In short — the answer in one paragraph
Clients' biggest fear is that Google will ban them for AI. It's a myth. Among top-ranking pages, 86.5% contain machine text, and the link between AI share and position is statistically zero (0.011). Google penalizes not the machine but the empty conveyor with no fact-checking. But there's a catch costlier than any ban: in the top 10, AI and humans run neck and neck, yet position 1 — the one that takes the bulk of the clicks — goes to a human about 80% of the time vs roughly 10% for pure AI. And even in paid ads, where AI matches humans on clicks, an ad the audience perceives as "made by AI" converts worse. The machine is great at distributing content. But the top spot and trust go to whoever has real experience behind the text. This is the second of three parts in my breakdown of AI content effectiveness.
A three-part series
This is part 2 of 3
Part 1 was about perception and trust: people can't tell machine text apart blind, but hang a "made by AI" label on it and trust collapses. This part is about distribution: where AI content gets let in and where it doesn't. Search, ads, social feeds. In part 3 I'll pull it all together — why a human + algorithm hybrid is what actually works.
From me — before you read the numbers
There's a story here that taught me a lot. I'd assembled a slick table for this part on conversion by average order value and a neat "human vs AI" split test — all from a single research doc I was handed. But when I sat down to verify every figure against primary sources, as I promised in part 1, half of it fell apart. The loud percentages traced back to one anonymous blog citing itself. The most-quoted "Meta study about +12% to clicks" actually said the opposite. So what's here is only what I found in real primary sources: Ahrefs, Semrush, an academic study with Taboola, the platforms' own engineering blogs. What I couldn't confirm, I cut. Even when it was pretty. Especially when it was pretty.
The big myth: "Google bans you for AI text." It doesn't
The most common client fear. In short — no. Ahrefs ran 600,000 top-ranking pages through its machine-text detector. Purely human pages in the top came to just 13.5%.
Link between AI share and search position: 0.011 — that's statistical zero. In plain terms: Google doesn't care whether a human or a neural net wrote the text. It judges the value of the text itself, not its author.
So what does get penalized. And this matters
The empty conveyor. In March 2026 Google rolled out a Spam Update aimed at "scaled content abuse" — when a site churns out hundreds of articles a day on autopilot, with no editing or fact-checking, just to game the rankings. Those got hit: organic traffic down 50–80% and pages dropped from the index. What got punished wasn't AI. It was the absence of a human at the output.
And to understand why pure AI text underperforms even without any penalty, there's a phrase — Information Gain. Informational novelty. It's a real concept from a Google patent: how much new a page adds beyond what's already in the results. On a standard prompt, a language model outputs an average of what's already indexed. So why would a search engine surface a rehash? Hence the tightening of the E-E-A-T framework, where the first "E" is Experience. Google officially added it to its rater guidelines. That's the main antidote to machine spam: AI can mimic an expert's style, but it didn't test the product with its hands, run a live interview, or break down a real case. It has no experience — only a rehash of someone else's.
The key point: the top 10 is parity, but position 1 is human
This is the heart of the whole part. Semrush took 200,000 URLs across 20,000 queries, filtered down to 42,000 blog articles, and ran them through an AI-text classifier.
That's the whole truth about SEO in one chart. On the lower slots of page one, AI and humans run neck and neck — hence the illusion that "the machine ranks just as well." But position 1, the one that takes the lion's share of clicks, goes to a human about 80% of the time vs roughly 10% for pure AI. Eight times more often. An honest caveat, as is my habit: Semrush did the "human or AI" classification with an automated detector, and those things — as we recall from part 1 — get it wrong both ways. So I treat 80.5% as a strong signal, not a law of physics. But the direction is backed by the logic of Information Gain and by common sense: position 1 goes to whoever has something to say beyond the average.
You can write it all yourself — whoever wants to will still decide it's AI
A personal story. I once ran my own 2016 articles — the ones I wrote myself for the magazine Finansovy Director — through the most popular AI detectors. The services confidently flagged them: "written by AI." In 2016, mind you, there was no ChatGPT in sight.
The takeaway is simple and uncomfortable. Today neither a human nor a detector can tell machine text from live text (covered in detail in part 1). So whether you write it all yourself or pay copywriters, whoever wants to will still consider your text machine-made. And most people simply won't care. And that already costs money. The largest ad study — Columbia, Harvard, the Technical University of Munich and Carnegie Mellon with Taboola, over 500 million impressions — found that on clicks AI ads matched human ones. They didn't beat them, as vendors promise. Even. But ads people perceived as AI-made performed worse. Not worse text — worse attitude.
Recognize it? It's part 1's plot, now in money. What sags isn't quality but trust in whoever the reader sees a void behind. AI delivers the message well. But the moment the audience senses there's no one behind it — they pay less.
Social: the platforms stopped feeding "smooth but empty"
The same logic from search has reached the feeds.
