A founder does not need thirteen automations. A founder needs a split: which loops may run without them, and which loops must stop because they spend money, speak as the house, recommend a product, or publish onto a ranking surface. The rest is tool shopping.
This piece runs three models on that split. Requisite variety counts whether you have enough distinct responses for the states the system can enter. Reputation decay treats trust as a stock that leaks when cheap output is the only flow. Ruin sizing puts a cap on the one loop that can end the account. Together they give you a card you can tape to the one automation you already run.
The machine has no taste. Taste is the human stop. Volume is not a personality.
Why these three
A bought playbook will sell you the opposite story. Assistance is for people still chained to the desk. Automation is freedom. Blog posts publish while you sleep. Ads scale the winners. A chatbot drops your affiliate links at 2 a.m. The picture is always the same: you on a hike, the dashboard green.
Two things broke that picture, and neither is a vibe.
First, the controller is under-equipped. Thirteen named loops (blog, email, social, ads, bonuses, reporting, offer rotation, chat, repurposing, research, review video, outreach, compliance) and one recommended response: set it and forget it. Ashby’s law does not care how expensive the software is. If the states the environment can enter outnumber the responses you can actually make, you are not in control. You are subscribed.1
Second, the platforms rewrote the rule. Google’s spam policies now name “scaled content abuse” in plain language, including pages generated with generative AI that add no value for users. The March 2024 Search update said the abuse can be actioned whether the volume came from a machine, a human, or both.2, 3 YouTube, from 2024 and more visibly from May 2026, labels realistic synthetic media and will apply the label itself when it detects photorealistic AI the creator did not declare.4, 5 The FTC’s endorsement guides still require an honest opinion and a material-connection disclosure close to the recommendation. “Paid link” next to the URL can do the commercial job. It does not turn a review of a product you have not used into an honest opinion.6
That is a matrix break. The 2019 content-farm base rate is not your reference class anymore.
The framework
1. The path: one response is not control
Requisite variety is a counting argument. List the distinguishable states of the thing you think you are running. List the distinguishable responses you can actually make. If responses are fewer than states, you cannot regulate the thing, however hard you work, and adding another vendor with the same “let it run” behaviour does not change either count.1
On an affiliate or content operation the states are not “on” and “off.” Each loop can be healthy, degrading, overspending, banned, or quietly lying. Ads can be profitable at 9 a.m. and ruinous at 4 p.m. A blog can be indexed on Tuesday and hit with a spam classification after the binge. A chatbot can answer a FAQ correctly and invent a bonus by Thursday.
The playbook’s response set is a single move: walk away. That is attenuation by abandonment. The two legal repairs are still only two. Attenuate what comes in (run one loop, not thirteen). Amplify what you can do (add response classes: approve, cap, disclose, kill). A human approval queue is an amplifier. A seventh AI writer is not.
African rails do not change the count. WhatsApp status, a Jumia cookie, a mobile-money payout: still four classes of act that need a stop. Spend, speech, recommend, rank. The costume of WarriorPlus is irrelevant. The stops are not.
Requisite variety, the equilibrium lens
– Assumes: a controller cannot regulate a system whose distinguishable states exceed its distinguishable responses.
– Fits because: thirteen loops, one recommended response.
– Breaks when: the environment is genuinely simple (one product, no ads, no ranking surface).
– Evidence: A (Ashby). Transported from cybernetics; the transport is the bound.
– Counteracts: buying another tool and calling it control.
– May reinforce: over-attenuating until the business has no surface at all.
2. The path: trust is a stock, posts are a flow
You cannot “increase trust” as an action. Trust is a stock. It changes only through deposits and leaks. A first-hand note on a product you used, a disclosure the reader actually sees, a bonus that arrives when you said it would: deposits. A faceless review of something you have not held, a footer disclaimer nobody reads, a 404 on the affiliate URL: leaks.
Unattended output is a leak with a schedule. The audience does not see your Zapier scenario. They see the post. The FTC’s own test is what a significant minority of consumers would not expect.6 Platforms see volume and pattern, not your intent. YouTube will now write the synthetic label for you if you do not.5 Amazon requires you to identify yourself as an Associate on the site, which is a minimum deposit, not a strategy.7
Reputation also has a next period. The list you burn this month is the list you cannot mail in November. Start small is not a motivational poster here. It is the only way a leaking stock gets a chance to refill: one loop, one week, visible deposits.
Reputation decay, the cycle lens
– Assumes: trust is a stock that leaks without replenishment, and cheap output accelerates the leak.
– Fits because: the audience and the platform only see the output.
– Breaks when: the audience was never going to form a trust stock (pure coupon clippers).
– Evidence: B+ as a diagnostic. Not a forecast of the week a given list dies.
– Counteracts: treating a burst of posts as a deposit.
– May reinforce: publishing nothing, which is a different leak.
3. The path: the ads account can end the game
Ruin is the only gate that vetoes a positive expected value. An ads account with platform auto-scale and no daily ceiling is an absorbing barrier. “The platform will pause the losers” is not a cap. It is a hope about someone else’s optimiser, pointed at someone else’s objective.
The same shape applies to a network ban and a Search manual action. They are slower than a Meta spike and just as absorbing. You do not size those bets by average case. You size them by whether the bad case is survivable.
The cheap version of this model is ugly and sufficient. Write the number that, if it left the account today, would change the next payroll or the next school fee. That number is the daily cap. Put it in the ads UI before you generate the twentieth creative. If you cannot name the number, you are not running ads. You are running a lottery with a company card.
A disposable test domain with prepaid spend is the honest exception. Most founders do not have one. They have the site their name is on.
Ruin sizing, the random lens
– Assumes: an absorbing barrier exists. Expected value does not licence a bet that can hit it.
