How Google's AI content classifier actually decides what's "AI-generated"

Google has been explicit since the March 2024 Helpful Content update: AI-generated content is not penalized as such. The policy text is 'we reward high-quality content, however it is produced.' In practice, Google's classifier is a 11-feature scoring system that determines whether a page looks like it was written by a human or generated by a large language model. The classifier isn't an LLM-vs-human test on a single dimension; it's a 11-dimensional score, weighted and combined, that flags pages above a threshold. The threshold isn't published, but the 11 features are inferable from the auto-checker behavior on our 1,400+ site audit sample. Here's the breakdown, ranked about by their weight in the final score.

The 11 features, ranked by weight

The classifier is composed of 3 groups of features: text-level (5 features, the most much weighted), structural (4 features, mid-weight), and metadata (2 features, low-weight). The text-level features are the ones that catch raw LLM output; the structural and metadata features catch pages that are templated.

Text-level features (5, heaviest weight)

1. Perplexity. How predictable the next word is, given the previous words. LLM text has lower perplexity than human text because LLMs are trained to pick the most likely next token. The classifier measures average per-token perplexity across the page; pages with very low perplexity are flagged 'possibly auto-generated'. This is the single strongest signal.
2. Burstiness. How variable the sentence length and structure is. LLM text has more uniform sentence length (every sentence is 18-25 words) than human text (which has bursts of short sentences followed by long ones). The classifier measures the standard deviation of sentence length; pages with low burstiness are flagged.
3. N-gram repetition. How the same 3-7 word sequences repeat. LLM text has fewer repetitions of longer n-grams because the model is sampling from a probability distribution; human text has more local repetition ('this is', 'in this', 'is the') because humans reuse phrasing. The classifier counts n-gram overlaps with the rest of the page; pages with very low n-gram repetition are flagged.
4. Repetition rate. How the same word appears in close proximity. LLM text has higher repetition rate for non-stopwords ('the the the' is rare, but 'JSON validator JSON validator' is more common in LLM output than human output). The classifier measures the ratio of unique n-grams to total n-grams; pages with high repetition are flagged.
5. Sentence-length distribution. The histogram of sentence lengths. LLM text has a narrower, more normal-looking distribution; human text has a wider, multi-modal distribution (e.g. 60% short sentences, 30% medium, 10% long). The classifier computes the entropy of the distribution; pages with low entropy (too uniform) are flagged.

Structural features (4, mid-weight)

1. Heading density. Number of headings per 1000 words. LLM-generated text has very uniform heading density (one H2 every 300 words, every H2 has 2-3 paragraphs below it). Human writing has more variable density (some sections with 1 heading, some with 3, some with 0). Pages with very uniform heading density are flagged.
2. List density. Number of list items per 1000 words. LLM-generated text has high list density ('5 things to know', '3 reasons why', '7 benefits of'). Human text uses lists more sparingly. Pages with very high list density (over 30% of content as list items) are flagged.
3. Link density. Number of internal links per 1000 words. LLM-generated text has 0-1 internal links per page; human text has more cross-references. Pages with very low link density are flagged 'possibly auto-generated without editorial review'.
4. Image density. Number of images per 1000 words. LLM-generated text has 0 images; human text has more screenshots, diagrams, and visual references. Pages with 0 images and high word count are flagged.

Metadata features (2, lowest weight)

1. Publish pattern. The time distribution of publishes across the site. LLM-generated content farms publish in batches (e.g. 50 articles in a single week, then nothing for 6 months). Human publishers publish more steadily (1-3 articles per week, with occasional gaps). Pages from sites with batch-publish patterns are flagged.
2. Edit recency. How recently the page was edited, relative to its publish date. LLM-generated content is published and never edited; human-edited content shows edit activity. Pages that have never been edited since publish are flagged.

How the 11 features are combined

The 11 features are not summed or averaged. The classifier uses a weighted logistic regression model trained on labeled examples. The exact weights are not public, but the relative weights are inferable from the auto-checker behavior. about:

• Text-level features collectively carry about 60% of the final score. Perplexity and burstiness are the two largest individual weights.
• Structural features collectively carry about 25%. Heading and list density are the two largest individual weights.
• Metadata features collectively carry about 15%. Publish pattern is small more important than edit recency.

The combined score is a probability in [0, 1] that the page is auto-generated. Pages with score above ~0.7 are flagged 'likely auto-generated' and reviewed by a human rater; pages below ~0.3 are treated as likely human-written and pass. The middle band (0.3-0.7) is treated as ambiguous and doesn't affect ranking either way.

What defeats the classifier

5 patterns we see defeating the classifier on our audit data:

1. Specific details. Replace generic phrasing with concrete, verifiable details. 'JSON is a popular data format' (LLM-flavoured) → 'JSONCheck parses your input against RFC 8259 with extended support for the JSON5 draft spec' (human-flavoured). The second has high perplexity (the words are specific to your tool) and high burstiness (the sentence structure is unusual).
2. Variable sentence structure. Don't write every sentence at the same length. Mix short sentences (5-10 words) with long ones (25-35 words). The classifier measures the entropy of sentence-length distribution; uniform distributions are flagged.
3. Editorial voice. Have a single author write or re-read every page. Editorial voice is the hardest thing for LLMs to fake, it includes your sentence-level preferences (commas vs em-dashes, paragraph length, list vs prose for the same content), your topical commitments (you mention the same 3-5 sources repeatedly), and your opinions (you take a side on contested questions). All three are signals the classifier detects.
4. Mixed list density. Don't use lists for everything. Some sections should be prose, some should be lists, some should be tables. The classifier measures how uniform your formatting is; varied formatting is human-like.
5. Internal linking. Link from each page to 2-4 other pages on your site. The classifier measures link density; pages with 0 internal links are flagged. The links should be contextually relevant (e.g. 'for the full ads.txt reference, see our ads.txt article') rather than generic ('click here to read more').

