---
title: "What it took to build a self-learning content engine — Atlan Frontier Labs"
url: "https://atlan.com/frontier/essays/a-self-learning-content-engine/"
description: "Between 2022 and 2026 we rebuilt our content engine four times. What each version cost us, what it bought, and what the fourth one — Helix — now does on its own."
---

> Atlan is hosting Context Conference, bringing together the leaders and builders at the frontier of giving AI the context it needs to understand their business. It runs online on October 28, 2026, from 11:00 AM to 2:00 PM ET. Atlan co-founder Prukalpa Sankar opens and closes the day. Leaders from AstraZeneca, BNY and Verizon share why they invest in context and what they get from it. Registrants get early access to The AI Context Gap, a new study from MIT Technology Review Insights. Register: https://atlan.com/context-conference/

An essay from Atlan Frontier Labs (Operating Notes, essay 005) by Labani Biswas, Content, Atlan, dated 2026.08.25, 2,482 words. Between 2022 and 2026 Atlan rebuilt its content engine four times: what each version cost, what it bought, and what the fourth one, Helix, now does on its own. Part of [Becoming Frontier](https://atlan.com/frontier/).

## Intro

Human civilization built institutions, processes, even the idea of value, on the assumption that intelligence is scarce. Now it is available on tap. GPT-3.5 was magical, yet stupid enough to elicit one of two reactions:

- Reaction one: AI will always be a plaything; it will fail to seriously disrupt the physics of our work or drive true outcomes.
- Reaction two, more common: this is great, we can do things a lot faster, cheaper, but perhaps never better.

At Atlan, Moore's Law had dictated too much to bet against it. Organic search was by far the most predominant, efficient and scalable channel, and increasingly the most disrupted. What kept the team from being paralyzed was one enduring mission: helping humans of data (and now agents) do their lives' best work. Between 2022 and 2026 the content engine was fundamentally reimagined at least four times. The essay is also about the team's assumptions, unconscious biases and "the mythical power of saying 'Why the hell not?'"

## The only lever to organic distribution is value to the reader

The job is simple: find out what people care about, help them understand it, form an opinion about it, give them perspective or courage toward their goals. Teams hyperobsess over 200+ Google ranking signals, feed changes on Twitter/LinkedIn, and now query fan-out and chunking for LLM search, but all these platforms are geared toward engaging readers with what they care about. That is the guiding principle behind Atlan's content function.

The audience wants to understand things, clarify thinking, learn the product and connect it to their systems. Atlan must reach them where they ask and serve intellectually honest answers, unique perspectives and helpful examples; the essay calls this one of the most consequential bets the marketing function (even the company) has made. The biggest change: readers moved from Google's 10 blue links to chat interfaces (ChatGPT, Claude, Gemini, Perplexity and others) that give answers, not a bridge to answers.

Running now: tests of whether the answer engines cite Atlan when a reader asks how to make AI understand their data. Those results feed back into what Helix plans next.

## V1: Quality control of thought, not words (Q1 2022)

First insight: understanding the why, the intent, of a piece is the limiting factor to success; everything else is downstream.

V1 was one person working with external partners (agencies and consultants), with a strict check at the outline phase: no word written until the team aligned on what to write, the flow of information and the value of the piece. This often stalled publishing, and pieces stuck at outline, but it was necessary. "Getting some traffic on the website was not the goal — getting the right people to trust you and see value in our content was."

That was Q1 2022; it set up an advantage for a world where words would be unlimited but insight, perspective and understanding would stay constrained. On December 1st that year, Sam Altman announced ChatGPT.

## AI-assisted → AI-augmented → AI-native

### V2: AI-assisted (April 2023)

Built with GPT 4o. The team wanted to believe GPT generated gibberish, but the results were undeniable. A parallel pipeline on the same principles: humans controlled what to write and defined the value; GPT 4o wrote the words. Hallucination was still a big problem, so external partners moved from writers to editors.

Results: tripled the highest-ever production volume at one-fifth the initial cost, gave Atlan its first hockey-stick graph in qualified traffic, and fed the business pipeline with ICP personas for at least 6 quarters after. Website traffic by year (figure): 176,468 visitors in 2021, 452,276 in 2022, 878,000 in 2023, just shy of a million in the first full year of the AI-assisted pipeline. That content still stands after several Google helpful content updates.

### V3: AI-augmented (Q2 2025)

V2 was bottlenecked by humans' ability to package context: value sits at the intersection of market, customer and product understanding, and a human had to hand-stitch it for an LLM. By Q2 2025 there were better tools. The team moved to Claude Projects: custom instructions with strict dos and don'ts and everything known about ranking on Search, project files with everything a human would read before outlining, and a search tool call to the company knowledge base (Glean).