– Fits because: auto-scale and bans are absorbing.
– Breaks when: spend is prepaid and capped, and the property is disposable by design.
– Evidence: A as a constraint. Not a prediction that ruin will occur.
– Counteracts: “the platform will pause the losers.”
– May reinforce: refusing any paid test, including one that cannot hit the barrier.
The four stops, cheapest first
Do not buy a stack. Classify what you already run.
- Spend. Daily cap in the ads UI. Kill switch you can hit from a phone. No auto-scale above the cap.
- Speech. Anything that talks as the house (email promo, chatbot, outbound) sits in a hold for a human pass. FAQ deflection can run. Invented bonuses cannot.
- Recommend. No review, no “best of,” no pinned affiliate link on a product you have not used. If you have not used it, say so, or do not publish.
- Rank. Nothing auto-publishes to a surface Google classifies. Drafts may be AI. The publish button is not.
Link-health crawls, bonus delivery, and inbox reporting can run. They do not speak, spend, recommend, or rank. That is the keep list from the playbook, stripped of the walk-away religion.
WhatsApp status is still speech. A Jumia or Amazon cookie is still a recommendation. A mobile-money payout does not make the four stops optional. If the first loop you pick sits on a rail your buyers already use, you will feel the stops sooner, which is the point. A costume stack (WarriorPlus, a US bonus page, a hike-and-dashboard story) delays the feeling and does not change the count.
Disclosure is not a fifth stop. It is the sentence that makes the third stop visible. Place it next to the URL, in the same post, before the click. “Paid link” is enough for the commercial relationship. It is not a substitute for honest opinion.6
What to do in the next 28 days
Do now. Write the four stops on a card. Put the card on the one loop that already runs. If that loop is ads, set the daily cap today. Effect on spend arrives the same day. Effect on the trust stock arrives with the next clean posts, call it T+14.
Hedge. Do not turn on unattended blogging or faceless review video. The premium is the posts you did not publish. That cover has to be live before the next binge, not after Search Console gets interesting.
Defer, with a trigger. New vendors (another writer, an avatar studio, an Instantly seat) wait. The trigger is a 30-day sample, human-reviewed, that beats the current baseline on revenue after refunds, not on post count. Size the spend when you set the trigger, not when a sales page is open.
If the do-now effect arrives after the deferred purchase, the portfolio is upside down. Queue first. Shop second.
What usually happens to people who skip the split
Name the shape before the story. Content operations that industrialise thin pages follow S-shaped growth with overshoot, or overshoot and collapse. The collapse is a classifier, a reviewer, or a ban, and it arrives on a delay the founder sized as if it were zero.
Has a rule changed? Yes. Google said so in March 2024. YouTube said so in 2024 and again in May 2026. That is the break test. Do not take the 2019 blogger base rate into a plan.
The reference class is sites that industrialised thin affiliate pages, not “people who use ChatGPT.” Google’s own sentence is the match: many pages, generative tools, little or no value for users, purpose to manipulate rankings.2 Present-state modifier: the binge is cheaper than it has ever been, so the class is larger and the classifier is hungrier. Subtract the banned state before you book the hours you think you saved.
What this ensemble cannot see
Merchant-side quality. A first-hand, well-disclosed review of a product that then refunds at a brutal rate still poisons the list. None of the three models watch the merchant. You will have to.
There is also a lag the cards do not name. You know the approval queue exists. The reader only sees the post. For a while those two pictures disagree, and a user-visible join (a complaint, a refund, a “this person never used it” comment) is the moment the lag becomes an inconsistency. Until then you can look identical to the founder who automated the witness.
The action that survives that ignorance is small. This week, write the four stops. Put them on the loop that already runs. Do not buy the stack.
That is the decision.
Sources and notes
- W. Ross Ashby, An Introduction to Cybernetics (Chapman & Hall, 1956), the law of requisite variety. Primary statement of the counting argument. Transported here from general control to affiliate operations; the transport is stated in the card.
- Google Search Central, “Spam policies for Google web search,” section on scaled content abuse, including the generative-AI example. https://developers.google.com/search/docs/essentials/spam-policies. GET 200 and claim present, 17 August 2026.
- Google, “New ways we’re tackling spammy, low-quality content on Search,” 5 March 2024. Strengthens the policy so scale-without-value can be actioned whether produced by automation, humans, or both. https://blog.google/products-and-platforms/products/search/google-search-update-march-2024/. GET 200, 17 August 2026.
- YouTube, “How we’re helping creators disclose altered or synthetic media,” 18 March 2024. Realistic synthetic content must be disclosed; scripts and ideas are exempt. https://blog.youtube/news-and-events/disclosing-ai-generated-content/. GET 200, 17 August 2026.
- YouTube, “Improving AI labels for viewers and creators,” 27 May 2026. More visible labels; auto-apply when systems detect significant photorealistic AI the creator did not declare. https://blog.youtube/news-and-events/improving-ai-labels-viewers-creators/. GET 200, 17 August 2026.
- U.S. Federal Trade Commission, “FTC’s Endorsement Guides: What People Are Asking.” Honest opinion; material connection; “paid link” next to the URL; closer to the recommendation is better. https://www.ftc.gov/business-guidance/resources/ftcs-endorsement-guides-what-people-are-asking. GET 200, 17 August 2026. This is guidance, not legal advice.
- Amazon Associates, “Why do I have to identify myself as an Associate?” Legally compliant disclosure plus on-site identification as an Associate. https://affiliate-program.amazon.com/help/node/topic/GHQNZAU6669EZS98. GET 200, 17 August 2026.
- Google Search Central, “Creating helpful, reliable, people-first content.” Method is not the crime; unhelpful pages are. https://developers.google.com/search/docs/fundamentals/creating-helpful-content. GET 200, 17 August 2026.