There is a thirty minute job most founders reading this have never done, and it will cost them traffic. Open the last ten products you publicly recommended. Not from memory, actually open them. Then put exactly one of three marks on each: used, not used, unclear. Most people find that the honest count of “used” is somewhere between two and five, and that the rest were written in a confident voice about things they had only read about. That gap was always possible. What changed in the last two years is that it is now free, fast, and indistinguishable from the real thing, because a model will write the review for you and it will read better than yours did.
Key Takeaways
- A recommendation is not a content format. It is a loan of your name, made to someone who is about to spend money they worked for, and the loan is called in whether or not the product was good.
- The new problem is not that AI writes badly. It is that it writes well about objects it has never touched, and nothing in the draft signals the difference to a reader.
- The fix is a register, not a policy: every live recommendation carries used, not used, or unclear. Unclear is not a third category. It is a slower no.
- Disclosure is placement, not paperwork. The money gets named in the same sentence as the link, before the click, and a footer notice does not do that job (1).
- The platforms have converged on the same line from the other direction: unattended, empty content at scale is a named spam category, and undeclared synthetic media now gets labelled whether you declare it or not (2)(3)(4)(5).
- Most of the automation is still fine. Crawls, delivery, reporting and scheduling never speak in your name. Four acts do: spend, speech, recommend, rank.
What is actually new here
For most of the history of small business, praising a product you had not used required effort. You had to write the thing. The effort was a natural brake, and it kept the honest ratio roughly honest, because faking enthusiasm at volume was more work than being enthusiastic about the few things you actually owned.
That brake is gone. A model can now produce four hundred fluent words on a mixer, a plugin, a supplement or a software tool, in your cadence, with your sentence rhythms, referencing the objections you usually raise. It will be better written than the review you would have produced on a Tuesday evening. And the only thing separating it from an honest review is a fact that exists nowhere in the text: whether anyone opened the box.
I want to be precise about the claim, because there is a lazy version of it going around. The claim is not that using AI to write is dishonest. Drafting with a model is drafting. The claim is narrower and harder to dodge: the machine cannot hold the used flag for you. It has no access to whether you have handled the object. It will produce the sentence “I have been using this for three months” with exactly the same confidence whether that is true or false, because the sentence is a pattern and not a memory.
Everything else in this essay follows from that one asymmetry.
Where this came from
A thirteen recipe automation playbook, bought and mined this month, that teaches affiliate marketers to remove themselves from thirteen business loops. It is not a stupid document. Its catalogue of loops is genuinely useful, and its most honest page says to pick one system, run it, and only then add the next.
Its problem is the destination. The pack treats “never having to touch the posts again” as the mature end state, and one recipe explicitly builds faceless review videos for products the operator has not used. That was a workable 2021 position. It is now the exact pattern that ranking surfaces and endorsement regulators spent 2024 to 2026 closing.
I am naming the specimen rather than citing it, because everything in it is instructor assertion with no evidence behind it, no numbers of any kind, and no spreadsheet. It is worth reading the way you read a competitor’s brochure.
What the platforms already decided
The founder question and the compliance question turn out to be the same question wearing different clothes, which is the part most people miss.
Google’s spam policies name scaled content abuse, including generative pages produced at volume that add no value for the person reading them (2). The March 2024 update made the position explicit: the abuse is actionable whether the volume came from a machine, a person, or both (3). And the helpful content guidance is careful about the thing everyone gets wrong: the method is not the crime (4). Nobody is penalised for using a model. They are penalised for publishing emptiness.
The endorsement side arrives at the same place from a different direction. The FTC’s guidance asks for two things: that the opinion is honest, and that a material connection is disclosed close to the recommendation (1). Read those two requirements together and you get a rule the playbook cannot survive. The disclosure is the easy half, and it is the half everyone attempts. The honest opinion is the hard half, and it cannot be automated, because an opinion requires having had an experience.
YouTube closed the last exit in May 2026: labels on realistic synthetic media became more visible, and they are applied automatically where the creator does not declare them (5). A faceless review of an unused product is no longer a clever workaround. It is a labelled object, and the label arrives without your permission.
Amazon, if you touch that network, requires you to identify yourself as an Associate on your own site (6). That is the lowest resolution version of naming the money. It is a floor, not a standard.
The three marks
Here is the whole system. It is not a framework, it is a register, and you can run it in a spreadsheet you already have open.
Used. You have handled the thing yourself. Not a demo, not a trial account you opened to write the post, not a colleague’s account. You used it for its purpose, more than once. This one may stand.
Not used. You have not. That is not a crime, and it does not mean the recommendation has to disappear. It means the sentence has to tell the truth. “I have not used this. I am pointing at it because a founder I trust runs her whole invoicing on it” is an honest and useful post. “This changed my workflow” is not, if it did not.
