Is AI content bad for SEO? What Google actually penalizes, and what still ranks.

1. Does Google penalize AI content? Google's own words.

Google's AI content policy is not a separate rulebook. It is the old one, applied to a new tool. On February 8, 2023, a few months after ChatGPT launched, Google Search Central published its guidance on AI-generated content and answered the question head on: appropriate use of AI or automation is not against its guidelines. The same post said Google focuses on the quality of content rather than how it was produced, and that using AI gives content no special gains in ranking either way. Most of the AI SEO statistics people argue over make more sense once you start from that line.

The current version lives in two documents. The guidance on generative AI content (last updated October 1, 2026) says generative AI can be particularly useful when researching a topic and to add structure to original content, and then sets the bar: it is critical to manually fact-check and review all AI-generated content for accuracy and trustworthiness before publishing. The people-first content guide (updated October 5, 2026) draws the other line: using automation, including AI, to produce content for the primary purpose of manipulating search rankings is a violation of Google's spam policies.

So does Google penalize AI content? Not as a category. Search Console's list of manual actions has no entry for AI content. The two that AI-heavy sites actually receive are Thin content with little or no added value, and Major spam problems, whose description names scaled content abuse. Both existed before ChatGPT, and both hit hand-written pages just as hard. The practical question is never “was this written by AI?”. It is whether the page is worth ranking, which is the same question behind every explainer on what AI SEO means and every honest use of AI in search.

2. Scaled content abuse: the policy that actually bites.

If AI content gets a site in trouble, this is almost always the policy that does it. Google's spam policies define scaled content abuse as many pages generated for the primary purpose of manipulating search rankings and not helping users, typically large amounts of unoriginal content that provides little to no value, no matter how it's created. The examples Google lists are specific:

• Using generative AI tools or 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 or translating.
• 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 introduced the policy on March 5, 2024, alongside a core update aimed at unhelpful pages, and said it expected the changes to cut low-quality, unoriginal content in results by 40%. In an April 26 update, after the rollout finished on April 19, it reported 45%. Note the wording in the announcement: content at scale to boost rankings, whether automation, humans or a combination are involved. A content farm of underpaid writers is just as exposed as a script. What the policy targets is volume without value, which is why the same tactic that fails as AI spam can work as programmatic SEO with real data on every page.

Google's human raters are told the same thing. The January 2025 update to the search quality rater guidelines added a definition of generative AI and told raters to give the Lowest rating when all or almost all of a page's main content is copied, paraphrased, auto or AI generated with little to no effort, originality or added value. Rater scores do not move rankings directly; they grade how well Google's systems are doing. They are still the clearest description of what those systems are built to push down. The usual trigger is not one bad post. It is a site that pointed auto blogging software at a keyword list with no checks and published faster than anyone could read.

3. What gets hit and what still ranks.

For AI generated content, SEO outcomes split along one line: did the page add anything. The table pairs the patterns that get pushed down with the version of the same idea that still ranks. Notice that the right-hand column does not say “written by a human”. Every item on it can be drafted with ChatGPT, Claude or Gemini. What it cannot be is drafted by them alone.

The best public data points the same way. Graphite, a growth agency, published a study in October 2025 that ran the first two pages of Google results for 31,493 keywords across ten categories, collected in June 2025, through an AI detector. It classed 86% of the ranking articles as human-written and 14% as AI-generated, and found that when AI-generated articles did appear, they tended to rank lower. Two caveats matter. Detectors misfire, and the study did not test AI drafts with heavy human editing. Read it as evidence that raw AI output underperforms, not that AI help does. Our own AI SEO case study publishes the raw Search Console numbers for 61 new pages on this site, including the pages that lost.

There is also a site-level effect to plan for. Google's guide to its ranking systems says they work on the page level, and that site-wide signals and classifiers also contribute to how pages are understood. Good site-wide signals do not guarantee every page ranks, and poor ones do not sink every page, but a pile of thin pages is not free. It is part of how the whole site is read, which is why recovering from a drop usually starts with fixing or merging the weak pages you already have, the work our guide to optimizing existing content with AI walks through, rather than with publishing more.

4. Why raw AI drafts fail the test.

Google's guidance on generative AI content explains the core problem in one sentence: generative models don't retrieve facts, they predict a likely sequence of words based on their training data. That is why a raw draft fails in four predictable ways. It states numbers nobody measured. It states facts that were true when the model was trained and are not now. It says what every other page on the topic already says, because those pages are what it learned from. And it reads like it, with the stock phrasing our guide to making ChatGPT write like a human lists.

