Build Your Own Autonomous Marketing Agent

What it does (and doesn't do)

The agent runs a three-step cycle, either on a weekly schedule or when you @mention it in Slack.

1. Sense. Pull traffic from your analytics, search performance from Google Search Console, and referral sources. Flag anything that moved more than 20% week on week. Check Google Trends to see whether your target queries are growing or shrinking as a category.
2. Maintain. Search the web for product mentions, and extract the "People Also Ask" questions for your target queries (these are the questions AI systems actually answer, so they make good article headings). Then hit the OpenAI, Claude, and Perplexity APIs directly to see whether they recommend your product when someone asks about your category. Call it your AI visibility, and it increasingly decides whether buyers find you at all.
3. Ideate. Propose up to three things to do this week, each with a data-grounded reason and an effort estimate.

It does not write, publish, or send anything. It observes and suggests, you decide what to act on. An earlier version had a content-creation step, dropped because the editing overhead ate the time savings.

Before you start

Services you'll need

Most of these have free tiers that comfortably cover a weekly run.

You'll also need a Slack workspace for the agent to post to and take questions from.

On the analytics tool: pick whatever you already use, as long as the agent can query it. Google Analytics 4 has an official MCP server. The examples here use Lodd, an MCP-first analytics tool I built. Plausible has an API you can wire up over HTTP, but no MCP server yet.

Choose a model that can tool-call

Settle this before anything else in the build. The wrong model here doesn't fail loudly, it fabricates.

The agent is a multi-tool loop: it decides which of eight tools to call, calls them, reads the results, and calls more. That ability is tool calling, and not every model does it well. It tracks how hard the lab tuned the model for agentic use, not how clever the model sounds or how big it is.

Do this:

• Pick a model that scores well on public function-calling benchmarks. A current Qwen flagship (qwen/qwen3.7-max) is a solid default, near the top of the benchmarks and cheap on output tokens. Claude Haiku and the GPT-mini tier also tool-call reliably. Rock-bottom free and quantised models mostly don't.
• Verify before trusting any marketing page. OpenRouter exposes a supported_parameters list per model, you want tools in it.
• At one weekly run plus a few ad-hoc questions, the cost gap between capable models is cents. Optimise for tool-calling reliability, not token price.

Step 1: Confirm your analytics has an MCP server (or an API)

The agent needs to reach your analytics programmatically. The cleanest path is an MCP server it can point at, which auto-exposes the tools. If your analytics doesn't have one, a plain REST API works too, you just wire it up by hand. Either way it has to be able to ask "how many visitors this week, which pages, and where from?" and get an answer back.

• GA4: its MCP server handles this.
• Lodd: MCP endpoint https://api.lodd.dev/mcp with Bearer token auth.
• Anything else: check for an MCP server first (easiest), otherwise a plain REST API you wire up with an HTTP Request node (more manual). Either works.

Step 2: Collect your API keys

Gather these before opening n8n:

1. Analytics token. For Lodd: Dashboard → Settings → API keys.
2. Google Search Console. Verify your site at search.google.com/search-console if you haven't.
3. Tavily API key. From your tavily.com dashboard.
4. Supadata API key. From supadata.ai.
5. OpenAI API key. platform.openai.com → API keys → Create new.
6. Perplexity API key. docs.perplexity.ai → Settings → API keys.
7. OpenRouter API key. openrouter.ai → Keys. This is the agent's brain, and it also powers the Claude visibility check. You can use a provider's own key instead, but OpenRouter lets you swap models by changing one string, which is handy while you're finding one that tool-calls well.
8. Slack bot token. Needs a few more steps, see Step 3.

Step 3: Set up the Slack bot

Go to api.slack.com/apps and create a new app. Name it something like "Marketing Agent" and select your workspace.

1. Under OAuth & Permissions, add these bot token scopes:
   • app_mentions:read
   • channels:history
   • channels:read
   • chat:write
2. Under Event Subscriptions, enable events and subscribe to app_mention. The request URL is your n8n webhook URL, which you'll get in Step 5, so you may need to come back here.
3. Install the app, copy the Bot User OAuth Token, and invite the bot to the channel where you want reports.

Step 4: Import the workflow into n8n

Grab the finished workflow from the Get the build panel (or download the JSON directly). It's the real workflow, stripped of credentials and personal values, so you bind your own keys in Step 5 and swap the placeholders ([YOUR PRODUCT], [YOUR CATEGORY], [YOUR TARGET QUERIES], YOUR_SLACK_CHANNEL_ID) for your own.

In n8n, go to Workflows → Import from File and select the JSON. You'll see the full workflow with two trigger paths feeding into the AI agent.

Step 5: Configure the workflow

The workflow has a model node plus several tools that each need credentials. Bind them one at a time and test after each.

