Does Google penalize AI content? Across 600k pages the correlation with rankings is 0.011. The real risk is scaled emptiness. 4 edits that de-risk your AI blog.
Does Google penalize AI content? No, not for being written with AI. The largest public study on this, an Ahrefs analysis of 600,000 pages, found a correlation of 0.011 between how much AI a page contains and where it ranks. That is statistically indistinguishable from zero. In the same dataset, 86.5% of top-ranking pages already contained some AI-generated content. Authorship is not the thing Google grades.
What Google does penalize is scaled, unhelpful content produced to game rankings, and it applies that judgment "no matter how it's created." So the real risk with an AI-written blog is not that a model touched it. The risk is that the model wrote something empty and you shipped it without checking. This post separates the two, with the data, and ends with the four edits that turn raw AI output into a page that ranks. Those four edits also happen to be exactly what a fact-checked, human-merge pull-request pipeline enforces on every post.
No. Google does not have a penalty for AI authorship, and there is no evidence in the data that it ranks pages down for being AI-assisted. It penalizes two things that have nothing to do with who or what wrote the text: scaled manipulation, and pages that do not help the person who searched. A page can hit both of those while being written entirely by a human. A page can avoid both while being drafted by a model. The production method is not the variable.
This is the same line Google drew during the scaled-content-abuse crackdown, and it is worth internalizing before you read another "AI content is dead" take. The pages that lost rankings in 2024 and 2025 were not punished for using automation. They were punished for being unoriginal at scale. If you run content as a growth channel, which is the whole premise of SEO for SaaS, the distinction decides whether automation compounds for you or quietly sinks your domain.
The penalty target is intent and value, not tooling. Google's spam policy and its public statements on AI both say the same thing in different words: produce unhelpful content at scale to manipulate search, and you are in violation, however you made it. Produce genuinely useful content, and the tool you used to write it does not count against you.
Google defines the violation precisely. In its spam policies documentation, scaled content abuse is "when many pages are generated for the primary purpose of manipulating search rankings and not helping users." The policy is "typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it's created."
Read that last clause twice. "No matter how it's created" means a human content farm and an AI content farm are judged identically. The method is explicitly removed from the equation. What is left is purpose (manipulating rankings versus helping users) and value (original and useful versus thin and unoriginal). Google published this language alongside its March 2024 core update and later reported the result: "You'll now see 45% less low-quality, unoriginal content in search results versus the 40% improvement we expected." That is a quality cleanup, not an anti-AI campaign.
Google has been unusually direct about AI specifically. In its guidance on AI-generated content, it states: "Rewarding high-quality content, however it is produced, is key to what we do." In the same guidance it draws the boundary: "using automation, including AI, to generate content with the primary purpose of manipulating ranking in search results is a violation of our spam policies," while making clear that not all use of automation, including AI generation, is spam.
So the rule, in Google's own framing, is not "do not use AI." It is "do not use AI (or anything else) to flood search with pages built to rank rather than to help." That is the same standard applied to automated content creation generally: automation is a multiplier, and what it multiplies is your decision. Point it at real substance and it scales an asset. Point it at filler and it scales a liability.
The strongest evidence that authorship is not the penalty comes from the largest public analysis of the question. The picture it paints is not "AI is safe so publish anything." It is more specific and more useful than that.
On July 7, 2025, Ahrefs researchers Si Quan Ong and Xibeijia Guan published an analysis of 600,000 pages, the top 20 ranking URLs for 100,000 random keywords, run through Ahrefs' own AI-content detector. The correlation between a page's AI-content share and its ranking position was 0.011. For context, correlations run from -1 to 1, and 0.011 is the kind of number you get from noise. There is no meaningful relationship in either direction. AI content is not a ranking boost and it is not a ranking penalty.
The second headline number is the one that ends the debate for most people: 86.5% of top-ranking pages contained at least some AI-generated content. The pages already winning in search are, overwhelmingly, pages that used AI somewhere in their production. If Google were penalizing AI authorship, that number would be near zero. It is the opposite.
Averages hide the interesting part, so look at the distribution. Of the top-ranking pages Ahrefs studied:
| Content type | Share of top-ranking pages |
|---|---|
| Pure AI (no detectable human edits) | 4.6% |
| Pure human (no detectable AI) | 13.5% |
| Human and AI blend | 81.9% |
The blend is the real story, because it is most of the web that ranks. Inside that 81.9%, the split was 13.8% minimal AI (1 to 10% of the text), 40% moderate (11 to 40%), 20.3% substantial (41 to 70%), and 7.8% dominant (71 to 99%). The winning pattern is not "all AI" or "no AI." It is a human working with a model, in proportions that vary widely. Drafting with AI and editing as a human is not a workaround. It is what the top of the results already looks like.
