AI Watermarking and Slop Detection: What It Means for SEO

Search Engine Land published Kevin Indig's piece "Slop antibodies: The link between AI slop, watermarking, and commodity content", which lays out how fast platforms are building defenses against low-value AI content. LinkedIn now runs systems that identify slop and cap its distribution to a poster's immediate network, is testing a "Seems like AI slop" report button, and killed its own "Enhance post" feature in favor of a proofreader that fixes grammar without rewriting your voice. All of that happened over a short stretch of months.

That is the headline. The rest of the article is a census of the same immune response everywhere else, plus one upstream development that made a lot of founders nervous: Anthropic announced machine-readable watermarks on Claude's text and file output at the model level, worldwide, persisting across the API, Claude, Claude Code and cloud access through AWS, Google Cloud and Microsoft Foundry.

If you are a solo founder who publishes AI-assisted articles, or you have been considering it, this is the week to read carefully rather than panic. The two things are related but they are not the same thing, and the practical conclusion is narrower than the headlines suggest.

What actually happened

Photo: Homedust · BY

Indig's article documents the pattern across platforms. Substack rolled out site-wide use of Pangram to detect AI use in writing. YouTube demonetizes repetitive, low-effort, emotionally manipulative video, and has terminated and removed batches of slop channels accounting for billions of lifetime views and millions of dollars in annual revenue. Reddit shipped AI-based detection of manipulated and spammy content. TikTok requires AI labels, embeds invisible metadata watermarks, and started testing detection aimed at accounts dedicated to AI spam. Spotify removed more than 75 million spammy tracks. Meta labels AI content across its apps.

On the search side, Google shipped a spam update in June aimed at scaled content abuse. Wikipedia went further: speedy deletion for suspected LLM-generated articles, then an outright ban on using LLMs to write or rewrite article content.

The watermarking part is regulatory, not moral. As Indig notes, the EU AI Act's Article 50 requires providers of generative systems to mark output in a machine-readable format, with penalties up to €15M or 3% of global turnover. Anthropic is complying. Google has been applying SynthID watermarks to Gemini output for a good while now with far less noise about it.

Why watermarking is not the thing to worry about

Indig is explicit that he is not worried about watermarking's impact on marketing, and his reasons are worth repeating because they cut against the doom takes:

A watermark proves a model touched the text. It says nothing about how much, and nothing about quality. Detectors also need a minimum amount of text to work, on the order of a hundred or more tokens in published benchmarks. Editing, paraphrasing, translating or chaining models can weaken or remove the mark. Published paraphrasing attacks have achieved very high success rates against recent watermarking methods at trivial cost per million tokens. And open-weight models apply no mark at all, because watermarking happens in the sampling pipeline at inference.

So the nightmare scenario a lot of founders imagine, where Google reads a hidden watermark in your blog post and deletes you from the index, is not the mechanism in play. Nobody in this story is penalizing "AI touched this." They are penalizing generic.

The actual dividing line: commodity vs. non-commodity

The most useful part of the article is where three organizations independently land on the same definition.

Danny Sullivan shared a slide at Search Console Live Toronto describing non-commodity content as unique (brings a viewpoint, information or content others lack or can't easily replicate), specific (talks about a specific instance, situation or thing, not general rules or generic information) and authentic (demonstrates first-hand knowledge or experience).

LinkedIn's VP Laura Lorenzetti described systems trained to recognize content that adds perspective, context or expertise versus content that feels generic or repetitive, "even if it appears polished on the surface." Reddit CEO Steve Huffman said on the Q2 earnings call that as AI makes information abundant, the challenge is no longer finding content, it's finding context, personal opinion and first-hand accounts.

Read those three again. Not one of them mentions authorship. All three describe the same failure mode: text that any competitor could have produced.

There is even research on the shape of it. A recent paper from University of Maryland and Google DeepMind researchers ran tens of thousands of stories through a classifier deliberately denied every style signal and told to work from structure alone. It kept almost all of the accuracy of models allowed to see word choice and sentence rhythm. Slop has a shape. The paper also found AI states the moral outright far more often than human writers, and that AI stories are much more likely to contain no subplots at all. Over-explaining and single-track tidiness.

