AI Content Detection and Affiliate Sites: What Detectors Actually Measure and Whether It Matters
Someone runs your guide through a detector and it comes back "92% AI". The question that follows is whether that number means anything for a site that lives on Google traffic, and the answer needs three parts: what the number measures, how often it's wrong, and whether Google looks at it. Short version: predictability, often, and no.
This post covers what AI content detectors actually score, why their false positives are a structural problem rather than a bug, what Google has said and done, the situations where a detector score does matter, and what to do with a flagged guide. It's the detector-specific companion to the Google and AI content post, which covers the ranking policy.
The short answer
Detectors score how predictable the text is, not where it came from. Predictable text includes most clear, edited, formal writing, so they flag humans at a rate that makes any single score unusable. Google doesn't run them; it judges whether the page helps.
A score matters only when a person or a network you work with has made it a rule. The fix for a flagged guide is the same edit that makes it better: a real opinion, a specific example, a sentence only you could write.
What detectors measure
The common tools estimate two things: how likely each next word is given the words before it (often called perplexity, lower meaning more predictable) and how much that varies across the text (burstiness).
Language models tend to produce low-perplexity, even-paced prose; so do technical writers, non-native speakers writing carefully, students writing to a rubric, and anyone who's been edited for clarity.
None of it is a record of who typed the words. There's no watermark in ordinary model output and no signature a detector can read; it's a statistical guess from the surface of the text, and it gets less reliable as models improve and as humans write more plainly.
The false-positive problem
| Text | Typical detector result | Why |
|---|---|---|
| A spec-heavy product section, human-written | Often flagged | Specs are predictable by nature |
| Formal or academic prose | Often flagged | Low variance, standard phrasing |
| Non-native English, written carefully | Flagged at higher rates | Simpler vocabulary, regular structure |
| Generated text with one strong opinion added | Often passes | The opinion is unpredictable |
| Generated text run through a paraphraser | Often passes | Scrambled phrasing raises perplexity, and the text is worse |
The last two rows are the tell. A tool that passes a scrambled paragraph and flags a careful human one isn't measuring authorship. Vendors publish accuracy figures on their own test sets; independent tests and the US Constitution (which several detectors have flagged) say the number on a single document isn't something to act on.
Google's position
Google's guidance on AI content says appropriate use of automation isn't against its policies and that it rewards high-quality content however it's produced.
Its scaled content abuse policy targets pages generated in volume to manipulate rankings, with no value to readers, and says explicitly that this applies regardless of how the pages were made. Google has never said it uses detectors, and the way its systems are described (helpfulness, experience, originality) is about the page, not the author.
What this means in practice: a detector score and a Google outcome are unrelated. Sites have been hit with 100% human content that was thin and duplicative; sites with heavily generated guides that carry real opinions and accurate picks rank. The E-E-A-T post covers what Google is actually looking for on a review page.
When a score does matter
- A client who's made it a rule. If you're producing content for someone who runs a detector before accepting it, their rule applies. Know which tool they use and edit to it.
- An affiliate network or ad network with a policy. A few ask about AI use in applications; answer honestly, and note that most are asking about scaled junk, not assisted writing.
- Marketplaces that ban AI content. Some content marketplaces and some publications do, and run detectors to enforce it. If you sell there, that's their call.
- Not: your own site's rankings, Amazon's Operating Agreement (which says nothing about how text is produced), or readers, who judge the page by whether it helped.
What to do with a flagged guide
- Run it through two more detectors. They'll disagree, which tells you what the number is worth.
- Read it as a reader. If it's generic, the detector is accidentally right about the thing that matters: the guide has nothing of yours in it.
- Add what's missing. A verdict with a reason you hold, one product told off, a sentence of real experience where it's true, a specific example from your own use of the category. The opinions post is the method.
- Don't paraphrase-scramble. Tools that rewrite to beat detectors produce worse text and the pattern Google's scaled-content policy describes.
- Keep the score out of your workflow unless someone you work with requires it. Time spent chasing a detector is time not spent on the edit that improves the guide.
Doing it with AffSERP
Guides written by AffSERP's X Best Review tool are built around the things detectors can't predict and readers want: a verdict with a reason, a caveat per product drawn from its reviews, and your own notes placed where they belong.
Nothing in the tool is designed to beat a detector, because that isn't the goal; the goal is a guide with a point of view, which happens to be what both readers and the statistics respond to.
FAQ
Does Google use AI detectors to rank content?
No. Google has said it rewards helpful content however it's produced and penalises content made at scale to game rankings, whichever way it was made. Its systems judge the page (does it answer the query, show experience, add something) rather than the authorship. A page that fails a detector and helps readers ranks; a page that passes and says nothing doesn't.
How accurate are AI content detectors?
Not accurate enough to act on for a single document. They score how predictable the text is, and predictable text is also what careful, plain, well-edited humans write; published studies and the vendors' own disclaimers put false positives on human text in the range of several percent to well over ten, higher for non-native English writers and for formal prose. OpenAI withdrew its own classifier for low accuracy.
Can AI content detectors be bypassed?
Yes, and the point is that this shows what they measure. Adding a specific opinion, an unusual example, a sentence in your own voice or a real experience changes the score, because those are the things generic text lacks. That's also the edit that makes the guide better. Paraphrasing tools that scramble the text to beat a detector make the guide worse and are the pattern Google's scaled-content policy describes.
What should I do if a client or network flags my content as AI?
Ask which tool and what score, run the same text through two others (they'll disagree), and offer the edit above rather than an argument: add the opinion, the experience sentence, the specific example. If the relationship depends on a detector score, that's the relationship's rule to follow; it isn't Google's.
Keep reading