The model node

Connect an OpenRouter Chat Model node (or your provider's node) to the AI Agent. Paste your OpenRouter key and set the model to one that tool-calls well (e.g. qwen/qwen3.7-max). See "Choose a model" above for why this is the highest-stakes choice in the build.

MCP tools (analytics, Tavily, Supadata)

Each connects via n8n's MCP Client node. Add the endpoint URL, set authentication, and n8n auto-discovers the tools.

• GA4: follow Google's MCP setup guide. OAuth auth.
• Lodd: endpoint https://api.lodd.dev/mcp, Bearer token auth. 42 analytics tools auto-discovered.
• Tavily: endpoint https://mcp.tavily.com/mcp/?tavilyApiKey=YOUR_KEY. Gives web search and People Also Ask extraction.
• Supadata: endpoint https://api.supadata.ai/mcp, Header auth (X-API-Key). Handles scraping and YouTube transcripts.

Google Search Console

Two routes, depending on how you host n8n.

Self-hosted: use mcp-gsc, an open-source GSC MCP server exposing search analytics, URL inspection, indexing checks and sitemap status. It's a command-line (stdio) server, so: create a Google Cloud service account with Search Console API access, download its JSON key, add the service account's email as a user on your GSC property, clone mcp-gsc and point it at the key, then add an MCP Client node using stdio transport. Cloud n8n can't reach a stdio server, so this is self-hosted only.

n8n Cloud: skip the MCP server and call the Search Console REST API from an HTTP Request tool. Configure it like this:

• Method/URL: POST https://searchconsole.googleapis.com/webmasters/v3/sites/sc-domain%3Ayourdomain.com/searchAnalytics/query
• Property identifier: sc-domain:yourdomain.com is a domain property (the %3A is the URL-encoded colon). For a URL-prefix property, use the encoded https://yourdomain.com/ form. Check which you have in Search Console.
• Auth: generic credential type → Google OAuth2, scope https://www.googleapis.com/auth/webmasters.readonly.
• Body (JSON), using n8n date expressions so the window is always current:

{
  "startDate": "{{ $today.minus(28, 'days').toFormat('yyyy-MM-dd') }}",
  "endDate": "{{ $today.toFormat('yyyy-MM-dd') }}",
  "dimensions": ["query"],
  "rowLimit": 25
}

Once connected, the agent can pull impressions, clicks, CTR, and average position by query and page. High impressions with low clicks usually means Google shows your page but the title or description isn't working.

Google Trends

No official API, and this is the one tool to treat as optional.

• Self-hosted: the free pytrends library. n8n's Code node can't pip-install it, so run it via an Execute Command node calling a small Python script, or wrap that script in a tiny HTTP endpoint.
• n8n Cloud: SerpApi's Google Trends API via an HTTP Request tool. Needs a key (httpQueryAuth credential, key goes in as api_key). The free tier covers a weekly run.

Either way the agent asks the same thing per target query: is interest UP, DOWN, or FLAT over the last 3 months. That's worth having because it separates "my site is declining" from "the whole category is declining", which are different problems with different fixes. But it's the flakiest tool, so if you want to ship faster, leave it out and add it later.

AI visibility checks (OpenAI, Claude, Perplexity)

Three more HTTP Request tool nodes, calling the chat completions APIs directly to see whether the assistants mention your product. Claude reuses your existing OpenRouter key, so it needs no extra credential.

OpenAI:

POST https://api.openai.com/v1/chat/completions
Authorization: Bearer YOUR_OPENAI_KEY

{
  "model": "gpt-5.4-mini",
  "messages": [{
    "role": "user",
    "content": "What [your category] tools exist? List specific products."
  }]
}

Perplexity:

POST https://api.perplexity.ai/chat/completions
Authorization: Bearer YOUR_PERPLEXITY_KEY

{
  "model": "sonar",
  "messages": [{
    "role": "user",
    "content": "What [your category] tools exist? List specific products."
  }]
}

Claude (via OpenRouter): no new key, reuse the OpenRouter credential from the model node. Pick a current Claude slug from OpenRouter's model list.

POST https://openrouter.ai/api/v1/chat/completions
Authorization: Bearer YOUR_OPENROUTER_KEY

{
  "model": "anthropic/claude-sonnet-4.6",
  "messages": [{
    "role": "user",
    "content": "What [your category] tools exist? List specific products."
  }]
}

Replace [your category] with what your product does: "web analytics tools", "equipment tracking software", "teleprompter apps for Mac", whatever fits.

Memory

Connect a Simple Memory node to the agent's Memory input, set to the last 10 messages. This gives the agent context for the Slack @mention path, so you can ask follow-ups like "tell me more about that traffic spike" and it remembers what it just reported. The weekly scheduled run starts fresh and doesn't need it.

Output routing

The agent's output goes through an If node that checks which trigger started the run. Set up two Slack Send Message nodes connected to the If node's true/false outputs: scheduled runs post to a dedicated weekly-report channel, ad-hoc questions reply in the channel where you asked.