Unclear. You cannot remember. This is where most of the corpus will land, and the temptation is to leave those standing, because taking them down costs clicks and nobody has complained. Treat unclear as not used. If you cannot remember opening it, the reader has no reason to believe you did, and you are relying on the fact that they cannot check.
The register has one column that does the real work, and it is not the mark. It is the action, computed from the mark, so that the decision is made once rather than argued about ten times.
The grind you may automate, the acts you may not
The purge is not an argument against automation, and I would not want it read that way. Most of the hours are in the boring half, and the boring half is safe.
These may run without you. Link health crawls that catch a 404, because a dead link in front of a buyer is a small broken promise and a weekly scan fixes it. Delivery of a file you already made and already promised. Reporting pulled into one digest you read on a Monday. Trend alerts written into a sheet for you to decide on later. Scheduling of copy a person already approved.
Notice what those five have in common. None of them speaks in your name, and none of them spends your money.
These four keep a person. Spend, because a platform optimiser is not a cap; it is trying to spend your budget well, and it is not trying to keep you solvent. Speech, because an invented offer is still your offer. Recommend, for the whole reason above. Rank, because anything a search engine will index is something you will answer for.
That is four acts, against thirteen loops. The ratio is the point. The playbook is not wrong that most of this can be automated. It is wrong about which part.
The African rails make this less abstract
A founder in Gulu with a WhatsApp status and a customer list is not running a content farm. She is telling people she actually knows what to buy, and in a market where institutional buyer protection is thin, her word is doing work that a returns policy does elsewhere.
WhatsApp is a room, not a billboard. Mobile money is a sending that does not unwind easily. A recommendation made on those rails sits much closer to a word spoken across a market stall than to a 2019 listicle, and the person receiving it has fewer ways to check and less recourse when it is wrong.
Which means importing an American autopilot pack into that room without the used flag is not localisation. It is a costume. The pack assumes a reader who will bounce, forget, and never see you again. Her buyers see her on Sunday.
AVODA has sat with 543+ founders in formation and helped launch 150+ start ups. The pattern I am describing turns up in the room every time a cohort starts talking about generating the reviews, and it turns up as a genuine question rather than a cynical one. Nobody is trying to defraud anybody. They are trying to keep up with a content cadence that a machine made cheap, and the used flag is what keeps the cadence from eating the name.
What this week is for
A principle that cannot survive a delete key is decoration.
- Open every property where your name or your house recommends a product. Site, YouTube, WhatsApp broadcast, the email archive you still mail.
- Mark each recommendation used, not used, or unclear.
- Rewrite or remove everything that is not used. The honest version of a not used recommendation is often a better post than the original.
- On what remains, put the money in the same sentence as the link, before the click. “Paid link” is enough words. Placement is the duty.
- Strip affiliate URLs out of any chatbot that can speak without you.
- Do not switch on an unattended blog or a faceless review recipe while the purge is open.
The triple bottom line applies cleanly here, and it is worth saying out loud because the first line is genuinely uncomfortable. Revenue: the purge may lower this month’s clicks, and if it lowers them a lot, that number is telling you where the revenue was actually coming from. Impact: the buyer is not used. Stewardship: the house does not sell a weight it does not carry. If those three cannot sit at the same table, the affiliate line was never work worth doing.
You do not need another tool to start. You need the list, and the nerve to delete.
Sources
Guidance pages are not legal advice. Where a decision carries legal weight in your market, price the advice in.
(1) U.S. Federal Trade Commission, “FTC’s Endorsement Guides: What People Are Asking.” https://www.ftc.gov/business-guidance/resources/ftcs-endorsement-guides-what-people-are-asking. GET 200, 17 August 2026.
(2) Google Search Central, “Spam policies for Google web search.” https://developers.google.com/search/docs/essentials/spam-policies. Scaled content abuse, generative example. GET 200, 17 August 2026.
(3) Google, “New ways we’re tackling spammy, low-quality content on Search,” 5 March 2024. https://blog.google/products-and-platforms/products/search/google-search-update-march-2024/
(4) Google Search Central, “Creating helpful, reliable, people-first content.” https://developers.google.com/search/docs/fundamentals/creating-helpful-content
(5) YouTube, “Improving AI labels for viewers and creators,” 27 May 2026. https://blog.youtube/news-and-events/improving-ai-labels-viewers-creators/ See also “How we’re helping creators disclose altered or synthetic media,” 18 March 2024. https://blog.youtube/news-and-events/disclosing-ai-generated-content/
(6) Amazon Associates, “Why do I have to identify myself as an Associate?” https://affiliate-program.amazon.com/help/node/topic/GHQNZAU6669EZS98