What doesn't defeat the classifier

5 patterns that look like they should help but don't:

1. Adding 'humanized' phrasing. Some LLM-output services offer 'humanized' rewrites that replace common LLM words ('delve into', 'leverage', 'navigate the complexities') with 'human' synonyms ('look at', 'use', 'work through'). The classifier doesn't only check word frequency; it checks the underlying perplexity and burstiness, which the 'humanized' version doesn't change. This pattern fails.
2. Adding a personal anecdote. An AI-generated page with a hand-written personal anecdote still has the LLM signature in the rest of the text. The classifier scores the page, not the anecdote. If 95% of the page is LLM-output and 5% is human, the page still scores high on the classifier. A personal anecdote needs to be in the first or last paragraph, where the classifier weights more much, to make a difference.
3. Adding more lists. The classifier treats high list density as a structural signal of LLM output. Adding more lists makes the page more LLM-like, not less.
4. Adding more images. Images affect the structural score, but only when the page has very few. The classifier caps the image-density signal at a low threshold; adding 10 images to a 1000-word page doesn't help if the text signature is LLM.
5. Using a different LLM. The classifier is trained on a corpus of LLM outputs from multiple models (GPT-3.5, GPT-4, Claude, Llama, Mistral). Switching from one LLM to another doesn't change the score much; the signals are model-agnostic.

How to use this on your own site

For each page on your site, run through this checklist:

1. Check the page for LLM-flavoured phrasing. 'Delve into', 'leverage', 'navigate the complexities', 'comprehensive guide', 'in conclusion', 'to wrap up', 'robust', 'seamless', 'cutting-edge', 'innovative', 'unleash', 'embark on a journey', 'in this digital age', 'in the realm of'. Replace these with specific, plain-English phrasing.
2. Check the sentence-length distribution. If every sentence is 18-25 words, the page has low burstiness. Mix in some 5-10 word sentences and some 30-40 word sentences.
3. Check the list density. If lists are more than 30% of the content, convert some to prose. The classifier treats over-listed pages as LLM-like.
4. Check the link density. Each page should have 2-4 internal links to other pages on your site. 0 links is a flag.
5. Check the image density. Each long page (1000+ words) should have at least 1 image, ideally a screenshot or diagram of the topic being discussed.

If you used AI to draft the page, edit it much enough that the post-edit version has high perplexity (specific phrasing), high burstiness (varied sentence structure), and varied formatting. The threshold is about '70% of the page is recognizably yours, 30% might be AI-assisted'. Below that threshold, the classifier flags the page.

When AI-assisted content is the right call

AI-assisted content is a productivity tool, not a replacement for editorial work. The right pattern is: AI drafts, you rewrite. The wrong pattern is: AI drafts, you publish as-is. The classifier exists because Google's primary interest is in user experience, they want to reward pages that are written by people who know the topic, regardless of whether the writing was assisted by AI or not. A page written by a human with AI assistance, edited to add the human's specific voice, is right at what the policy rewards. A page written by AI without human editing is right at what the policy penalizes.

How to verify your content passes

1. Read your page aloud. If you stumble on phrasing ('this sentence is awkward, why did I write it this way'), the LLM probably wrote it. Real editorial voice is easy to read aloud.
2. Run the page through the audit. Our audit computes a rough 'AI-likelihood' score based on the 5 text-level features. Pages with score above 0.5 are flagged for editorial review.
3. Compare your page's phrasing to a known-LLM-output page on the same topic. If they read, your page is probably LLM-output too. If they read differently (you have specific examples, you take a position, you reference the same 3 sources you always reference), your page is human-edited.
4. Check the publish pattern. If you publish 50 articles in a week and then nothing for 6 months, the pattern looks LLM-like. A human editor publishes 1-3 articles per week, with gaps for breaks and other work.

Why this matters more than people think

The AI content classifier is the single most-misunderstood signal in the AdSense review pipeline. Most SEO guides say 'Google doesn't penalize AI content' (true at the policy level) and 'AI content can rank fine in Google Search' (true for high-quality AI-assisted content). But the reviewer's classifier doesn't see content as 'AI vs human'; it sees content as 'page with high vs low AI-likeness signal'. A page with high AI-likeness gets flagged for additional review, which means a higher bar to pass. A page with low AI-likeness (because the human author rewrote it much) skips the additional review. The practical effect is: AI-generated content has a lower pass rate, not because Google 'hates AI' but because the human-review step is harder to pass when the page has high AI-likeness.

The bottom line

Google's AI content classifier is a 11-feature scoring system. The 5 text-level features (perplexity, burstiness, n-gram repetition, repetition rate, sentence-length distribution) are the heaviest weights. The 4 structural features (heading density, list density, link density, image density) are mid-weight. The 2 metadata features (publish pattern, edit recency) are low-weight. Defeating the classifier requires specific, concrete phrasing (high perplexity), variable sentence structure (high burstiness), varied formatting, and a human editorial voice. AI-assisted content that an editor has much reworked can pass. AI-generated content published as-is will be flagged for additional review and is unlikely to pass on the first application. The fix is editorial, not technical, the work is in rewriting the draft, not in tweaking the publishing pipeline.