Everything on this blog, this post included, comes out of the system I'm about to describe. That's a disclosure, and it's also my excuse for writing in the first person. I built the thing badly once and then rebuilt it, and I watch it run every day, so each failure mode below cost me something.
My view is that most companies with a content engine are really running a conveyor belt. Topics go on at one end and articles come off the other, and nothing those articles do once they're out in the world changes what goes on the belt next month.
A single feedback wire separates the two. Nobody talks about it much, and it's the least glamorous part of the system, but it's also the most valuable. Below I build the loop one stage at a time, then cover where these systems die and the thirty-day version I'd run if I had to start again from zero.
Engine versus conveyor belt
Start with the belt, since that's what I ran first without knowing it. A belt has the three middle stations and skips the first and the last. Its topics come from a brainstorm, or from a keyword dump somebody made back in January, and nothing further down the line ever reports back.
In hindsight the giveaway was our planning meetings. We talked about what to write and never about what had happened to the things we'd already written, because nothing was set up to tell us. If your planning conversation has no scoreboard in it, you're running a belt too.
An engine puts the missing stations back at both ends, which makes it a loop of five. It opens by mining what your buyers actually ask. You plan titles against those questions on a calendar and build each article for extraction, so an AI engine can lift an answer straight out of it. Then you publish natively on your own domain. The fifth station measures, per engine, which answers cite you, which rankings moved and which pages are going stale.
After that comes the wire, where the measurement decides a chunk of next month's plan. Gaps it finds turn into titles. Winners that are decaying get refreshed, and when a citation goes missing, somebody investigates.
The surprise for me was cost. I'd assumed production was the hard part and that whoever cracked article quality would win, and I had it backwards: writing is the cheapest station in the loop. It's also the most automatable of the five, and it got more automatable every quarter. The scarce inputs sit at the two ends, in knowing which questions matter and knowing what happened after you answered them. That's where the judgment lives, and judgment is the one station machines haven't eaten.
Why your own domain, operationally
The philosophical argument for owned content can have its own post. The operational one is short, and it starts with the fact that engines cite pages. Ask ChatGPT or Google's AI Mode a buying question and the references point at durable URLs. Every durable URL your loop produces keeps working while you sleep. It collects internal links and citations, and when it ages you refresh it instead of rewriting it.
Pour the same effort into a social feed and it circulates for two days. That's fine, as long as you know it's doing a different job: distribution circulates the asset, and it should never become the asset.
The compounding part is boring arithmetic. Four durable pages a week comes to two hundred citable assets a year on one domain, and each new page strengthens the internal link graph for all the others. Put the same output across Medium, LinkedIn and a newsletter and you end up with three rented audiences and zero owned authority. I've watched both patterns from the inside, and the gap between them at month twelve is embarrassing.
26 content formats. Published on your domain. Matched to your brand.
Guides, comparisons, listicles, case studies, and more. RankControl generates content that gets cited by ChatGPT, Perplexity, Claude, Gemini, Grok, and Google AI Mode.
The five stations, as actually built
Everything starts with questions, never keywords. We begin from what buyers actually ask rather than from a keyword dump, and the question-inference methods are worth learning because volume tools miss most of what routes B2B deals. What this station produces is a scored pool of question clusters, refreshed every week instead of once a year.
Those clusters turn into titles, and the titles get dates. People underrate the dates. A "publish when ready" pile dies in the first sprint where something urgent lands, while a backlog that's future-dated across a calendar survives your busy weeks. I came to think of the calendar as the engine's flywheel, because it converts a pile of intentions into a cadence, and cadence is the input that compounds.
Production means building every article extraction-first. The question goes in a heading and the verdict sits in the first two sentences, with depth after that and one intent per page. AI does more work at this station than anywhere else in our system, and going by public threads, the same holds for everyone else.
In a recent AMA, a consultant running an AI-scaled blog in e-learning described the setup plainly: seven thousand monthly clicks, custom tooling wired to the CMS, and a human who still does the topic research and feeds the system.
Judgment going in and production coming out is the split you'll find in every version of this that works.
Publishing has to be automatic, native to your CMS and at your own URLs. Copy-paste is where cadence dies. If a person has to move files every time, that person becomes the bottleneck and your calendar turns into fiction. I'd draw the automation line firmly here, too. Pages headed to your own domain can go out on autopilot once the system has earned your trust, but anything outbound under your name, outreach and social posts included, stays reviewed or opt-in.
Measurement is the wire I keep mentioning. Each week, engine by engine, it tells you which of your tracked buying queries cite you and which cite a competitor, what your rankings did, and which pages are decaying. For our own site we track fifty queries across six engines, and the weekly change is what writes next month's plan. Without that station you can't tell an engine that's working from one that's only warm, and most belts stay warm for a year before anybody asks.
What I got wrong the first time
Writing in the first person is only worth it for the scar tissue, so here are the three mistakes from my version one, ordered by what they cost me.
