Does Google Penalize AI Content? No - But It Punishes This

Does Google Penalize AI generated content? Short answer: No.

Google has no penalty for AI-written content, and it has said so publicly since February 2023. What it does penalize is content it detects to be produced at scale to manipulate rankings, rather than written with the intention to help anyone. AI now makes it easier to do that badly, but the Google policy cares more about the intention rather than who or what typed the words.

That's the short version, and if it's all you came for, you can stop there. The longer version is useful though, because "there's no penalty" is not the same as "you can publish anything", and the gap between those two statements is key to retaining strong rankings. So read on to dig into this more.

Why does everyone think there's a penalty?

Because a lot of sites full of AI content really did lose their rankings, and it's difficult to watch that and not draw a certain conclusion. As a result, a number of marketers now understandably feel semi-hesitant about using AI to generate content.

The March 2026 core update was the loudest example. Sites that had been publishing mass-produced articles saw large drops, the SEO world lit up, and the story that spread was "Google is cracking down on AI". I understand why - the change was "we started using AI" and what happened next was "our traffic fell", which seems like clear cause and effect.

But look at what those sites had in common beyond the AI: hundreds of almost-identical pages, no first-hand experience, and nothing new that you couldn't get from the first result.

However, plenty of sites using AI didn't drop traffic at all, which is the part that tends to get left out of the story. (If you're trying to work out what actually happened to a specific site, our breakdown of the May 2026 core update walks through how to tell a quality problem from a volatility one.)

What does Google's policy actually say?

This is worth reading closely.

Google's spam policies define scaled content abuse as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users". The examples underneath it include "using generative AI tools or other similar tools to generate many pages without adding value for users".

Notice the structure of that sentence. AI is named - but the violation isn't the AI. It's many pages plus without adding value. If you take either of those away then you're safely outside the policy.

Google's February 2023 guidance on AI-generated content, linked at the top of this piece, says the same thing from the other direction: "however content is produced, those seeking success in Google Search should be looking to produce original, high-quality, people-first content". And more bluntly: "not all use of automation, including AI generation, is spam".

So the all-important question isn’t "was this written by AI", it's "was this made to help someone, or to fill a slot". The second question is the one Google's systems are actually asking.

What does the ranking data actually show?

Here's where it gets slightly more complex because two large studies point in directions that seem contradictory until you put them together and dig a little deeper.

What Ahrefs found. In June 2026 they sampled a million pages from 100,000 SERPs, and ran AI detection on the 150,000 that had enough text to analyse. The average level of detected AI content barely shifts across the whole top 10 - 27.1% at position one, 30.9% at position ten (and remember, this is only the percentage of AI content they’re able to detect, high-quality or well-edited AI content goes way under the radar vs more generic AI content).

If Google were suppressing AI content you'd expect that number to climb steeply as you go down the page, and it doesn't really move. In fact, 5% of the pages positioned 1-3 were detected as fully 100% AI-generated content.

What Semrush found. Their April 2026 analysis of 42,000 blog posts offers a slightly different perspective. Content classified as purely AI-generated took the top spot 9% of the time. Whereas predominantly human-written content held it 80% of the time. The gap narrows further down the results, but at position one it's stark.

Taken at face value, that second study is the one that should worry you. Read it quickly and it says: write it yourself.

But look at the word doing the work. Classified. Both studies rest on detection, and a detector can only tell you what a page reads like — not who or what actually made it.

That isn't a small asterisk. A classifier spots text that reads like AI (usually flat vocabulary, even rhythm, certain grammatical choices), the signs you probably spot yourself, or would certainly start recognising after a week of reviewing drafts. Content that started as an AI draft and was then properly researched, edited and given a real voice doesn't read that way any more; so it doesn't get counted that way either.

Which changes what Semrush actually found. Not "human-written content wins 80% of the time", but "content that reads as human-written wins 80% of the time." How much of that 80% was drafted with AI and simply edited well enough that nobody could tell? Nobody knows. Neither do they.

Do the same to the Ahrefs figure and it stops meaning "the winners are 27% AI" and starts meaning "27% of the text at position one still reads as AI". The true figure is almost certainly higher.

So the two studies aren't contradicting each other at all. They're finding the same thing twice: what wins isn't human content, it's content that doesn't read like a machine wrote it. Which is quietly the most useful thing in either study because it's a description of quality and insight provided, not just of authorship.

What it isn't is a penalty. Nobody is being punished for using AI. It's that pages made entirely by generic AI, with nobody checking, editing or offering an insightful new perspective, rarely turn out to be the best result for the query, and they read like it too.

A small confession since this piece is partly about quality gates. An earlier draft of this article carried a figure attributed to the same Semrush analysis I've just quoted - AI and human content supposedly reaching the top ten at near-identical rates. A tidy stat that made a strong argument.

When I went to verify it against the source before publishing, it wasn't there. Not misquoted, but entirely absent. I’d used some AI-assisted research notes for this piece which was where the stat came from, and it hadn’t been traced back until that point. Lesson: always check your sources!

So how much AI content is acceptable?

This is the question people actually want answered.

The honest answer: Google enforces no threshold. There is no single percentage that trips a filter, because there's no filter of that kind - what gets measured is whether the page is useful, not how it was made.

