This is a box that sat at the bottom of my post about AI writing tools, copied exactly as it appears in the export of the site I took on 19 September 2026:
OUR TOP PICK
Our #1 pick for AI writing: Jasper
Affiliate link · we may earn a commission at no extra cost to you.
The button under it went to jasper.ai. Just the home page: no affiliate ID, no referral code, nothing that could have paid anyone. I have never used Jasper. And the article above the box had named a different tool, Claude, as the best overall.
That box is a fair summary of how I did affiliate marketing on this blog: most of the textbook mistakes, made in public, without a single programme having approved me. I exported all hundred posts through the WordPress REST API and counted. What follows is each mistake with its evidence, why it was wrong, and what replaced it. The old version of this page was a generic list of ten mistakes that promised the reward for avoiding them was scaling “to five and six figures”. That line is gone, along with every other number in it I couldn’t source.
Mistake 1: a disclosure label on links that paid nothing
There were 29 of those boxes, on 29 different posts. Every one carried the line “Affiliate link · we may earn a commission at no extra cost to you.” None contained an affiliate ID. They pointed at jasper.ai, midjourney.com, surferseo.com, claude.ai, chatgpt.com and the rest: front pages, every one.
It is tempting to call that harmless, since nobody paid anybody. It isn’t. The FTC’s guidance says that if you link to a retailer as a convenience and aren’t paid, “there isn’t any connection you need to disclose” (FTC endorsement guides FAQ, read 21 September 2026). So the label was not a cautious disclosure. It was a false statement about the link, telling readers a recommendation was paid when it wasn’t, and Google’s publisher policies specifically rule out content that “misrepresents, misstates, or conceals information” about the publisher or the content. It also told every reviewer that the page existed for the sake of the button.
What replaced it: all 29 boxes come out. Links to tools now sit inside the text, with rel="nofollow", no parameters, and no disclosure, because there is nothing to disclose. The day a programme approves me, that particular link becomes rel="sponsored nofollow" with a plain label next to it. The rules behind that are in my guide to how affiliate programmes and disclosure actually work.
Mistake 2: winners picked from tools I had never opened
The same 29 boxes each crowned a winner. Jasper was the “#1 pick” for AI writing, and also the top pick for small businesses, the top pick for content creators, and “The one tool worth paying for”. Surfer SEO was the “#1 pick for bloggers”. GitHub Copilot was the top pick among AI coding assistants.
I have never used Jasper or Surfer. I don’t use Copilot either; I build with Claude Code most working days. So the most prominent recommendation on those posts contradicted either the article it sat under or my own practice. The boxes came from a template with a slot for a tool’s name, and something always went in the slot.
“Top pick” is a claim of comparison. It says you have used the alternatives and chose one. Without that, it is an advert with a trophy drawn on it. Google’s guidance on writing high-quality reviews asks for evidence of first-hand experience, and Amazon’s own programme policies ask for “significant commentary, analysis, or transformation” of anything you borrow. A crown on a product you never opened meets neither standard.
What replaced it: no winners at all unless I have used the things being compared. Tools I haven’t used appear, when they appear, marked “not tested” and compared only from their own documentation, with the date I read it. The coding post shows the change best: it now covers Claude Code from daily use, and Copilot and Cursor on paper only, and says which is which.
Mistake 3: one product, two prices
Every one of the hundred posts ended with a box selling my own twelve-page PDF on Gumroad, at $12. Gumroad charged $5.00. That was the price in the listing’s own product data when I read it on 19 September 2026, and again on 21 September.
Advertising a higher price than you charge sounds like the harmless direction. It isn’t, for two reasons. A reader who notices stops trusting every other number on the page, and this was a blog full of numbers. And a site where the same product has two prices looks unattended, which is the last impression you want to give someone deciding whether it deserves ads. The box also promised a four-figure monthly income, which is a separate problem, covered below.
What replaced it: the box comes out of every post, and nothing on this site links to the PDF until the listing, the file and this blog say the same thing. If I mention the product again, the price will be the one on the listing that day, with the date next to it.