"V3 felt like going from a bullock cart to a rocket." Multiple back-and-forth prompts became one: generate an outline for topic x. Humans could feed 10-15X as many external partners at once. Words were in free supply, so there was no longer a need for people who are both technical and good writers; subject matter experts pack in insight and lived experience, and the system handles editing and optimizing for visibility and distribution.

### V4: AI-native (March 2026)

Between November and December 2025, coding agents went mainstream and grounding LLMs in context leapt ahead (MCPs, skills and more), while the economics of digital marketing stopped working. Atlan set up Marketing OS ([explained here](https://atlan.com/frontier/essays/we-rebuilt-marketing-around-it/)): a context repo where strategy and knowledge are versioned and expertise is encoded as skills agents run on, so anything built starts out knowing what the team knows. In March 2026 marketing was organized around Marketing OS, and V4, Helix, followed soon after.

The value of content when intelligence is on tap still has two components: (1) credible, helpful information, ideally with an original PoV or unique insight, and (2) most importantly, being the answer when and where the reader looks. The engine was rebuilt around using Atlan's understanding of why a topic matters, all its collective knowledge, and existing credible literature, and around progressively cracking distribution via LLMs.

## Helix is the name for the process of making ourselves redundant

Helix is both a group of skills and the agent that executes them. It is not built to assist or augment the content engine: "It is built to become us." Architecture (figure): three inputs feed Marketing OS, which holds Helix and its phases with an approval gate between each; a market research agent hands it briefs, a social agent takes the published page onward; human approval comes before a live page, and the live page loops back into the inputs. Nothing goes live without a human.

It publishes any cohort of pages through six phases, each a gate a page can be sent back from:

1. Plan: what is worth writing, and why it matters to the reader.
2. Research: three searches at once: what ranks, what the sources say, what practitioners complain about.
3. Outline: the argument and the order, fixed before a word is written.
4. Write: a draft, then a second model attacking it for claims nothing supports.
5. Edit: four passes: voice, every fact against a live source, every link fetched, then a fold-in.
6. Assemble: the code that turns the draft into a page, opened as a pull request.

It also runs skills taught since: competitor comparison, coverage of an event while it is still on, a brief for an external writer. It picks up a new skill every few days and keeps its own schedule (tell it to do something every Monday at 9:00 and it writes that into its calendar). Its inputs via Marketing OS: demand signals from GA4 and Google Search Console, buzz on Twitter and Reddit, results of experiments on increasing citation on LLM surfaces, and a log of what customers care about from customer conversations. Helix holds knowledge no individual holds, and learns continuously and faster than any of the team.

Helix hands each part to the model best suited:

- Planning: Claude reasons across everything known to build the topic tree; Gemini checks what already ranks.
- Research: Gemini on search results, Perplexity through the sources, Grok on what practitioners complain about.
- Writing: Claude drafts; Grok attacks it for unsupported claims; Claude revises.
- Editing: voice first, then every factual claim against live sources, then a plain HTTP request to confirm each link resolves, then a pass folding in what the first three found.

In essence, Helix figures out what readers care about, gathers credible information, generates a unique PoV from Atlan's reality and experiences, and writes the code that makes the piece a live URL.

## Evolution isn't trying to get somewhere, it is trying to make the best version of ourselves

Are the team's jobs done? In a way, yes; nobody could keep doing what they did 3 years or even 6 months ago. "The risks that need our attention are far more existential, our opportunities that peak from the horizon are generational. Helix is doing our job so we can move to what's next."

Helix makes most past constraints on content quality and quantity obsolete:

- An opportunity outside the content calendar takes hours, not days. Example: a request to publish a recap of the NVIDIA GTC 2026 keynote and rank for it, asked at 8:37 AM, live on atlan.com at 10:54 AM.
- Testing demand across competing theories of how the market defines a problem no longer takes quarters; it can run in one sprint. Example: one cohort of eight pages on context engineering sitting at six different stages at once, nothing queued behind the piece in front.
- A newsroom-like experience for live events, by connecting Helix to the right sources and people.

The biggest value is not spending a third of the originally approved content budget, going academic on topics with the best subject matter experts, or cutting time to market. It is the option to work on the most impossible problems and exciting opportunities that otherwise get sidelined by keeping the lights on.

Closing: evolving the content engine for a growing business and a changing world also meant the team had to evolve into different animals that could devise and maintain each version, and imagine what's next.

## Author

Labani Biswas, Content, Atlan. Labani runs content at Atlan and the engine this piece is about. She has rebuilt it four times since 2022, and now spends most of her time on the problems Helix cannot pick up on its own.

More essays: [Becoming Frontier](https://atlan.com/frontier/).