None of the four is fixed by switching models. A newer ChatGPT or Claude writes more fluent prose, and fluent prose makes an invented statistic harder to spot. Web search helps, because Claude, ChatGPT search and Perplexity show the pages behind an answer, but a citation next to a sentence is not proof the cited page says it. The fix is in the workflow: give the model real inputs and check what comes out. The ChatGPT for SEO setup and the Claude writing workflow both start there.

The fifth failure is the byline. Faced with E-E-A-T advice, some sites invent an author: a stock name, an AI headshot, a bio full of credentials. Google's people-first content guide now names this directly. Fabricating creator profiles, such as by using AI-generated headshots, made-up names or false credentials to make content look written by human experts, is a form of deception, and any deception makes a page untrustworthy to readers and to Google's quality systems. Its 2023 FAQ added that giving AI an author byline is probably not the right fix either. Put a real person or team on the page, and where readers would wonder how it was made, say so.

5. How a fact check and a quality gate change the answer.

Here is the real answer to “is AI content bad for SEO”: it depends on what happens between the draft and the publish button. Without a check, AI content carries every risk in the sections above. With one, most of those risks are gone, and what is left is an ordinary editorial question about whether the page is good. The check has two parts. A fact check, which catches what is wrong. And a gate, which decides whether the page is ready at all.

Start with the gate. It is six questions, written down so nobody skips them on a busy week, and a page that fails any one of them does not go live. The first two come straight from Google's people-first guidance, which asks whether content is created primarily to help people and whether it provides original information, reporting, research or analysis. The last one keeps a site out of scaled content abuse territory and answers another question on the same page, whether you are using extensive automation to produce content on many topics: you publish what you can check, and nothing more. Add it to whatever SEO checklist you already run.

BEFORE ANY AI-ASSISTED PAGE GOES LIVE

1. Purpose     Would this page exist if search engines did not?
               If the honest answer is no, do not publish it.
2. Added value Name one thing on this page the top five results
               do not have: data, a test, a worked example, a
               first-hand detail. None = not ready.
3. Claims      Every number, date, price, name and quote is
               paired with a source sentence you opened.
               Unsupported = cut or soften.
4. Metadata    Title, meta description, structured data and
               image alt text get the same check as the body.
5. Byline      A real person or team. No invented experts.
6. Pace        Publish only as many pages as you can run
               steps 1-5 on. The rest wait.

The fact check is step 3, and it is the one people do badly, usually by asking the model that wrote the draft whether it is accurate. Run it as a separate pass in a fresh chat, give it the draft and the source pages you opened, and make it pair every claim with the sentence that supports it. Anything it cannot pair gets cut or softened. Google's guidance says the review covers titles, meta descriptions, structured data and image alt text as well as the body, so paste those in too. It is the same discipline behind an AI fact checker that reads claims against saved sources instead of trusting the writer.

Check this draft before it is published.

1. List every factual claim: numbers, dates, prices, product
   names and features, rankings, quotes, "studies show" lines.
2. For each claim, quote the sentence in SOURCES that supports
   it. If none does, write NONE. Do not guess.
3. Separately, list any sentence that only a model could have
   written: claims of personal experience, invented examples
   presented as real, or experts who are not named in SOURCES.

Output: claim | supporting quote | keep / soften / cut

DRAFT:
<paste the draft>

SOURCES:
<paste the pages you opened, each with its URL>

Step 3 in that prompt matters more than it looks. Models will happily write “in our testing” or “when I set this up for a client” about tests nobody ran. A sentence like that is worse than a missing source, because it fakes the experience signal Google asks for. Cut it, or replace it with something you actually did. That swap, from invented experience to real experience, is most of the gap between an AI draft that ranks and one that does not.

6. AI content in AI Overviews, ChatGPT and Perplexity.

The same question now has a second scoreboard: does AI content get cited in AI answers? For Google's own AI features the answer is short. Its guide to AI features says the best practices for SEO remain relevant for AI Overviews and AI Mode, and that there are no additional requirements or special optimizations to appear in them. A page that is thin or wrong for blue links is thin or wrong for an AI Overview, and our guide to ranking in AI Overviews covers what helps beyond that.

Answer engines outside Google show the same pattern in the Graphite data. Of the articles cited by ChatGPT with web search, 82% were classed as human-written, and the same 82% held for articles cited by Perplexity. The engines are not checking for AI. They pick pages that answer the question with something specific, and unedited AI drafts rarely have anything specific to offer. If you want the mechanics of how those engines choose sources, start with how to rank in ChatGPT and the wider LLM SEO approach.

