How to Use AI as a Marketer

In 2024 I wrote a post on this site called "How to Use ChatGPT as a Marketer." It was a long transcript of me asking a chatbot to describe personas, build an AARRR funnel, and write a blog post. It was genuinely useful at the time, and people liked it.

Reading it now is a little embarrassing. Not because the model got worse. Because I was thinking about the whole thing the wrong way. I treated AI like a very fast intern who writes copy. That is the smallest, least interesting thing it can do.

The way I work has changed completely since then. AI is not a tool I open to generate text. It is the layer the rest of my work runs on. It is connected to my ad accounts, my analytics, and my creative pipeline, and it does the parts of the job that used to eat my week. So this is the honest, updated version. Not prompts to copy, but how I actually use AI as a marketer today.

I was using it as a copywriter. That was the mistake.

Go back and look at what everyone did with the first models, myself included. We asked them to write subject lines, ad variants, and blog intros. That was the first thing every model could do and, it turned out, the least valuable. If the whole industry can generate the same email in one prompt, that email is not your edge.

The real unlock was not AI writing marketing. It was AI doing the parts of marketing that were never marketing to begin with. Pulling the funnel. Reconciling two dashboards that disagree. Launching a campaign from a brief. Generating fifty creative variants. Drafting the weekly narrative. That work used to eat half my week, and almost none of it was differentiated.

The tools were never the moat. The judgment about what to look at was the moat. We just paid for the tools, and the hours, because the judgment had nowhere else to run.

Once you stop thinking "AI writes my content" and start thinking "AI runs my plumbing," everything reorganizes.

The setup: how I actually work with it

A few things sit underneath everything else I do.

Claude is a thinking partner, not a search box. I do not open it to get an answer and close it. I think out loud with it, argue with it, ask it to attack my own plan. I talk instead of type, because Wispr Flow turns voice into text as fast as I can think, and thinking out loud is faster than writing.

Context is the whole game. A model with no context guesses, and a confident guess is worse than no answer. So I feed it the real thing: the product, the ICP, past campaigns, what already failed and why. I keep that in a linked, LLM-managed knowledge base in Obsidian, my second brain, so I am not re-explaining my company every morning. Garbage context in, confident garbage out.

MCP is what connects it to reality. The real shift is the Model Context Protocol, an open standard that lets a model reach into my actual systems instead of imagining them. Analytics, the ad accounts, the warehouse, the CRM. A connection is just a small config that points the model at a system it is allowed to use.

{
  "mcpServers": {
    "analytics": {
      "command": "npx",
      "args": ["-y", "@google-analytics/mcp"],
      "env": { "GA4_PROPERTY_ID": "properties/XXXXXXXX" }
    }
  }
}

Once that link exists, the model stops being a clever text box and becomes an operator with hands. Everything below is what those hands can actually do now.

Wiring AI into the ad platforms

This is the part that changed my job the most, and it is very new. As of 2026 the big ad platforms ship their own MCP servers, which means an agent can operate directly on the account instead of me clicking through five different ad managers. The important nuance is how much rope each platform gives you, because it tells you exactly how much to trust it.

Meta Ads

Meta shipped a first-party Meta Ads MCP, and it is the most capable of the three by a distance. It is read and write, authenticated straight through Meta Business login, and an agent connected to it can genuinely run the account:

• Create campaigns, ad sets, ads, and creatives. Everything lands in a paused state, which is the correct default.
• Build and update custom audiences, including lookalikes off a source audience.
• Set up and read A/B tests and lift tests, at the campaign, ad set, or creative level.
• Manage the product catalog for dynamic ads, including feeds and product sets.
• Create and read pixel and conversion events, and check dataset quality.
• Boost an Instagram post, generate an ad preview per placement, and pull insights, benchmarks, and an opportunity score.

The detail I like most is that every write action carries a field that records the human's verbatim intent. Accountability is built into the server, not bolted on. If you do not want to run Meta's own server, the community pipeboard-co/meta-ads-mcp is the popular open-source option and covers campaign and audience work plus reporting.

