Marketing engineering, ranked by evidence: which growth loops are actually worth automating
Anyone can build software now. That was true a year ago and it is more true today. What has not changed is that nobody shows up to use it.
I learned this the expensive way with StudentGrounds, a college social network I built with zero distribution plan. The product worked. Nobody came. Everything I have built since starts from the other end: who is going to find this, and through what channel.
So when Greg Isenberg put out two episodes back to back on "marketing engineering," I pulled both transcripts out of my data lake and sat with them. Then I ran a research pass on every indie hacker growth tactic I could find real numbers for. This post is the result: a frame for what a marketing agent actually is, the two concrete agents Cody Schneider walked through, and thirteen growth loops ranked by evidence quality and automation potential.
If you only read one section, read the table.
The frame: signal in, pipeline out
Greg's episode on the marketing engineer role (What Is a Marketing Engineer?, Aug 31) lays out four eras. Traditional marketing made people care. Digital marketing acquired customers through channels you could measure. Growth hacking pulled marketing into the product. Marketing engineering is the fourth era: using agents, data, code, and taste to build a marketing system that keeps learning.
His one line definition is the part worth memorizing. A marketing engineer turns market signal into pipeline.
The tactical claim underneath it is that most companies already have the signal. Sales hears one version of the market, support hears another, product sees usage, marketing sees clicks, and the founder remembers the one call that hit emotionally that week. Everyone walks into the growth meeting with a different reality. The job is to pull those into one place.
Three ideas from that episode map directly onto how I want to build.
The growth repo. Before any agent, a folder. Customer truth (call notes, tickets, churn reasons), founder voice (hooks that worked, banned phrases), outbound (ICP, trigger events, angles that got replies), creative tests, and agent job specs. The point is memory. Most people use AI in disposable chats, so every week the model starts from zero. A repo means the prompt becomes "read the customer truth file, read the last five posts that got qualified replies, draft five around the pains buyers mentioned this week."
The job spec. Every agent gets written up like a hire: data source, run cadence, what to filter out, expected output, what good looks like, what needs human approval, the metric that matters, and where results get written so the system learns. The metric point is the one people skip. Messages sent is activity. Qualified replies is signal.
The example workflow. Beginner version: ask AI to write a post about a keyword. Marketing engineer version: read Search Console, pull keyword data from Ahrefs, check the CMS for an existing post, rank by volume and buyer intent, read what currently ranks, add the founder's POV, draft the post plus meta title plus internal links, send for approval. Same LLM, completely different output, because the inputs come from the business.
Every tactic below fits that same skeleton. Signal source, filter and rank, draft with founder POV, human approval, outcome log. Build the skeleton once and each playbook becomes a config.
Two agents, concretely
The second episode (These AI Marketing Agents Get You Customers, with Cody Schneider) is the hands on version. Two agents, every tool named.
Agent one: LinkedIn engager outbound. Cold email reply rates are down across the board because AI slop has flooded every inbox. Cody's answer is to stop targeting firmographics and start targeting hand raises. Someone who liked or commented on a post about your category just told you they care about that topic.
The loop:
- Pick 10 to 20 creators or company accounts your ICP engages with. A handful of outliers get most of the engagement in any niche, so you get most of the coverage from a small watchlist.
- Daily cron: scrape net new posts from those accounts (Apify, specifically the API Maestro LinkedIn actors), then scrape reactions and comments on each.
- LLM step: research each person and company, drop the ones that do not fit the ICP.
- Waterfall enrichment: GetLeads first (cheapest), then Apollo for the misses, then something like Origami or Prospeo for the remainder. Verify with Million Verifier before sending anything.
- Send cold email from burner domains via Instantly or similar, LinkedIn DMs via HeyReach or BotDog. Never from your core domain. He separates cold, marketing, transactional, and business domains entirely.
- Webhook on positive reply back to a small hosted agent whose job is to answer questions and push toward a booked call. Plug in Cal.com so it can see whether the meeting actually happened.
Infrastructure cost to start is roughly two hundred dollars a month.
Agent two: transcript to content engine. Record a weekly conversation with each person on the team (or use sales calls, Slack, Gong). Extract insights. Write posts. Schedule to each person's LinkedIn via Ordinal's MCP. Pull the post analytics back into the loop so the next batch is weighted toward what got impressions.