That is the pattern your generic "ultimate guide to X" has. Five headings, each one restating the intro, a tidy conclusion that summarizes what you just read. It reads fine. It is also structurally indistinguishable from the other 400 versions of it.

What this means if you are running a one-person content operation

Photo: Generationbass.com · BY

Here is the part that matters for this audience, and it is not comfortable.

Indig's framing is that production cost has collapsed to near zero, which puts pressure on distribution, and distribution has moved under machine control. So the scarce asset is selection. Production efficiency is worthless if the content doesn't reach anyone. He estimates the distribution cost of running an all-AI account at roughly two thirds of LinkedIn reach, around 7% on short-form video based on TikTok field data, and 40% to 95% in SEO if you scaled a content operation.

That last range is the one to sit with. It is a range, not a number, and the width of the range is the point. Two publishers using the same tool can land at opposite ends of it depending entirely on what went into the drafts.

So the honest version of the advice, from a company that sells AI-assisted publishing: the tooling is not the risk. The input is. If you feed a generator nothing but a keyword, you get commodity content, and commodity content now has a floor of roughly zero. This is true of every AI article writer worth comparing, not just the cheap ones. Google is unlikely to index it. LinkedIn may cap it at your immediate network. Almost nothing cites it.

What to feed the machine, and what to add yourself

Practical version, in the order we'd actually do it.

Feed it things a model cannot produce on demand. Indig's test is the right one: would it be expensive for someone else to fake? Your own numbers. Your churn rate, your conversion rate on a specific page, what happened when you changed pricing. Your support inbox. The five questions customers ask before buying. Your failed experiments, which almost nobody publishes and which are the cheapest unique content you own. Screenshots from inside your own product. A model can write about your category. It cannot know that 40% of your trial signups come from one Reddit thread.

Be specific to the point of discomfort. Sullivan's middle criterion is the one most AI content fails. "Improve your onboarding flow" is commodity. "We cut our onboarding from 6 steps to 2 and activation went up, here is the exact screen we deleted and why we were wrong about it" is not. Names, dates, versions, prices, numbers.

Use the approval step for the things only you can add. This is the whole argument for an approval-based workflow rather than bulk autonomous publishing. The draft handles structure, research synthesis, keyword coverage and the tedious parts. You spend fifteen minutes adding the two or three paragraphs that carry your first-hand experience, and you delete the paragraph that states the moral outright. That COLM finding about over-explaining is a free editing checklist.

Attach a real name. Indig lists attributable identity as one of three antidotes: a named author with a track record. Not "Team" or "Admin." Your name, your face, a bio that says what you actually built.

Own a channel where no filter sits between you and the reader. Email, a community, direct relationships. Your blog is one of the few surfaces you control, which makes it more valuable than it was two years ago, not less — which is also why abandoning it after three posts costs more than it used to.

The shortcuts that are now actively dangerous

Photo: zeevveez · BY

Publish-and-forget on autopilot with no human pass. Reworded competitor articles. Machine translation at scale, which cost Reddit dearly: its ?tl= translated pages nearly vanished from ChatGPT's Reddit citations within a couple of months while Reddit rose in aggregate. Volume plays where the plan is to publish 2x more to offset AI Overview click losses. Indig's read on the last 24 months is blunt: production gains from AI do not offset the losses in distribution.

One more thing worth flagging honestly. LinkedIn's system reportedly runs in the mid-90s for accuracy, which Indig points out is orders of magnitude worse than Gmail's spam filters, with no published false positive rate. These filters will catch legitimate work sometimes. That is an argument for making your content obviously non-generic, not for avoiding AI.

The plain conclusion

Nothing in this story says stop using AI to publish. It says the bar for what gets distributed moved, and it moved in a direction that happens to favor founders. You have proprietary data. You have customer conversations. You have a product you built and opinions you earned. Big content teams optimizing for volume have none of that, and they are the ones with 400 interchangeable guides about to get throttled.

Use the machine for the 80% that is structure and research. Spend your fifteen minutes on the 20% nobody else can write. That has always been the deal. It is just that until this year, you could get away with skipping it.