Mistake 4: Amazon boxes that recommended nothing
There were 71 “Recommended on Amazon” boxes. Forty-five linked to a search results page for “artificial intelligence” with a tracking tag attached. The other 26 linked to amazon.com/s, a search page with no search term and no tag at all. On top of those, 99 posts carried a box for my own Kindle book. Not one of the Amazon boxes named a product I had used and was recommending for a reason.
Under each one sat the line “As an Amazon Associate, we earn from qualifying purchases — at no extra cost to you.” The “we” was one person. Amazon’s own required wording is “As an Amazon Associate I earn from qualifying purchases”, and its policies say a participating site “must contain original content”. A box pointing at a generic search is neither a recommendation nor content. It is a hole in the page with a logo on it.
What replaced it: every Amazon box comes out. I am not in the Associates programme. If I ever link to something on Amazon, it will be a specific product I have used, inside a sentence that says why, with Amazon’s disclosure next to it.
Mistake 5: income figures as the sales pitch
The old version of this very post told readers what monthly affiliate income to expect after six to twelve months, put a dollar figure on what most new affiliates earn in their first six months, and claimed an email list “converts 5-10x better than any other channel”. None of it had a source. Across the site, one four-figure monthly income figure appeared 108 times, at least once in every post, because it was printed in the box that sold the PDF.
Nothing on the site backed those figures, from me or from anyone else. On a site about making money, an income claim is not decoration: Google’s publisher policies single out “Get Rich Quick” schemes as deceptive. The rest of that story, with the counts, is in the export behind this blog’s four AdSense rejections, so I’ll keep this one short.
What replaced it: one rule. An income figure goes on this site only if it is mine, it carries a date, and I could show you the statement behind it. Today that means none.
Mistake 6: advice to add screenshots, on a site with none
This one is almost funny. The blog had a short post, 122 words of text, on how to write affiliate reviews with AI. Its key advice was to “Include your own testing and screenshots”. The site it was published on had no testing and no screenshots of anything. The advice was right. The site just didn’t follow it.
That post now redirects here, and this is how I write about a product with an AI assistant in the loop:
- Use first, write second. If I haven’t used it, the post says “not tested” and sticks to what the company publishes, with the link and the date.
- Notes before prompts. The model gets my notes: what I did, what broke, what it cost, on which date. It helps with structure and wording. It does not supply experiences.
- Every number gets a source or goes. Prices come from the pricing page, with the URL and the date I read it, and the entry price and the renewal price stay apart.
- A hunt for sentences that sound lived-in but aren’t. Models are very good at the plausible detail: support that “answered in minutes”, a dashboard that “feels intuitive”. If I can’t point to when it happened to me, it goes.
- A short note on how it was tested, on posts where there was a test: the tools, the plan, the tasks and the date.
The longer version of that process, including what the model is genuinely good at, is in how I write articles with AI without publishing filler.
How the clean-up was done
Not by hand, post by post. The REST export put every post’s full HTML into one file, and a short script counted each kind of box and listed where it appeared. The 29 “Affiliate link” boxes went into a spreadsheet with the post, the tool, the headline and the destination URL, and every row got the same decision: remove. Each surviving post then got a cleaned version with the boxes, the fake tables of contents and the fixed “Related Articles” strips taken out. Posts that don’t survive go to draft, with a redirect to the article that replaces them.
Finding the boxes was quick. Rewriting the posts that survive is the slow part, and that is the honest ratio: these mistakes were cheap to make and are expensive to undo. How affiliate links fit next to ads and your own products on a small site is covered in how a small blog actually makes money.
The textbook mistakes I never got far enough to make
The old list had ten entries. Several of the ones I haven’t covered here only happen to affiliates with a working programme: cloaking links or chaining redirects to hide them, relying on a single traffic source for your commissions, and never checking which page produced a sale. You need a sale before you can fail to track it. I can’t claim credit for avoiding those; I just never reached them.
One line from the old list I would keep exactly as it was: “Only promote products you have actually bought, tested and used for at least two weeks.” I published that sentence on this blog, and the same blog put a crown on a tool I had never logged into. The gap between those two things is the whole lesson of this post.