5 min read
No. The honest answer to "is AI content penalised by Google"is that there is no penalty for using AI, and Google has said so in plain language: how a page was produced is not the ranking question. What the page actually does for the reader is.
That answer takes fifty words. The useful part is everything after it, because sites do lose traffic after they start publishing AI-assisted content, and the reasons are boringly specific. Cannibalisation. Publishing spikes. Unchecked claims. Duplicate schema. None of those are penalties for using a language model. All of them are avoidable before you publish.
Why the question keeps coming back
Two changes in Google's own wording explain most of the confusion, and most people missed both.
In September 2023, the helpful content guidance quietly changed "written by people, for people"to "created for people". One word, but it removed the last line anyone could point to that implied a human hand was required. Then in March 2024, Google's spam policy update renamed "spammy automatically-generated content"to scaled content abuse. The old name described a method. The new name describes a behaviour: producing lots of pages primarily to manipulate rankings, whether a person typed them, a tool generated them, or both.
Read those two changes together and the position is clear. Google stopped caring who or what held the pen, and started measuring volume, intent and quality.
"Penalty"is also the wrong mental model, and it sends people looking for something that will never appear. A manual action is a human reviewer at Google flagging your site, and it shows up in the Manual Actions report in Search Console with a description and a reconsideration route. An algorithmic demotion shows up nowhere. There is no message, no flag, no appeal. So when a founder says "I think we've been penalised for AI content", the first thing to do is open that report. It is nearly always empty. Which means the traffic went somewhere else for reasons the site owner can usually find themselves.
For a small business publishing one or two AI-assisted articles a week, briefed properly and edited by someone who knows the subject, the ranking risk from the AI part is close to zero. The risk lives in the process around it.
What Google's spam policies actually target
Read the spam policies in Google Search Essentials and the pattern is consistent: they describe outcomes, not tooling.
Scaled content abuse covers producing many pages for the primary purpose of manipulating search rankings, regardless of how they were made. The word doing the work is "primarily". Fifty location pages with the town name swapped and nothing else changed qualify, and always did, even when a human copy-pasted them in 2014. One 1,200 word article a week, briefed against the live top 10 for the query, edited, fact-checked and linked to the rest of your site, does not.
Two other policies get misread as AI penalties. Site reputation abuse is third-party content published on a strong domain to borrow its ranking strength, the classic case being coupon or casino sections bolted onto a news site. Expired domain abuse is buying a domain for its history and repopulating it with unrelated content. Neither has anything to do with who wrote the words, but both get bundled into the same anxiety.
The other structural change worth knowing: the helpful content system is no longer a separate named update. Google folded that signal into its core ranking systems with the March 2024 core update. Practically, that means quality assessment now moves continuously rather than arriving as one dramatic event you can point at on a chart. Nobody is going to publish a blog post telling you what happened to your site.
Can Google tell that AI wrote it, and does it matter?
Wrong question. Google is not running a detector and deducting marks; it is assessing the page. Whether a paragraph came from a model, a junior copywriter or a founder at 11pm makes no difference to how it is measured.
Third-party AI detectors have no bearing on rankings at all. They are not what Google uses, and a 98 per cent "AI"score on some tool tells you nothing about whether the page will rank. Cleanly edited human copy gets flagged all the time, because the thing detectors are really measuring is predictability, not authorship.
Here is the part worth taking seriously. The traits detectors latch onto are the same traits that make a page useless:
Uniform sentence length, paragraph after paragraph, with no change of pace.
Vague hedging that commits to nothing: "can be beneficial", "may vary depending on your needs".
Recycled connective phrasing stitching sections together.
Claims with no source, no date, no first-hand detail and nothing only your business could say.
Most AI writing has a tell. It is competent, it is grammatical, and it sounds like nobody in particular. That is what fails on quality signals, and that is the real mechanism behind lost traffic. Fixing the writing fixes both problems. Gaming the detector fixes neither, and the paraphrasing tools sold for that purpose usually make the prose worse.
The patterns that really cost sites traffic
These are the ones we see repeatedly, in rough order of how much damage they do.
Cannibalisation, which is the actual risk for most SMEs
This is the big one and almost nobody names it. A team gets access to an AI writing tool, generates a list of twenty topics, and publishes. Three of those topics are near-duplicates of pages already on the site, and two more overlap each other. Now you have several URLs targeting one query, splitting internal links, relevance signals and click data between them. Google picks one, often the weakest, and the rest sit on page three.
The fix costs nothing. Check every planned title against your existing pages and everything you have already published, before a word is written. That single step removes most of the risk people attribute to AI.
The cold-start publishing spike
Cadence is a signal people ignore. A site that has posted nothing for two years and then puts out forty pages in a fortnight has changed shape overnight, and it looks like a different site to a crawler than one that has been publishing weekly for months. There is no published threshold and anyone who quotes you one is guessing. But ramping gradually, with a consistent weekly rhythm, is a deliberate safeguard rather than a limitation, and it costs you nothing except patience.
Unchecked facts surviving into the published draft
Models invent statistics with total confidence. They also produce figures that were true three years ago: old VAT thresholds, superseded standards, prices that have moved. If nobody reads the draft sceptically, those go live under your company name. The reputational damage lands long before the ranking damage does, usually when a client emails to ask where the number came from.