So the key things to note are:

• 100% unedited generic AI output is often a losing strategy. Not because it's banned, but because it almost never turns out to be the best answer on the page.

• Some AI in the mix is clearly survivable and even advisable - 27.5% of top-three pages sit between 20% and 50% detected levels. If AI generation is the key to helping you create valuable, high-quality content faster, keep it up!

• The number itself isn't the lever. Nobody at Google is counting your percentages of detected AI content. What a moderate blend correlates with is high-quality content that has been created with real consideration (by human or AI) as to what makes this content useful to readers.

What separates AI content that ranks from AI content that tanks?

Almost none of it is the writing itself.

I say that as someone who has always a lot of time on the writing, so it took me a while to accept. But when you compare AI-assisted content that performs against AI content that sinks, the difference is nearly always decided before a single sentence is drafted:

• Does it contain something that isn't already on the SERP? First-hand data, a real example, a unique take or experience, a position. If a page is a competent summary of the current top three, it has no reason to outrank them.

• Is there genuine experience behind it? The first E in E-E-A-T. A model can't have used the product, run the campaign, or made the mistake. (Unsure of your E-

• Is it built on real research, or on what the AI model already knew? This is the one people skip when they're moving fast. A draft generated from a model's training data can only reassemble what it already absorbed, which is often a year out of date and averaged across everyone. A draft built on what's actually ranking today, what those pages miss, and what the question underneath the query really is - that's a different starting point entirely, and you can feel it by the second paragraph.

None of those are settled by whether a human or a model produced the prose. They're settled upstream, in the research and the brief - which is a less satisfying place to look, because it means the fix isn't actually as simple as just "write it yourself".

The bit that makes the AI question go away

The reason the "should we use AI?" debate never resolves is that it's being argued at the wrong stage.

If the quality layer is set at the brief - what the page must cover, what it has to say that nobody else does, what evidence it needs, what quality research it will pull in, how it's structured - then the drafting step becomes an entirely personal choice. Write it yourself because you enjoy it. Have AI draft it and you edit. Or split it up and give some sections to AI and you take care of the first-hand experience. To Google it genuinely doesn’t matter.

The brief creation and the optimisation score are doing the quality work either way, and the page that comes out the other end is held to the same standard regardless of who typed it. Leveraging AI in these stages is where you’ll save time and set your content up for success.

Plus, if you do want to lean on AI for the writing, in Frase you’ll find Brand Hub: you build a voice profile by feeding in samples of your own writing rather than describing your tone in the abstract, and it's applied while content is drafted and while it's edited, not bolted on at the end. Next to it is a terminology layer with three tiers - terms you want reinforced, terms to avoid, and never-terms that make a page unpublishable until they're removed.

None of this is about slipping past a ‘detector’, it’s about giving you the tools to scale the production of genuinely valuable, useful content.

So how would I actually approach a page?

Whether you end up writing it yourself or not, the order is the thing:

1. Start with the research, not the draft. What's ranking now, what those pages are missing, what the person searching actually wants. This is the step that decides whether your page has a reason to exist, and it's the first thing to go when you're in a rush.

2. Decide what only you can add. The example, the number, the client conversation, the thing you learned the hard way. Write that down before anything else - it's the part no model can generate for you, and it's much harder to retrofit into a finished draft than to build around.

3. Then pick who writes it. Draft it yourself, have AI draft it and edit properly, or split it and take the sections that need your experience. By this point it genuinely doesn't matter, which is the entire argument.

4. Be honest about scale. Are you producing lots of similar pages, quickly, with little between them? That's the one thing in Google's policy actually pointed at you - and it's a decision about how you're operating, not about which tool you opened.

What I wouldn't do is count percentages. No threshold exists, and chasing one is effort better spent on the four steps above.

If you want to see what the quality layer looks like on your own pages, there's a 7-day Frase trial with no card required.

FAQs

Does Google punish AI-written content?

No. There's no penalty for AI authorship, and there never has been.

How much AI content is acceptable to Google?

No threshold is enforced, so no percentage is "safe" or "unsafe" in policy terms. For what it's worth, Ahrefs found position-one pages average 27.1% detected AI. That describes what's currently winning. It isn't a number to write towards.

Can Google detect AI-generated content?

Probably some of it. Detectors identify text that reads like AI, which means well-edited AI content in a distinctive voice can pass straight through them - the published studies can only count what they caught. But detection isn't the enforcement mechanism anyway, so it matters less than people assume.

Is AI content bad for SEO?

Unedited, mass-produced AI content is bad for SEO. Semrush found purely AI-generated content takes the top spot only 9% of the time, against 80% for human-written. AI-assisted content that a person has directed, checked and added to competes perfectly well - the distinction is human involvement, not tooling.

Does Google derank AI content after a core update?

Core updates don't target AI authorship; they reassess quality across the board. That's why sites full of thin, duplicative AI pages tend to fall while sites using AI carefully often don't move at all. If you dropped during one, look at information gain and duplication before you look at your tooling.

What should I check first if my AI-assisted content isn't ranking?

Information gain. Open your page next to the current top three and find what yours offers that theirs don't - original data, first-hand experience, a genuine position. If you can't find anything, no amount of rewriting by hand will fix it.