One thing that does not matter: whether a crawler can tell your page was AI-assisted. GPTBot and ClaudeBot collect pages for model training, while OpenAI's OAI-SearchBot and Anthropic's Claude-SearchBot build the indexes that ChatGPT search and Claude search answers draw on. OpenAI's documentation says each setting is independent, so blocking GPTBot does not take you out of ChatGPT search answers. What gets you cited is the same thing that gets you ranked. The ClaudeBot guide walks through the robots.txt rules for one operator's crawlers.

A worked example: forty pages or eight.

Tallow is a fictional invoicing app at tallow.example, and every page, claim and decision below is invented to show the shape of the work. Tallow wants traffic from “invoice template for <job>” searches. Plan A is the one an AI makes easy: one prompt, a job name swapped in forty times, published in an afternoon. Here it is run through the gate.

TALLOW (FICTIONAL) · PLAN A · "INVOICE TEMPLATE FOR <JOB>"
pages              40  (plumber, electrician, photographer, tutor, ...)
prompt             1, with the job name swapped per page
unique per page    the job name, two sentences of intro
claims checked     0
time to publish    one afternoon

gate result
  purpose          fail  (pages exist to catch 40 query variants)
  added value      fail  (same template, same advice, 40 times)
  claims           fail  (tax and late-fee lines from model memory)
  pace             fail  (40 pages, nobody read them)

Plan A fails on every line, and it is a textbook match for the first example in the scaled content abuse policy. It would not be penalized for using AI. It would be pushed down for being forty copies of one page. Plan B keeps the topic and the AI drafting, and changes what goes in and what comes out.

TALLOW (FICTIONAL) · PLAN B · SAME TOPIC, EIGHT PAGES
pages              8   (jobs whose invoices actually differ)
unique per page    a sample invoice built in the product, the
                   line items that job bills for, two questions
                   pulled from that job's support tickets
dropped            32  (same invoice, nothing to add; the 8 pages
                   link to one general template page instead)
claims ledger      61 claims listed
                     kept 44 · softened 9 · cut 7 · fixed 1
time to publish    four working days

Eight jobs had invoices that genuinely differ (a photographer bills usage licences, a tutor bills packages of sessions), so eight pages got a sample invoice built in the product and the questions those customers actually ask support. The other 32 jobs got nothing new, so they got no page. Then each draft went through the claims ledger in a fresh chat.

CLAIMS LEDGER (excerpt) · "INVOICE TEMPLATE FOR PHOTOGRAPHERS"
claim                                        source              verdict
"most photographers charge a 50% deposit"    NONE                cut
"usage licences belong on the invoice"       2 of 3 guides read  soften
"Tallow adds late fees automatically"        product docs: no    fix
"invoices must show the client's address"    NONE (varies)       cut
"Tallow exports invoices as PDF"             product docs        keep

Look at what the ledger caught. Two confident lines with no source at all, one rule that varies by country stated as universal, and one claim about Tallow's own product that was simply wrong. That last kind is the most common and the most damaging, because it is the line an AI engine will repeat to a buyer. Plan B took four working days instead of one afternoon, and it is the version that matches what Google's guidance asks for. That trade, time for safety, is the real cost of AI content, and it is the same trade behind every sensible way to use AI for SEO.

Common mistakes with AI content and SEO.

1. Writing to beat an AI detector. Detector scores feel like a measure of risk, so people rewrite until the score drops. Google's published guidance on AI content never mentions a detector score; it asks about value, accuracy and purpose. A page rewritten to fool a detector is still thin. Spend the time adding something real instead.
2. Publishing everything the AI can produce. When a page costs nothing to draft, every query variant looks worth a page. That is the exact pattern the scaled content abuse policy names. Publish what you can check and merge the rest into fewer, better pages.
3. Inventing an author. A fake expert seems like cheap E-E-A-T. Google calls fabricated creator profiles a form of deception. Use a real byline, and describe the process where readers would want to know it.
4. Checking the body and skipping the rest. Titles, meta descriptions, structured data and alt text feel too small to check, and they are what shows up in results. Google's review guidance covers them by name.
5. Deleting all AI content after a traffic drop. Panic says the AI was the problem. Check Search Console first: which pages fell, and on which queries. Prune and merge the thin ones, fix the wrong ones, and keep the AI-assisted pages that still earn clicks. The blog SEO guide covers the basics a recovery plan needs.
6. Treating disclosure as the fix. A “written with AI” note is useful where readers would wonder how the page was made. It does not make a thin page helpful or a wrong fact right.