Does Google penalise AI content? What the data shows

The question arrives in our inbox every month, usually from a marketing manager who has just been quoted ₹40,000 for "200 SEO articles" and wants to know whether Google will punish them for buying it. Google has published its position in plain English, several times, and almost nobody selling content at that price has read it.

This post reads what Google says, then what has happened since, and turns it into a checklist a small Indian business or agency can follow. All quotations below are from Google's own pages or from named industry coverage, fetched on 3 November 2026, and listed in the sources.

What does Google actually say about AI-generated content?#

Google's clearest statement is a Search Central blog post from February 2023, titled with a heading that does most of the work: "Rewarding high-quality content, however it is produced." The key sentence is this one: "Appropriate use of AI or automation is not against our guidelines."

The same post draws the line that still applies today. Using automation "with the primary purpose of manipulating ranking in search results is a violation of our spam policies." Note what the test is. Not the tool. The purpose.

The post also connects the question to what Google calls E-E-A-T, saying its ranking systems aim to reward "original, high-quality content that demonstrates qualities of what we call E-E-A-T," listed as "expertise, experience, authoritativeness, and trustworthiness." On authorship it says: "You should consider having accurate author bylines when readers would reasonably expect it." On disclosure, it says AI or automation disclosures "are useful for content where someone might think" the reader would want to know how it was made, and adds that "giving AI an author byline is probably not the best way" to make that disclosure.

Google's newer guidance page on using generative AI content is shorter and more practical. It says generative AI "can be particularly useful when researching a topic, and to add structure to original content," tells publishers to "focus on accuracy, quality, and relevance, especially when automatically generating the content," and repeats the warning that "using generative AI tools or other similar tools to generate many pages without adding value for users may violate" the spam policies. On disclosure it says that "sharing information about how a piece of content was created can help give your readers more context."

That is Google's whole stated position on AI as a method. It is permissive. The restrictions arrive from a different direction.

What is the scaled content abuse policy?#

On 5 March 2024 Google added three named policies to its spam documentation: expired domain abuse, scaled content abuse and site reputation abuse. The first two took effect immediately; site reputation abuse was deferred to 5 May 2024.

The spam policies page defines the one that matters here: "Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users. This abusive practice is typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it's created."

The examples Google lists are worth reading in full, because most "SEO content packages" sold in India fit at least one:

• "Using generative AI tools or other similar tools to generate many pages without adding value for users"
• "Scraping feeds, search results, or other content to generate many pages (including through automated transformations like synonymizing, translating, or other obfuscation techniques) where little value is provided to users"
• "Stitching or combining content from different web pages without adding value"
• "Creating multiple sites with the intent of hiding the scaled nature of the content"
• "Creating many pages where the content makes little or no sense to a reader but contains search keywords"

Google's announcement on The Keyword made the method-neutrality explicit, describing the policy as targeting content produced at scale to boost search ranking "whether automation, humans or a combination are involved." The older spam policy this replaced was called "spammy automatically-generated content." Renaming it was a deliberate signal that a room full of underpaid writers producing the same output is treated the same way as a script.

The consequence is stated on the Search Essentials page: a violation "can lead to a page or an entire site being ranked lower or completely omitted from Google Search."

What does the helpful content guidance ask you to check?#

Google's "Creating helpful, reliable, people-first content" page is a list of self-assessment questions rather than a rule book, but the questions are specific enough to use as an editorial checklist. The ones most relevant to AI-assisted publishing:

• "Does the content provide original information, reporting, research, or analysis?"
• "Does the content provide substantial value when compared to other pages in search results?"
• "Does the content provide insightful analysis or interesting information that is beyond the obvious?"
• "Is this content written or reviewed by an expert or enthusiast who demonstrably knows the topic well?"
• "Are you producing lots of content on many different topics in hopes that some of it might perform well in search results?"
• "Are you using extensive automation to produce content on many topics?"