In March 2026 it rebuilt the feed: instead of counting likes, a system built on large language models (internally it's called the Generative Recommender, based on their 360Brew model). It looks not at the like but at behavior: how long a post is actually read, whether it gets saved, whether real discussion starts. And in its own blog the platform says plainly it suppresses "recycled" content and "engagement bait" — the generic post about nothing. A smooth machine post scrolled past in three seconds simply stops getting reach. The platform doesn't ban AI. It bans emptiness.
TikTok
Went for transparency. It was the first major platform to adopt the C2PA standard (Content Credentials) and began auto-labeling AI-made content from metadata — and for realistic faces, voice clones and photorealistic scenes it introduced mandatory disclosure. And here's the direct link to part 1: hidden AI that gets exposed hits harder than honesty up front. Label it yourself and the platform doesn't penalize you, the content lives. Hide it and get caught, and the penalties escalate. Meanwhile text AI helpers (scripts, captions) aren't subject to disclosure: the platform cares about the authenticity of face and voice, not who helped write the text.
What a business should take from this
A simple map emerges. AI is a great distribution machine. But the top spot and trust go to a human.
Don't fear AI — fear emptiness
Google doesn't penalize the machine, it penalizes the unchecked conveyor. 86.5% of the top is written with AI. The question isn't "use it or not" but whether the text has experience the model doesn't.
Experience pays for position 1
In the top 10 the machine gets in on equal terms. But position 1 goes to a human 80% of the time. Want the top of the results — add what AI can't fake: your case, your data, a live interview. That same Information Gain.
In ads it's not "AI vs human" but the label
On clicks they're even. But the perception "this was made by AI" costs you conversion. So: drafts and tests to the machine, and to the human what the audience reads as alive — voice, selection, the final pass.
On social, label it yourself
Honest disclosure isn't penalized by the platform. Caught concealment is. The arithmetic is obvious.
Notice what all four points have in common. Everywhere, the machine does volume, speed, testing, distribution brilliantly. And everywhere the top spot or trust is at stake, a human has to stand beside it — with experience, voice and the final edit. Not "AI instead," but "AI plus." More on that in the third and final part.
Where your money is leaking right now
The costliest mistake is putting the machine on the wrong stretch. Filling a blog with pure AI and waiting for the position 1 it almost never takes. Or keeping a human hand-building a hundred ad variants while a competitor runs tests by algorithm. I help build your content factory on this map: where the machine brings reach and speed, and where a human has to back it up. No taste-based guesswork — on verified data.
Key numbers and sources
| Metric | Value | Source |
|---|---|---|
| Top Google pages containing AI text | 86.5% | Ahrefs (Ong & Guan, 2025), 600k pages |
| Correlation: AI share ↔ position | 0.011 | Ahrefs, same analysis |
| Traffic drop for content farms after Spam Update | 50–80% | Google March 2026 Spam Update + reviews |
| Top 10: human / AI | 58% / 57% | Semrush (2025), 42,000 articles |
| Position 1: human / AI | 80.5% / ~10% | Semrush (2025) + Search Engine Land |
| AI ads vs human on CTR | parity | Columbia/Harvard/TUM/CMU + Taboola |
| Ad study sample | ~500M impressions | same study |
| Ads perceived as AI | convert worse | same study |
| LinkedIn: LLM feed rebuild | March 2026 | LinkedIn Engineering (Generative Recommender) |
| TikTok: auto AI labels via C2PA | 2026 | TikTok Newsroom + Content Authenticity |
Sources
- •Ahrefs (Si Quan Ong, Xibeijia Guan), 2025 — analysis of 600,000 top-ranking pages; 86.5% containing AI, correlation 0.011; via Search Engine Journal and eMarketer
- •Semrush "Does AI content rank in search?" (2025) — 42,000 articles, 200,000 URLs; top 10 57/58%, position 1 80.5% / ~10%; classification by GPTZero (a method with known limitations); corroborated by Search Engine Land ("Human content is 8x more likely to rank #1")
- •Google Search Status / March 2026 Spam Update — crackdown on scaled content abuse; 50–80% traffic-drop estimates per industry reviews
- •Google Search Quality Rater Guidelines + the "Contextual Estimation of Link Information Gain" patent — E-E-A-T (Experience) and Information Gain as real concepts (not direct ranking factors)
- •Columbia / Harvard / Technical University of Munich / Carnegie Mellon + Taboola — ~500M impressions; AI ads ≈ human on CTR, ads perceived as AI perform worse; via bestmediainfo and others
- •LinkedIn Engineering Blog + news.linkedin.com (2026) — feed rebuilt on LLMs (Generative Recommender / 360Brew), dwell time / saves signals, pressure on recycled content
- •TikTok Newsroom + Content Authenticity Initiative (C2PA) — automatic labeling of AI content and mandatory disclosure for realistic faces and voices