The reason source material matters: if you ask a model for "good LinkedIn content" with no input, you get the most mid thing imaginable, and LinkedIn now flags it. Real human conversation is where the non-obvious ideas are. A prospect saying why they did not buy is a better post than anything the model invents.
Cody's framing of what an agent even is stuck with me. It is code, maybe a thinking loop, and a live data stream. Nothing more. And the corollary: do not pay tokens every time an action runs. Pay for the model to write the software, then let CPU run the software. Only use inference where judgment is required. Most "agent frameworks" are bloat for problems this narrow.
That is almost exactly how I already think about VibeDraft and the data lake, so hearing it said plainly was useful.
Thirteen loops, ranked by evidence
Here is where I wanted more than podcast confidence. I ran a deep research pass looking for documented numbers per tactic: MRR attributed, conversion rates, reply rates, timelines. Two biases show up everywhere and I want to name them before the table.
First, most "playbook" numbers are published by vendors selling the tool in question. Social listening stats come from social listening tools. Dunning recovery stats come from dunning tools. Second, founder MRR is self reported and unaudited. The most defensible numbers came from Indie Hackers post mortems, Hunter.io and Belkins cold email datasets, PostHog's own writeups, and the Northwestern Spiegel reviews study.
| Tactic | Evidence | Time to result | Agent can run | Platform risk |
|---|---|---|---|---|
| Intent listening and reply | High | Days | Draft only, human ships | High |
| Competitor "alternative" pages | High | 2 to 6 months | Most of it | Medium |
| Programmatic SEO / free tools | High | 6 to 12 months | Most of it | Very high |
| Build in public / reply guy | High (self reported) | Months to years | Assist only | Medium |
| Launch stack and directories | High | Launch day | Most of it | Medium |
| Signal based cold outreach | Medium | Days | Most of it | High |
| Lifecycle and churn save email | Medium (vendor, large N) | Immediate | Nearly all | Low |
| Testimonial harvesting | Medium | Days | Most of it | Low |
| AEO / GEO | Low to medium | Weeks to months | Most of it | Medium |
| Content repurposing | Medium | Weeks | Most of it | Low |
| Open source as distribution | High (dev tools) | Months to years | Some | Medium |
| Newsletter sponsorship and affiliates | Medium | 2 to 8 weeks | Some | Low |
| HARO, giveaways, roast threads | Anecdotal | Varies | Some | Medium |
The numbers behind the rankings:
Intent listening. Octolens analyzed 522M Reddit mentions and found about 23% of B2B SaaS discussions show active buying intent, and authentic replies get roughly 8x the engagement of self promotional ones. Responses after 24 hours get buried. One documented case: an email tool founder monitored Mailchimp complaints, replied to 8 threads in a week, and three became paying customers within 30 days. Small numbers, high quality. This is the channel where the human has to press send.
Competitor pages. The strongest indie SEO evidence I found. Fifteen comparison pages took one company from 47 to 342 trial signups a month, with the best page converting at 13.8%. Another ranked first for "[competitor] alternative" and converted at 23% to trial, about 4x their normal landing page. This is the tactic closest to Greg's SEO workflow, and it is highest intent traffic in any category.
Programmatic SEO. Works when the pages are backed by a real tool or real data. HubSpot's Website Grader, Canva's background remover pulling 3M visits a month, Marc Lou's free logo tool funneling into ShipFast. Fails badly when thin: the 2024 helpful content changes reportedly cut G2 and ZoomInfo's organic traffic by 76 to 99% and deindexed 98% of one travel site's 50,000 pages within three months. Demotion is now sitewide and continuous.
Build in public. The highest lifetime ROI channel for a founder ICP and the least automatable. Pieter Levels' "$1M ARR in 17 days" and Photo AI's climb to $132K MRR sit on a decade and 600K followers. Marc Lou posts about 15 times a week with a median of 53 reactions; it is a volume game. Tony Dinh went from 100 to 180K followers and sold two products for $128K and $150K along the way. An agent can source ideas, schedule, and repurpose. It cannot be the voice.