The page's "Who, How, and Why" section is where AI comes in directly. "Who" asks: "Do pages carry a byline, where one might be expected?" "How" asks: "Is the use of automation, including AI-generation, self-evident to visitors through disclosures or in other ways?" and "Are you providing background about how automation or AI-generation was used to create content?" The page also states that of the E-E-A-T aspects, "trust is most important."

Read together, Google's three documents say one consistent thing. It does not ask whether a model touched the text. It asks whether a real person stands behind it, whether it contains something the reader could not get elsewhere, and whether the site is publishing to help people or to fill an index.

What actually happened after March 2024?#

Policy is one thing. Here is the documented enforcement, from Google first and then from the trade press.

Google's own account. The 5 March 2024 Keyword post said: "we expect that the combination of this update and our previous efforts will collectively reduce low-quality, unoriginal content in search results by 40%." The same post carries an update dated 26 April 2024: "As of April 19, we've completed the rollout of these changes," and "You'll now see 45% less low-quality, unoriginal content in search results." Search Engine Roundtable quotes a Google spokesperson in the same words. The Search Status Dashboard lists the March 2024 core update at 45 days and the March 2024 spam update at 14 days and 21 hours, both starting 5 March. Search Engine Land notes that with this update Google folded the helpful content system into its core ranking systems and stopped announcing separate helpful content updates.

The 45% figure is Google's self-reported measurement against its own definition of low quality; nobody outside Google can audit it, but it is the only quantified outcome Google has published.

Manual actions. Separately from the algorithmic update, Google issued a wave of manual actions in the first days of the rollout. Search Engine Journal's coverage from 6 March 2024 collected reports of sites removed from the index overnight; Lily Ray reported "10 websites completely removed from Google search overnight," and an affiliate publisher's network of eight sites went to zero impressions. Search Engine Journal's later report on the spam update's completion records that Google "issued numerous manual actions to address pure spam issues" after introducing the new policies, and distinguishes them from the algorithmic update: manual actions "are carried out by Google's human reviewers" and arrive with a Search Console notification, while algorithmic demotions do not.

Scale. The most quoted count comes from SEO Ian Nuttall, reported by Search Engine Journal: tracking "49,345 websites," he found "837 websites had been removed entirely from Google's search index," or "1.7% of sites got hit." That is one person's sample of sites already on an SEO watchlist, so it is indicative, not a population figure.

The AI correlation. The same Search Engine Journal piece reports an Originality.ai study claiming "100%" of the deindexed sites in its sample showed signs of AI-generated content, with "50%" having "90-100%" of posts AI-generated. Two cautions. Originality.ai sells an AI detector and ran its own detector to produce the finding. And the correlation runs the way the policy predicts: sites that mass-produce pages overwhelmingly used AI to do it, because that is the cheap way to mass-produce. Search Engine Journal's own reporting included a site started in 2013 with human-written content that also received a manual action. The common factor was volume without value, not the tool.

Where exactly is the line?#

The table below is our reading of Google's published documents against common publishing practices. Risk levels are a judgement, not a Google rating.

The low-risk rows still involve AI, and the high-risk rows would be just as risky written by hand. The axis is value and scale, not authorship.

What should an Indian SMB or agency actually do?#

Lacewing Technologies is an AI and software development company in Navi Mumbai. We build and operate our own products and we also build SEO content systems for clients, which means we have had to decide where this line sits in practice rather than in theory. This is the checklist we use.

Start from first-party data. The cheapest way to pass the "original information" test is to publish what only you know: your actual price list, how long a job really took, the three questions every customer in Thane asks, the mistake you see in half the invoices you receive. A model can draft around that; it cannot supply it.

Put a named person on it. Every post carries a real author with a real role, and that person has read the piece before it goes out. Google's guidance asks for bylines "where one might be expected," and on a business site a reader expects one. Do not list the AI as the author; Google's own post says that is "probably not the best way" to disclose.

Review before publishing, every time. A human reads the whole piece for factual errors, invented statistics and hedged filler. If the reviewer cannot say what the post tells a reader that the top five results do not, the post does not go out.