There is one faint trend worth reporting honestly because it cuts slightly against the easy conclusion. Ahrefs noted that #1-ranked pages tend to carry a little less AI content than pages ranking just below them, and independent coverage of the study framed it as a faint preference for the top spots among low-AI pages. The effect is real but too weak to treat as a ranking factor. The honest reading: heavy human involvement at the very top is more likely a sign that someone cared enough to do the work, not a dial Google turns.
Here is where "Google does not penalize AI" gets misused. It is true that authorship is not a penalty. It is not true that you can publish unedited AI at scale and rank. Those are different claims, and the data on the second one is brutal.
An experiment documented on Search Engine Land published 2,000 fully AI-generated, unedited articles across 20 brand-new domains and tracked them for 16 months. The pages indexed fast and even earned early impressions and clicks. Then they fell. By months three to six, only 3% of the pages still held a spot in the top 100. An August 2025 spam update briefly lifted that to roughly 20%, but the recovery did not hold. The conclusion the experiment reached: pure unedited AI content, with no authority, no expertise, and no unique insight, cannot sustain rankings.
Put the two studies side by side and the lesson is exact. Authorship is not the penalty. Emptiness is. The 4.6% of top pages that are pure AI are not winning because they are AI; they are winning because, in those cases, the raw output happened to be useful and original enough. The 2,000-article graveyard is what happens when it is not. The difference between the two outcomes is editing, and editing is something you control.
If the risk is emptiness, the fix is substance. These four edits are the difference between an AI draft that ranks and one that gets filtered. None of them are exotic. They are what a careful editor does to any draft, AI or not, applied without exception.
Put at least one original element on the page that exists nowhere else. Original data, a first-hand result, a real screenshot, a computed answer, a number you measured. This is the single most valuable edit, because it is the one thing a model genuinely cannot fabricate from its training data, and it is the thing both Google and AI answer engines reward.
The test is mechanical: strip the templated wrapper off the page, the intro, the recurring headings, the boilerplate, and look at what is left. If something unique and useful remains, you have an asset. If a find-and-replace of one variable is the only thing distinguishing your page from a generic article, you have thin content with extra steps. This is the same strip-the-wrapper test we use for programmatic SEO for SaaS, and it scales down to a single blog post just as cleanly. The original element is also what makes a page citable by AI answer engines, which prefer a specific, sourced number to a vague claim. One original thing per page pays off in blue links and in citations.
Write what the searcher actually wants, not just the keyword string. A model will happily produce a fluent page about the words in your title while completely missing what the person typing them is trying to do. That mismatch is exactly the "does not help users" failure Google's policy describes, and no amount of polish fixes it.
The check takes two minutes: look at the live results for your target query before you write. If the page-one results are comparison tables, the intent is comparison, and a 2,000-word explainer will lose. If they are step-by-step tutorials, the intent is how-to, and a thought-leadership essay will lose. Match the format and depth the results are telling you the searcher wants. Getting this right is most of the work in picking winnable topics for organic SaaS growth, and it is the edit AI drafts skip most often, because a model optimizes for plausible prose, not for the job the searcher came to do.
Show who wrote it and why they are worth trusting. Google's quality systems lean on experience, expertise, authoritativeness, and trust, and an anonymous, source-free page signals none of them. A named byline with real experience, primary sources cited inline, and first-hand evidence are what tell both a reader and a ranking system that a person with standing stands behind the page.
This is also the edit the 2,000-article experiment was missing by design: brand-new domains, no authors, no authority, no expertise. The pages collapsed precisely there. You do not need a famous byline to clear the bar. You need a real one, attached to someone who actually knows the topic, citing sources a reader can check. Experience and verifiable sourcing are also what make content quotable by answer engines, so this edit, like the first, pays off twice.
Never publish raw model output. This is the edit that catches what the first three cannot: the confident statistic that does not exist, the citation the model invented, the link that 404s. A language model predicts plausible text, not true text, so a fabricated number and a real one come out of the same process and read identically. The only defense is to check.
Pull every claim out of the draft and confirm it against a current source. Fetch every external link and confirm it resolves and actually supports the sentence using it. Re-check numbers, dates, and prices against the live source, and date anything volatile. Treat anything you cannot confirm as a blocker, not a footnote. We wrote the full mechanics in how AI content fact-checking works, and the one rule that makes it real bears repeating: verification only counts if it can stop a post from shipping. A confidence score the model assigns to its own output is the model grading its own homework.