Google Ads

Google went the opposite way on purpose, and I think they were right. The official google-ads-mcp is strictly read-only. Their docs say it plainly: it cannot modify bids, pause campaigns, or create assets. What it does give you is GAQL, Google's query language, so an agent can pull anything from the account and reason about it.

SELECT campaign.name, metrics.cost_micros, metrics.conversions, metrics.conversions_value
FROM campaign
WHERE segments.date DURING LAST_30_DAYS
ORDER BY metrics.cost_micros DESC

That one query, run by an agent, is a full spend-and-ROAS audit in a sentence. The popular community server cohnen/mcp-google-ads works the same read-only way, and GoMarble's adds keyword-planner research. If you want the agent to actually change the account, that lives in write-capable servers like promobase/google-ads-mcp or hosted products that gate every mutation behind a propose-and-approve step. Read is free. Write is a decision.

Apple Search Ads

If you run App Store campaigns, this matters. There is no official Apple MCP yet, but there are solid community servers wrapping the Apple Ads Campaign Management API, like AppVisionOS/apple-search-ads-mcp, ppcprophet, and gregtuc/asa-mcp. Through them an agent can manage campaigns, ad groups, keywords, and bids, and, the part I use most, pull the search term report and act on it: add converting terms as exact-match keywords, add expensive zero-install terms as negatives.

The landscape at a glance

A workflow that used to be a whole afternoon

Launching and split-testing a sale campaign on Meta now looks like one conversation. Create the campaign with a sales objective and a daily budget, paused. Spin up the ad set and two creative variants. Generate the ad preview for Reels and feed so I can eyeball both before a cent is spent. Set up a creative-level A/B test with cost-per-result as the primary metric, run it as a dry run first, then activate. The agent does the clicking. I do the deciding: the offer, the angle, the audience, and the call on which variant is actually on-brand.

Generating the creative

The other half of paid is the creative itself, and this is where the volume game changed. The model is a phenomenal first drafter and a terrible final judge. It will give me forty variants in minutes. Picking the three that match the brand and the moment is still a human act, and the gap between variant twelve and variant three is where the performance lives.

Here is the pipeline I reach for, depending on the job:

The agentic version of this is simple to describe and genuinely useful. An agent takes a product and a brief, then fans out across dimensions I care about, background, angle, market, hook, and generates dozens of candidates per SKU through Fal.ai, routing each job to the right model. It tags and stores them, and hands me a shortlist. I pick. Two rules keep this sane: generate then filter rather than rendering everything at full fidelity, because video billing per second and image billing per output add up fast, and never ship a variant the model chose. The model has no taste. That is still the job.

The analysis, where it actually earns its keep

This is where AI plus MCP quietly replaced the most tedious, most important part of my week. Every serious analytics platform now ships an MCP server. Google Analytics and Google Ads are read-only on purpose. Mixpanel is read and write and can build funnels, retention, cohorts, and even dashboards. Connect a couple of these and the Monday reporting ritual collapses.

What an agent actually does across them:

• Pull and reconcile the funnel across GA4, Mixpanel, and the ad platforms in one prompt, joining order events to revenue without me knowing which system holds what.
• Investigate anomalies. Not "the number moved," but "here is the endpoint that broke on Thursday, here is why, and here is the query that confirms it." Mixpanel's own team uses their MCP exactly this way.
• Compare cohorts and retention in plain language, so "what is retention for users who finished onboarding" is a question, not a ticket.
• Reconcile cross-channel ROAS, which is the one that saves you from lighting money on fire.
• Draft the weekly narrative on a schedule, pulling the data and writing the story, so no human is compiling a report by hand.
• Recommend budget shifts, then wait. The good pattern is recommend-then-approve: the agent flags which of eighty accounts has negative-keyword waste trending up, proposes the move, and a human ships it.

That reconciliation point deserves a real example. One team found a platform reporting 4.77x ROAS while a geo holdout test proved the actual incremental lift was basically zero. The platform number was confident and wrong. If you let an agent trust a single dashboard, you will scale a lie faster than you ever could by hand. Which is the whole reason for the next section.