Slack nodes

Paste your Bot User OAuth Token into the Slack Trigger node and both Slack Send Message nodes. Set the channel IDs for weekly reports and ad-hoc replies (these can be the same channel).

The agent prompt

The Edit Fields node after the schedule trigger holds the weekly prompt. Customise it for your product:

You are a marketing agent for [YOUR PRODUCT]. Run the following cycle.

TOOL USE IS MANDATORY. You have these tools: [list them]. You MUST
actually call the relevant tool for every section below. Do NOT answer
from memory or prior knowledge. If you have not called a tool, you do
not know the answer.

Every factual claim MUST be backed by raw evidence from a tool call:
an exact number, an exact date range, or a source URL. Bare statements
like "traffic is up" or "visibility unclear" are forbidden.

If a tool was not called, returned nothing, or errored, say so with the
literal token TOOL NOT CALLED or NO DATA for that line. Never paper over
a missing tool result with a guess.

SENSE:
- Call the analytics tools for the last 7 days vs the previous 7 days.
  Report exact visitor/pageview counts, which pages grew or declined
  (with numbers), new referral sources, and event counts.
- Call Google Search Console for the last 28 days. Report top queries
  by impressions with exact impressions, CTR, and position. Flag any
  page with high impressions but low clicks (title/description problem).
- Call Google Trends for: [YOUR TARGET QUERIES]. Report UP, DOWN, or
  FLAT over the last 3 months, citing the tool result.

MAINTAIN:
- Call the web search tool for mentions of [YOUR PRODUCT] in the last
  week. Include source URLs, or say NO DATA.
- Call the web search tool for the "People Also Ask" questions for the
  target queries. List the actual questions and flag which are not yet
  covered by existing content.
- Call the OpenAI tool: "What [YOUR CATEGORY] tools exist? List
  products." Quote the relevant part and state whether [YOUR PRODUCT]
  appears.
- Call the Perplexity tool with the same question.
- Call the Claude tool (via OpenRouter) with the same question.
- For any competitor claim, include the source URL or mark UNVERIFIED.

IDEATE: Based ONLY on the tool results above, propose up to 3 actions.
Each: what to do, why (citing a specific number or finding), effort
(quick/medium/significant). PAA questions not covered by existing
content are strong article candidates.

Rules: Lead with findings, not process. Every claim carries its
evidence. If a section has nothing to report, say "all clear", but only
after the tool was actually called. Under 500 words.

Update the channel ID and source field in the same Edit Fields node.

Step 6: Set the schedule and test

Set the Schedule Trigger to run weekly. Saturday mornings work well, so the report is waiting at the start of the week. Then run a manual test and confirm the agent:

1. Pulls analytics (real traffic numbers)
2. Pulls search data from GSC (impressions, clicks, top queries)
3. Reports trend direction from Google Trends
4. Extracts People Also Ask questions
5. Checks ChatGPT, Claude, and Perplexity for citation visibility
6. Posts a report to your Slack channel

If something fails, it's almost always a credential issue. Check the API keys, the MCP endpoints, and the Slack bot scopes.

Step 7: Ask it questions

Once the weekly cycle works, @mention the bot in Slack:

@marketing-agent what drove last Tuesday's traffic spike?

@marketing-agent compare our blog traffic to last month

@marketing-agent are any competitors running a sale right now?

The agent uses the same tools for ad-hoc questions as for the weekly cycle, and the Simple Memory node keeps the last 10 exchanges, so you can have a back-and-forth.

Customising for your product

The workflow is a template. The things you'll want to change:

• Target queries. The citation checks ask ChatGPT, Claude, and Perplexity about your category. Pick the 3-5 queries your ideal customer would type into an AI assistant ("best [category] for small teams", "[competitor] alternative", "how to [solve the problem]") and put the most important one in the HTTP tool prompts.
• Competitor list. Name specific competitors in the agent prompt so it knows what to watch for. Always require source URLs, mark anything it can't cite as UNVERIFIED. An early version attributed a competitor's negative review to the wrong product because two names were similar.
• Anomaly threshold. The prompt flags changes over 20%. Under 100 visitors/week, raise it to 30-40% so you're not alerted on normal variance.
• Report format. "Under 500 words" keeps reports scannable. Keep it tight, you can always ask follow-ups in Slack.

Going further

• Content creation. Dropped from this build, but if your product voice is straightforward and you're happy editing AI drafts, add a create step that writes posts from the ideation output. PAA questions make good headings.
• SEO health audits. Crawl your sitemap weekly, check every page for title tags, meta descriptions, JSON-LD, and heading hierarchy. Auto-fix the mechanical stuff, flag the rest. This works well as a separate workflow.
• Multi-product. Duplicate the workflow per product: different Slack channel, different analytics site, different target queries. The template is the same.