How it works

AutoFunnel isn't a single model, and it isn't a content generator with a scheduler bolted on. It's an orchestrator: a system that sits on top of several AI providers and a purpose-built data model, and coordinates them through a repeating cycle — analyze, build a strategy, execute, measure, learn, adjust.

Two layers make that possible. A static structure that holds what's true about a brand and its market. And a loop, decoupled from that structure, that actually runs the work.

Everything the system knows about a brand lives in one tree: the brand at the root, its products and markets as branches, each market broken into segments, each segment carrying its own competitor profiles, best practices, and marketing plans. A plan holds channels; a channel holds the actual assets — a post, a page, an ad, a video.

Brand
  ├── Product
  └── Market
      └── MarketSegment
          ├── CompetitorProfile
          ├── BestPractice
          └── MarketingPlan
              └── Channel
                  └── Asset  // post, page, ad, video…

Nothing here is a static template. Every branch gets re-read and re-evaluated as the loop runs, which is what keeps the structure current instead of becoming an onboarding snapshot that quietly goes stale.

The tree describes what's true. A separate engine decides what to do about it, structured as a chain: a hypothesis about what might work, one or more tickets that execute it, a learning that records what actually happened, which feeds the next hypothesis.

Hypothesis → Ticket(s) → Learning(s) → next Hypothesis

This chain is what makes the system genuinely iterative instead of merely automated. Automation without this loop just repeats the same action on a schedule. Closing the loop means every cycle is informed by the one before it.

Before a single asset gets generated, the system builds the segment it's about to speak to: who the competitors are, what's currently working for them, what a search-and-review scan of the market actually shows. That research isn't a one-time onboarding step — it re-runs as competitors change tactics and as the segment's own results accumulate.

Research produces observations. Strategy turns an observation into something specific enough to test: this angle, on this channel, for this segment, expected to move this metric. That specificity is deliberate — a hypothesis that isn't falsifiable can't produce a learning, and a strategy that can't produce learnings can't compound.

Once a hypothesis has a plan behind it, execution runs as a fixed pipeline, one step at a time, each step's input and output persisted so the whole run can be inspected or retried without starting over.

generate_brief → generate_copy → [generate_image] → compile_asset → validate

Each ticket also carries four independent status tracks — whether it's approved, where execution stands, whether the result passed review, and whether it's been published — because those four things don't always move together. A ticket can be approved and still failing execution; validated and still waiting to publish.

A compiled asset gets shaped for where it's going — a square crop and a caption for one platform, a vertical cut for another, a full page with its own domain for a landing page — rather than one piece of content stretched across every format. Publishing itself can run on approval or on autopilot, per channel, once a channel has earned the trust.

Results come back in, get compared against what the hypothesis predicted, and get written up as a learning — not a raw metric, but a conclusion: this angle underperformed because of X, this format outperformed because of Y. That learning is what the next hypothesis reads before it's written, which is the entire mechanism behind the system getting better instead of just staying busy.

No single step is locked to one AI provider. Research, strategy reasoning, copywriting, and image generation are treated as distinct jobs, each routed to whichever provider is actually earning that job right now — see Tools & platforms for how that gets decided.

Every generation runs on your own provider keys, at your provider's real rate — see Pricing for why. Your brand's data — research, strategy, history, results — belongs to the brand record you control, not to a shared model trained across customers.