Skills: turning a workflow into a button

Once you do the same analysis or the same creative brief for the tenth time, you stop prompting it from scratch and you package it. This is where skills come in, and the distinction between a skill and an MCP is worth getting straight because people conflate them.

• An MCP is a live connection to a system, with credentials. It is the hands. It reads and writes real data.
• A skill is a folder of instructions, and sometimes scripts, that an agent loads on demand. It is the playbook. It carries no credentials and touches nothing by itself. It is the repeatable know-how: which metrics to pull, how to structure the output, the tone, the rules.

The skill orchestrates the MCP. A weekly-ads-report skill defines the metrics, the anomaly thresholds, and the executive-summary-first format. The ad and analytics MCPs actually fetch the numbers. Put them together and a report that used to take an afternoon becomes a sentence, run the same way every time.

The ecosystem is real and public now. Anthropic maintains anthropics/skills as the standard, and there are growing community collections like whyashthakker/agent-skills-marketing with dozens of marketing skills across SEO, content, and lifecycle. Most of them lean toward content and SEO, so the ads and creative skills I actually want, I build myself. The ones that earn their keep for me:

• weekly-ads-report: pull spend, CPA, and ROAS across the ad-account MCPs, flag anomalies, output an exec summary.
• creative-brief: take a product and audience, return objective, angle, hooks, formats, and CTAs, ready to feed the creative pipeline.
• seo-audit: crawl a URL, check titles, headings, and Core Web Vitals, output prioritized fixes.
• keyword-clustering: group a keyword list by intent for ad-group or content structure.
• competitor-teardown: scrape a competitor and compare positioning and messaging.

Each is a small playbook that calls whatever MCP has the live data. That is the compounding part. Every workflow you package is a thing you never have to think through again.

The rest of the funnel

Ads, creative, and analysis are where AI shows up loudest, but it is everywhere else too.

• Research and positioning. I do not ask for invented personas anymore. I hand the model the real mess, sales-call notes from Granola, support threads, churn reasons, reviews, and it clusters hundreds of inputs into patterns. I decide which pattern is the wedge.
• Activation and onboarding. The agent reads the funnel in Mixpanel and the session recordings in Microsoft Clarity, tells me where users actually quit, and drafts the fix.
• Retention and lifecycle. Lifecycle still runs on a real platform, Insider in my case, but the agent drafts the sequences, the segment logic, and the win-back flows, and I edit for judgment.
• Outbound. Apollo supplies the data and the agent researches and personalizes at a volume I could never hit by hand, without sounding like a mail merge.
• Building. With Claude Code I ship landing pages, tracking, and internal tools by directing an agent instead of waiting on an engineering queue. A marketer who can build is a different animal now. I wrote about vibe coding for marketers separately.

The before and after

I do not believe in charts that pretend everything got replaced. Here is what genuinely moved.

The pattern is not "AI replaced marketing." Retrieval, synthesis, and execution collapsed in cost. Judgment did not. Everything the model made cheap was the part that was never the moat.

The principles that keep me out of trouble

I learned most of these the expensive way.

• Connect read, gate write. Let agents read everything and write nothing without a human in the loop. This is why I am glad Google's server is read-only and Meta's records intent on every write.
• Dry-run and propose before you spend. For anything that touches budget, the agent proposes and I approve. Never a blind mutation on a live account.
• Verify against a second source. The number comes from a query, not the model, and a number that matters gets checked before I believe it.
• Own the output. If the agent drafts a wrong narrative and you ship it, that is your wrong narrative. The byline does not transfer to the tool.
• Feed context, do not expect magic. The quality of what you get back is a mirror of what you put in.

What it still cannot do

When everyone has the same agents connected to the same data, the analysis stops being the edge. Two teams with identical setups will pull the same funnel and surface the same anomalies. The work that used to take a week and felt like progress is now a sentence.

So the edge moves up, to the two things that were always hardest and always undervalued. Asking the question worth answering, because the model will answer anything and has no opinion about which question matters. And having the taste to know which answer is right and which is confident nonsense. On top of that sits the work no model will ever do for you: walking into a room and convincing a company to actually ship the risky thing.

That was always the real job. There is just more of it now.