Launch stacks. Product Hunt is a backlink and credibility play, not a revenue play. An OpenHunts study of 387 launches found PH converts around 3% versus 23% for engaged Indie Hackers posts, and 89% of founders said they would not launch there again. A spring 2026 roundup of five founders showed rank did not predict revenue at all. The part worth automating is the directory sweep: 40 plus hours of manual submissions that yield 10 to 15 backlinks.
Signal based outreach. Generic cold email replies sit around 3 to 4% (Hunter.io, Instantly). Two custom attributes lifts that by half. Three touch sequences double it. Signal triggered campaigns report 15 to 25% in agency writeups and a more conservative 4 to 8% in SaaS specific data. This is Cody's LinkedIn engager loop with numbers attached. The caveat from an Indie Hackers post mortem: one founder got a 37% DM reply rate and zero sales. Replies are not revenue.
Lifecycle and churn save. The single lowest risk automation on the list. Baremetrics' May 2026 benchmark across 119 B2B SaaS companies shows median dunning recovery around 48% and median ROI of 808%. Churnkey reports 70% of involuntary churn recovered on average. Vendor data, but large N, and dunning is safe to run without a human in the loop.
Testimonials. The one strong independent number: Northwestern's Spiegel Research Center found purchase likelihood with five reviews is 270% higher than with none. Widget lift claims from testimonial tools are directional at best.
AEO/GEO. Early and volatile. ChatGPT was 87% of AI referral traffic in late 2025 and is fragmenting fast. The interesting number is conversion: AI referred traffic converts 4 to 5x organic across industries, and one B2B SaaS portfolio measured 14% versus 3%. Cheap to add llms.txt and structured comparison content now. Hard to attribute yet.
Open source. PostHog's "open source product analytics" Show HN got them 300 deployments in days and they rode "open source alternative" searches to $50M ARR. Plausible crossed $2M revenue. Very high ROI for dev tools, not applicable unless you open source a component.
The rest (repurposing, newsletter sponsorships, HARO, giveaways, roast my landing page) have mechanism evidence and thin attribution. Newsletter CPMs for good B2B lists run $25 to $100 and last click undercounts them 2 to 5x. Everything else is anecdote.
The two things every number agrees on
Distribution compounds over years. No agent manufactures a 600K follower audience. What an agent does is increase at bats and cut the time cost per reply, per page, per submission, so the compounding starts sooner and runs faster.
And every public facing action needs a human gate. Reddit bans, X link and slop penalties, Google demotions, directory spam traps: all of them are triggered by unsupervised posting. The agent's safe zone is finding and ranking and drafting. The human ships. Greg's job spec has a "needs human approval" field for exactly this reason and it is not optional.
What I am building first
I run VibeDraft, a content pipeline for founders on X. My ICP is founders building a product to sell. So the question is which of these loops I can build once, dogfood on my own accounts, and hand to that ICP.
Stage one: intent listening plus reply drafting. Reddit, HN, and X monitoring for problem keywords and competitor complaints, intent scoring, a reply drafted in the founder's saved voice, approval, outcome logged. It overlaps VibeDraft's existing drafting engine and it is the tactic with the fastest feedback. If founder approved replies convert at even 1 to 3% thread to signup and save half an hour a day, it earns the next stage.
Stage two: competitor page generator. This is Greg's SEO workflow almost verbatim, pointed at "[competitor] alternative" and "A vs B" queries. Ship five to ten pages, watch indexation, only scale if they rank without tripping the helpful content demotion.
Stage three: the boring high leverage stuff. Directory submission drafting and build in public repurposing. Low risk, reinforces the core product.
Stage four: dunning and win back. As soon as failed payment volume is worth recovering.
I am deferring programmatic SEO at scale, high volume cold outreach, and anything AEO beyond llms.txt. Not because they do not work, but because the failure modes are sitewide and reputational and I would rather earn the right to those.
One more thing, mostly a note to myself. I have a pattern of building infrastructure ahead of demand. The honest version of this plan is that stage one gets run by hand for two weeks before any of it becomes code. Greg's own 30 day plan puts "audit a real company" in week one and "first machine" in week three for the same reason. Learn the workflow first. Then automate the part you have already done ten times.
If you are building something similar, or you have real numbers on a tactic I marked thin, reply to this email. That is the whole point of the customer truth folder.