Two months ago, I wrote a post about how we built an AI agent that helped us hack SEO on TikTok, and even get the LLM-style search responses there to start recommending us. (You find this post here: https://llmmoney.beehiiv.com/p/how-we-hacked-tiktok-seo-and-got-the-ai-to-recommend-our-product-after-failing-to-go-viral)
This post is about what happened next: how we turned that agent into a system that could self-optimize. The agent didn’t just post content, but tracked performance, figured out what was actually driving signups and revenue, and improved itself every day, without us having to guess.
One thing I didn’t fully appreciate until we lived through it was that running organic social media strategies for a product can be deceptively exhausting.
Not because coordinating posting is hard, but because knowing what’s working is.
You post every day, check the views, you see some comments, cross-reference on Amplitude or Google Analytics, you maybe see signups go up… or not,.
And then you try to connect the dots.
That’s where things break down, and where the human brain starts to fail.
The Messy Reality of TikTok/Instagram Attribution
TikTok makes attribution brutal:
You can’t add links to posts.
You don’t even get a profile link until 1k followers.
Discovery happens mostly through search or the algorithm.
Users might see a post today but sign up days or weeks later.
The “normal” playbook with UTMs, links, and funnels just doesn’t exist here.
What we were left with was the daily grind: pull numbers from Socials, pull numbers from Amplitude, stare at dashboards, and guess which posts might be responsible for revenue.
Humans are not just built for this level of repetitive, biased, micro-analysis. And that’s when it hit me: maybe this is exactly the kind of thing an AI agent should be doing.
What If an Agent Did the Boring, Hard Parts?
I didn’t just want AI to write posts, that part is easy. I wanted it to:
Track every post, every day.
See which posts were actually being discovered.
Connect that to signups and revenue.
Pull data from Amplitude and Google Analytics using MCPs, check the customers' LTV, and analyze by cohorts.
Ask the one question humans struggle with: what is really moving the needle?
This is where AI shines since it doesn’t chase vanity metrics. It doesn’t forget what happened two weeks ago. It just looks at the data, figures out what works, and tweaks the strategy.
How It “Got Better Every Day”
Every day, we made the agent spit out trends and insights that it could spot from the whole data, and then we had someone just click apply or reject for the agent to act on those insights. It was also weird that there was no “GIT” for marketing, so we ended up building that as well. That way, we could easily revert those changes in case they caused things to go downhill.
The agent could also see the user journey in Amplitude, the usecase they’re using the product for, map revenue in Stripe, and notice trends humans would never catch consistently. Then it did the obvious thing we struggle to do: double down on what worked and drop what didn’t.
I think the craziest thing we noticed here was that the self-optimizing loop starts to compound really fast, cause this would have meant the team had to meet, review the results from the campaign in spreadsheets, go back and set up the tools with the right next steps, but all of this cycle could now be compressed into a day instead of 3 days.
Where This Is All Heading
With Agents in the picture, the future of marketing is really exciting.
Humans:
Come up with strategies.
Decide which bets are worth making.
Define what success looks like.
AI agents:
Execute those strategies daily.
Track everything.
Optimize continuously.
Possibly assign marketing tasks like changing copies and all to the engineering coding agent.
And eventually prove which assumptions were wrong.
The era of juggling spreadsheets, copying numbers, staring at dashboards, and manually recalculating everything… is ending. Not because humans don’t matter, but because humans can’t hold 10 metrics in working memory simultaneously, and optimization like this was never a human-strength problem to begin with.
The future that excites me is one where we can put the world's best marketing expert in the hands of everyone who has an idea, just like we’ve been able to give everyone a decent software engineer they couldn’t afford to hire with Claude code.
And that’s how an AI agent didn’t just drive $X,XXX in revenue, but it got better every single day doing it.
As always, I am available to chat through this with founders who would like to set something like this up. You can text me on 669-224-8202, or email [email protected]. Genuinely happy to help