The four edits are easy to list and easy to skip. The teams that get burned are not the ones using AI; they are the ones with no step between "draft generated" and "page live." So the durable fix is structural: put a gate between the draft and the index, and make the four edits the gate's job.
This is the shape Lyra is built around. She drafts each post in your blog's existing voice, then does the work most tools skip. She fact-checks every claim against a current source, fetches every link and drops or fixes anything broken, and verifies numbers and dates. She structures each post around the real question behind the query so it matches intent, and she keeps a named byline and sourcing on the page. Then she scores the draft and opens a GitHub pull request and tags you. Nothing auto-publishes. You read the diff, the score, and the fact-check notes, and you merge, or you send it back.
That last part is the point. The penalty risk in the data is unreviewed, unoriginal content shipped at scale, and a pull request is the oldest, most boring fix for exactly that: a human signs off before anything goes live. If your blog already lives in a repo, you can review a post the same way you review code in a pull request, where saying "not yet" is a normal part of the flow. Lyra runs on your own Anthropic key, encrypted at rest and never marked up, so the verification step is not a per-word tax. Want to see how she would gate your blog? Talk to the founder.
The fear that Google will penalize your AI-written blog is aimed at the wrong target. Google does not grade the author. It grades the page. Make every page carry one original thing, match the intent, show real authorship, and verify before you ship, and the question of whether a model helped you write it stops mattering, because the data already says it does not.
The data is clear that Google ranks helpful pages and filters empty ones, and the four edits that keep AI content on the right side of that line are exactly what Lyra enforces: original substance, real intent, real authorship, and a fact-check gate before any pull request opens.
Step by step
Add one thing only you have
Put at least one original element on the page that exists nowhere else: your own data, a first-hand result, a real screenshot, or a computed answer. Strip the wrapper and something unique must remain.
Match the actual search intent
Read the live results for the query and write what that searcher needs, not just the keyword string. A page that answers the wrong intent is unhelpful no matter how clean the prose is.
Put real authorship and E-E-A-T on the page
Add a named byline with real experience, cite primary sources, and show first-hand evidence. Experience and expertise are what separate a useful page from a generic one a model could have written about anything.
Edit the draft and verify every claim and link
Never publish raw model output. Check every stat against a current source, fetch every link to confirm it resolves and is relevant, and date volatile facts. Block anything you cannot confirm.
FAQ
Yes, when it is helpful and verifiable. Across 600,000 pages, Ahrefs found a 0.011 correlation between a page's AI-content percentage and its ranking, which is effectively zero, and 86.5% of top-ranking pages already contained some AI. Google ranks pages on whether they help searchers, not on whether a model helped write them. What does not rank is unedited, unoriginal AI published at scale.
Probably to some degree, but it does not rank on detection. Google's public position is that it rewards high-quality content however it is produced, and its spam policy applies 'no matter how it's created.' Detectors built by third parties are also unreliable. The practical takeaway is that trying to hide AI use is the wrong problem. Making the page genuinely useful is the right one.
No. Google states that using AI is not against its guidelines, and that appropriate use is fine. What violates the spam policy is 'using automation, including AI, to generate content with the primary purpose of manipulating ranking in search results.' The line is intent and value, not the tool. AI inside a real editorial process is allowed; mass unhelpful pages are not.
Make every post earn its place. Add at least one thing only you have (original data, a first-hand result, a real screenshot), match the actual search intent rather than just the keyword, put real authorship and E-E-A-T on the page, and verify every claim and link before publishing. The penalty risk is scaled emptiness, so the fix is substance plus a review step that can say no.
Built by the tool you're reading about
Lyra finds the topics worth ranking for, writes them in your repo's voice, fact-checks every claim, and opens a pull request scored and ready to merge. You review and hit merge. Want to see what she'd write for you? Tell us about your blog and the founder will walk through it with you.
Keep reading
How to get cited by ChatGPT, Perplexity, and Claude. One AEO checklist won't win all three engines, so here's the engine-by-engine playbook, backed by 2026 data.
Google AI Mode and AI Overviews cite different URLs (about 13.7% overlap). How each surface picks sources, and how to optimize for Google AI Mode and win both.
Schema markup for AI search is hygiene, not a lever. What Google says, what the one controlled test found, and which structured data to actually ship in 2026.