Technical own-goals blamed on the AI
A large share of "AI content penalty"cases turn out to be markup. A plugin or tool injects Article or FAQ schema into a page that already has Yoast or Rank Math outputting its own, and the page ends up with duplicate structured data, validation errors and lost rich results. The copy was fine. The schema was doubled. Detecting the existing SEO plugin before injecting anything is a five-second check that prevents an error which has nothing to do with who wrote the words.
Same category: pages published with no internal links in or out, and generic stock or AI-generated imagery that looks like everyone else's.
How to publish AI-assisted content that holds up
The method matters more than the model. This is how we run it, and the sequence is the point.
Research first, draft second. Every piece is planned against the live top 10 for its target query, so the article answers the intent Google has already decided on rather than the intent someone assumed in a meeting. Real UK search volume and difficulty data, not guesses. If the top 10 is all comparison pages, a listicle will not rank there, however well written.
Build on a brand-voice profile taken from your own writing. Not "friendly and professional", which is a vague adjective standing in for a decision. We read your website like a new hire would, we crawl it politely, and we build the profile from what is actually there: sentence length, vocabulary, how you open and close a point. Honest caveat: on a thin site with four pages of copy, there is not much to learn from, and the profile will be thinner too. Better to say that than pretend otherwise. Our page on how we write and edit every article sets out the full sequence.
Two grades, not one. Grading a draft on SEO alone will always reward the wrong article, because keyword coverage is easy to satisfy and says nothing about whether the page is worth reading. Every piece gets an SEO score and a content integrity score, and integrity wins any tie. A sceptical sub-editor pass hunts for unsupported claims, invented figures and stale dates. A banned-phrase list strips the tell-tale wording before it reaches you. The two-score grading and the rest of the editorial checks run before you ever see a draft.
Then make it sound like a person. Deliberate rhythm variation, short sentences against long ones. Named examples. First-hand specificity: the trade-off you hit, the client scenario, the decision that went badly. That is what genuinely makes AI content sound human, and no amount of synonym-swapping substitutes for it.
Ramp the cadence and check the gaps. A weekly rhythm that builds gradually, with keyword-gap checks against crawled competitors so each new piece earns its own slot rather than colliding with the last one.
A pre-publish check you can run in ten minutes
One thing only you could say. A number with its source, a client scenario, a decision you made, a trade-off you have hit. If the whole article could sit on a competitor's site with the logo swapped, it is not finished.
Duplication check. Does another page on your site already target this query? Does the title differ meaningfully from every existing URL? Site-search your own domain before publishing, every time.
Facts, dates, prices, standards. Checked against a current source, in the last week. Would you defend each one to a sceptical client on the phone? If not, cut it or make the point qualitatively.
Structure. One schema source only. Internal links to and from related pages. Meta title within length. Images that look like your brand rather than generic AI art.
If traffic has already dropped. Open the Manual Actions report in Search Console first. If it is empty, stop blaming an AI penalty and audit the thinnest 20 per cent of your pages: merge, improve or remove them before publishing anything new.
Google's own guidance on creating helpful, reliable content is worth reading in full once, ideally by whoever signs off your drafts.
So the dividing line was never human versus AI. It is expert and original versus generic and thin. Get the process right and the answer to "is AI content penalised by Google"stays a flat no for your site too. If you want to see where your current pages sit against that standard, start with the free content marketing snapshot, or read how the research, drafting and editing pipeline works end to end.
Frequently Asked Questions
Is AI content penalised by Google if I edit it before publishing?
No, and editing is not what makes it safe either. Google assesses the finished page against its quality and spam policies regardless of production method, so a lightly reworded but empty article is still weak, and a well-researched AI-assisted piece with real expertise in it is still strong. Edit for accuracy, specificity and originality, not to disguise the tool.
Can Google tell whether an article was written by AI?
Assume it can, and assume it does not matter. Google evaluates what the page delivers to the reader rather than how it was produced, and third-party AI detector scores have no bearing on rankings. The traits those detectors flag, uniform rhythm and vague hedging, are worth fixing anyway, because they are what makes a page unhelpful.
What counts as scaled content abuse?
Producing many pages primarily to manipulate search rankings rather than to help readers. Google renamed the policy from "spammy automatically-generated content"in its March 2024 spam update, deliberately shifting the focus from method to volume and intent. Fifty templated location pages with the town name swapped is the textbook case; a weekly researched article is not.
Do I need to disclose that an article was AI-assisted?
Google does not require it for ranking purposes. Disclosure is an editorial and trust decision, and it matters more in regulated sectors, on medical or financial advice, or where your audience would reasonably expect a named human author. What does matter either way is that a named person takes responsibility for the accuracy of what is published.
How many AI-assisted articles a week can a small site safely publish?
There is no published threshold, and anyone quoting a hard number is guessing. What we recommend is a steady weekly rhythm that ramps gradually rather than a standing start into dozens of posts, with every planned title checked against existing pages first. For most SMEs, one or two properly briefed and edited pieces a week is sustainable and avoids the two real risks: cannibalisation and thin filler.