Disclose how you work, once, somewhere findable. An editorial policy page stating that drafts may be AI-assisted and are reviewed by a named editor satisfies the "How" question and costs an afternoon.

Cap your volume to what you can review. If you can genuinely review two posts a week, publish two posts a week. The moment the pipeline produces faster than a person can read, you are describing scaled content abuse to yourself.

No templated location pages without local content. A page for each of 30 suburbs is fine if each one contains something true about that suburb: a client, a photo, a travel time, a regulation. Thirty pages that differ only in the place name are the first example in Google's list.

Keep the site's purpose legible. A chartered accountant's blog that suddenly covers cricket scores and crypto because "those keywords are easy" is answering yes to "producing lots of content on many different topics in hopes that some of it might perform well."

Our own blog runs on a 2-day publish gate: one post every two days, each one built against a quality check that rejects anything under 1,200 words, anything without sources, anything with the usual machine-sounding phrases, and anything that overlaps more than 20% with an existing post. It just makes the lazy version impossible to ship by accident.

Where do AI detection tools fit for a publisher?#

Not where most people put them. Google has never said it runs content through an AI detector and ranks on the result, and the detection industry's own documentation is careful about what a score means. Grammarly's detector page, for example, says "No AI detector is 100% accurate" and that you "should never rely on the results of an AI detector alone to determine whether AI was used."

Where a detector earns its place is upstream, in internal quality control. If you commission copy from freelancers or an agency, you have a contractual question (did I get what I paid for?) and an editorial one (is this thin?). A detector gives a fast first signal on both. Prose that scores as highly machine-like usually also reads as generic, because the same statistical smoothness that detectors measure is what unedited model output looks like. Treat a high score as a prompt to read closely, not as a verdict and not as a Google proxy.

Tools a publisher might use for that, with what each vendor says about itself:

1. Originality.ai positions itself for "Website Owners, Content Marketers, Writers and Publishers," bundles AI detection with plagiarism, fact and readability checks, and offers bulk scans and full-site scans. Pro is listed at $14.95 a month ($12.95 billed yearly); the Enterprise tier is described as "ideal for agencies & publishers."
2. GPTZero describes itself as "the most trusted AI content checker in education" and says it also serves "publishers, recruiters, copy-writers, marketers." It offers sentence-level highlighting, a plagiarism checker and a Writing Replay feature for documents written in its editor, plus an API.
3. Grammarly AI Detector is free, returns a percentage score, and includes the honest caveats quoted above. Paid Grammarly Pro adds passage-level highlighting. It states its training data includes AI text "created before 2021," which is worth knowing when checking copy from current models.
4. TextSight, which we build and operate, returns a score plus the passages that drove it, and is positioned explicitly as a signal for review rather than a verdict. The free tier scans up to 1,500 words with an account; Starter is $9.99 a month ($7.49 billed annually, ₹899 in India) for 5,000 words per scan, and Pro is $19.99 ($14.99 annually, ₹1,799) for 7,500 words per scan with API access. It shares the limits of every detector: short passages, heavily edited text and non-English content all reduce reliability.
5. HumanFlow, our other product, is primarily a rewriting tool for rhythm and specificity but includes a detection score on the free tier (10,000 detection words a month, 1,500 per scan). It is the right pick only if you also intend to rewrite flagged copy yourself rather than send it back to the writer.

None of these tells you what Google is measuring. They tell you whether a draft needs a closer read before a human puts their name on it, which is the step Google's guidance cares about.

The short version#

Google has said, in writing, that it does not penalise AI content as such and does penalise scaled, unoriginal content whoever produced it. It then demonstrated the second half with a core update it credits with 45% less low-quality content and a wave of manual actions that removed hundreds of sites in a week.

If you publish fewer pages than you can personally vouch for, built on data only you have, with your name on them, the method you used to draft them is not the thing that will decide your rankings. If you publish more pages than anyone has read, it will not matter who wrote them.