I start without spending much?

Start with one job and one cap. Pick the job from the table above that you do by hand today, give the agent a limit of fifty cents, and read every row of the first run. If the output is wrong, change the prompt and rerun, since a rerun of a small job costs almost nothing.

After three good runs, widen the job: more phrases, more competitors, more pages. Never widen before the small version works. A small job that fails teaches you the problem. A large job that fails teaches you the invoice.

Keep a short log of each job: the prompt, the date, the cost and what you decided. When you return in a month, you will know what worked, and the next run starts from your notes instead of from zero.

Run it: a prompt for your agent

This prompt runs steps one and two of the workflow on a real question and stops before it writes anything.

When the table looks right, ask for the draft in a second prompt, and tell the agent to quote only from the table.

What it cannot do

• It does not know your customers. It knows what the data says, and you know the rest.
• It does not publish or send anything. The agent drafts and you decide.
• It does not run on a timer. A weekly refresh needs scheduled runs, not available yet, so rerun the research by hand.
• It does not guarantee rankings or traffic. Volume data shows demand, not that your page will win.
• It does not replace taste. The agent can list ten angles, and picking the right one is your job.

For a longer look at an agent doing marketing work inside a coding tool, read Claude Code for marketing. To browse the endpoints these jobs use, open the public catalog.

Last checked 2026-10-03 against the live catalog. Author: Walid Boulanouar.