Marketing Autopilot for SaaS — manifesto
A ninety-day strategy, founder-voiced content, distribution across nearly eighty platforms. None of your prompts ever leave for OpenAI, Anthropic, Mistral or Gemini.
Public beta · Free plan, no credit card · 100 % sovereign inference
§ 00 — Demonstration
From brief to first submission : the method, the three distribution modes, sovereignty — shown, not told.
§ 01 — Method
01 — First
A few minutes to ingest your website, extract the promise, the ICP, the social proof you already have. Writing, analysis and review run on Sponge Theory’s own models (SPT Models). Only the GEO measurement asks public AI assistants (ChatGPT, Gemini, Perplexity) about your market — with the product’s name, its category, its competitors and your audience’s problems as your brief describes them, never your private data.
02 — Second
Gabriel Weinberg’s nineteen acquisition channels, plus GEO — your visibility inside AI assistants —, are scored against your product. At most three make the bullseye and are tested in parallel: three to four weeks each, eight for the channels whose effect compounds.
03 — Third
Your tone is inferred from your own site during the brief — register, vocabulary, the words you never use. You review and correct it in the app ; it is then applied to every piece of content. Never invented.
04 — Fourth
Official APIs where they exist. Soon, your own browser, from your IP, through an extension. For Reddit, Hacker News and Indie Hackers, you stay at the helm.
§ 02 — Tier 1 selection
A — Z
A — U
B — X
Alphabetical excerpt : thirty-one of the catalogue’s ninety-four listings. The full list, with each platform’s execution mode, lives in the app.
§ 03 — Architecture
Sovereign strategy
Gabriel Weinberg’s Bullseye, applied to your product in about ten minutes. Three acquisition channels run in parallel, scored against Dave McClure’s AARRR metrics. Whichever wins, we double down on it.
The thirteen-week plan is generated by SPT Models, on the GPU cluster Sponge Theory runs in France. None of your prompts, none of your metrics, none of your output ever transit OpenAI, Anthropic, Mistral or Gemini. The strategy is versioned, re-scored every two weeks against your measured results — you approve each new version — and stays readable : every channel carries its reasoning, not a black box.
Editorial voice
Short and long directory copy, X threads, LinkedIn posts, longform articles, newsletter pitches, video scripts. Six formats produced in the founder’s voice — inferred from their site, then recalibrated by hand the moment it rings false.
Every generation passes through an LLM-as-judge compliance gate that flags tone drift, factual hallucinations, and platform ToS violations. By default nothing leaves without the founder’s approval. Autopilot lifts that hand : on connected channels it generates, schedules and publishes on its own — a Pro option, armed workspace by workspace behind a dedicated live-publish switch, revocable in one click. Except on LinkedIn and Hashnode, whose terms forbid automated posting : there, every post waits for your approval, then goes out at its scheduled time.
Respectful distribution
Mode A — official APIs where they exist and ToS allow it : LinkedIn, X, Dev.to, Hashnode, Bluesky, Mastodon, GitHub Topics, plus YouTube and TikTok for video.
Mode B — coming soon : submission from your own browser, via a Manifest V3 extension you will install. Your cookies, your IP, your fingerprint. Undetectable because it really is you.
Mode C — semi-automatic on human-moderated communities. Rayonne drafts ; you press publish from your own account. Reddit, Hacker News, Indie Hackers : a hard line.
§ 04 — Video studio
The brief already knows the product, the promise and the voice. The studio turns them into a short video — plan, site capture, voiceover, music, captions, render — then pushes it where it counts.
Vertical for TikTok, landscape for YouTube, square for the LinkedIn feed. Each format is rendered for its venue : no letterboxing, no repost that smells recycled — a platform expecting 16:9 refuses a 9:16 render rather than squashing it.
The voiceover is synthesised on SPT Models, the music bed generated, the captions aligned to the audio. No studio to book, no sound library to license, not one byte sent to an external AI provider.
YouTube, TikTok, LinkedIn, X : the rendered video is dispatched like any other piece of content, with the same evidence capture, the same retries on failure, the same attribution link.
From the Pro plan up : twenty videos a month, sixty on Business, one hundred and fifty on Agency. The GPU is shared with audigeo and NeoRAG — so the video quota is real, and unlimited on no plan.
§ 05 — Numbers at October 2, 2026
§ 06 — Tariffs
Free
To validate the method on a single product.
Starter
For the solo founder shipping their first SaaS.
Pro
For the founder iterating on two or three products.
Business
For the small agency and the studio that resells.
VAT excluded. Free stays free, indefinitely. An Agency plan (unlimited workspaces, 150 videos / month) is available on request : contact@sponge-theory.ai.
« Rayonne does what I used to do on Sunday nights, but across a hundred channels — and without forgetting me on Monday morning. »
Christophe Verdier — founder, Sponge Theory — first dogfooder
§ 07 — Questions
Not for anything Rayonne writes, analyses or reviews : inference runs exclusively on SPT Models, the GPU cluster Sponge Theory runs in France, with no external fallback. The one exception, the GEO measurement (audigeo.ai, a Sponge Theory service), asks ChatGPT, Gemini and Perplexity, run from the United States, your market’s questions to measure whether they mention you : those questions carry the product’s name, its category, its competitors and your audience’s problems as your brief describes them — plus the ones you add —, never your private data.
Never. These communities explicitly forbid automated promotion. Rayonne drafts in your voice, surfaces ToS flags, and you publish from your own account. It is semi-automatic, not a bot.
Yes, and that is the point. Until you touch it, every piece stops for review : that is the default. Turning autopilot on for a workspace — a Pro option, then a second switch for live publishing — lets it generate, schedule and publish on its own, on the channels you have connected. On each platform, its first five texts go through you; after that it approves on its own, unless the second model that checks the facts has a doubt — the text then waits for you. Two exceptions written in their terms : LinkedIn and Hashnode forbid automated posting — the autopilot drafts and schedules there, but every post waits for your approval in the app, then goes out at its scheduled time. Mode C platforms still wait for you : no setting automates them.
A few minutes for the brief, about ten minutes for the ninety-day strategy on SPT Models, then a few minutes per piece — judge review included.
The extension (Mode B) is coming soon : until then, no submission depends on a platform’s DOM. It ships an audit tool that tests the declared selectors against the real page, filling and submitting nothing, and every release passes the full test suite before it ships.
Yes, at any time. The voice editor on the brief page opens three fields : tone, preferred vocabulary, forbidden words. What you write there applies to the next generation — no need to re-run the brief.
The crawler only reads public pages on your own domain. No external database is queried. Generated content stays private to your workspace ; it stays, cancellation included, as long as your workspace exists, and is deleted when you delete it or on request — except the GEO audits, kept by Audigeo until a deletion request ; the tracked questions and their readings are deleted with the workspace. The privacy policy has the detail.
§ 08 — What now
Create your free account, paste your product URL : a brief in a few minutes, a ninety-day strategy within a quarter of an hour, first submissions within the half-hour. A question before you start ? Write to us, we answer fast.