My first backlog came out of a single glorious weekend of question mining, and it was good. The mistake was treating that weekend as the research. I researched once and planned forever, drawing from the pool long after it had gone stale as though it were still fresh. Question pools age faster than keyword lists ever did, since the engines' follow-up suggestions and the communities' complaints change monthly.
What fixed it was making research a drip instead of an event, with a small standing block every week that feeds the same pool the measurement feeds.
I also reviewed for the wrong thing. In version one I read drafts the way an editor would, asking whether each was good, whether it was clear and whether I'd share it. Those are fine questions, and none of them is the one that decides citations.
The review that moved our citation rate asked something much dumber: could you lift the answer to the title from the first two sentences under each heading without losing its meaning? Quality gets a page read, while liftability gets it cited. They overlap maybe eighty percent, and the citations were hiding in the other twenty.
The third mistake was treating the scoreboard as a reward. Measurement felt like something you earn once the engine has produced enough to be worth measuring, so I put it off for two months. That was backwards. It's the cheapest station to build and the one every other station learns from, which means it should exist before your second article does.
Where these engines die
There's a pattern to the graveyard. One of the more useful recent examples is a builder who spent ten years on a hub-and-spoke platform for fully automated AI blogging, a "YouTube for texts," and then wrote the post-mortem himself. The platform never took off. Under his post, the thread fills up with people running fleets of hundred-percent-AI sites that pay a little and teach nothing.
Full automation is the first way these systems die. Every station gets handed to the machine, the two judgment stations included, and it keeps producing, confidently, into the void. Nobody notices that the topics stopped mattering, and when traffic sags eighteen months later there's no instrument to explain why. The fix is choosing which stations get automated. Keep a human hand on topic selection and review until the system earns looser reins.
Volume kills engines too. Speed is measurable and quality isn't, so the belt speeds up and thin variants multiply until a spam update or a quality system does what those systems do. The engines' answer surfaces made this worse by rewarding one page that answers completely over twelve that answer partially, so a volume strategy now loses twice.
Silence is the subtlest death, and it's why an engine that measures gets smarter every cycle while one without the wire just gets older. The engine runs and the articles are decent, but nobody wired up the measurement, so nobody notices which third of the output earns everything. Going by our own numbers, I'd estimate the split really is that lopsided, with a third of pages earning most of the citations. Without the wire you can't find out which third it is, and so you can't do more of it.
Your competitors are getting cited by AI. You're not.
Every day without citation tracking is a day your competitors pull ahead in ChatGPT, Perplexity, and Claude.
The thirty-day version
If I had to start from zero today, on my own, this is the month I'd run:
| Week | Focus | What happens |
|---|---|---|
| One | The question pool | Two hours of engine-suggestion mining, an hour in Search Console, an hour reading the communities where my buyers complain. Fifty question clusters, scored for buying intent, in a spreadsheet. No writing yet. |
| Two | The skeleton and the scoreboard | Run the twenty queries that decide deals across ChatGPT, Perplexity, Gemini and Google's AI surfaces, and log who gets cited (your baseline and your gap list). Write exactly one extraction-first article against the worst gap. |
| Three | Cadence | Four more articles, future-dated across the month |
| Four | Close the loop | Rerun the twenty queries, note any movement, and let the results pick week five's titles |
What week three tests is whether the production station can hold a pace without you staying up late, because a cadence you can't hold is a strategy you don't have. Long-tail citations really can appear that fast, so by week four you may already have movement to note. Once the results are choosing your titles, the belt has become an engine, and you can get there on day twenty-eight with a spreadsheet and an hour.
Don't overcorrect on pace, though. Five articles in thirty days is modest on purpose. I've seen solo founders launch at twenty a week because the tooling technically allows it, and within a fortnight the review station collapses and takes their trust in the whole system down with it. Start at a pace your judgment can actually inspect, and raise it only while the inspection stays easy.
After that, automate the stations in the order they earn it. Production goes first, then publishing, then the data-collection half of measurement, and the judgment stations go last or never. You can put that together from scripts and a spreadsheet or use a platform that runs the loop end to end. The loop is the same either way. The tooling only decides how much of your week it costs, which is why that build-versus-buy call depends more on your hours than on your skills.
The spreadsheet version needs a sheet for the question pool, your CMS's scheduler, one saved browser window with the engines open for the weekly runs, and Search Console. Priced honestly, that's zero dollars and roughly three hours a week, which is what the judgment stations you should be keeping cost anyway. A paid version swaps the browser window and the hand-run queries for automated per-engine tracking, and your drafting hours become review minutes. What it must never replace is the meeting where results pick titles.
One more first-person note, since this whole post has been a disclosure anyway. Most of an engine is plumbing, and the best change we ever made to ours was a pipe. We moved the weekly citation review into the same meeting where we pick titles, so the scoreboard and the plan can't be pulled apart again. No model upgrade or prompt tweak did as much for us, and the review takes ten minutes, in the same tab, every week.
200+ SaaS teams already track their AI citations.
They know exactly when ChatGPT mentions their brand, and when it stops. Do you?
