Does AI content get penalized? Straight answers, with sources.

No. Google does not penalize content for being AI-generated. Its policy has been explicit since February 2023: it rewards high-quality content "however it is produced," and appropriate use of AI or automation is not against its guidelines. What Google penalizes is scaled content abuse — many pages produced primarily to manipulate rankings rather than help users, "no matter how it is created." The distinction is value and originality, not the tool used to write the draft.

The two lines worth quoting directly: "Rewarding high-quality content, however it is produced," and "Appropriate use of AI or automation is not against our guidelines." The same post draws the line — using automation "with the primary purpose of manipulating ranking in search results is a violation of our spam policies" 1.

Google's spam policy names the offence precisely, and its first listed example is using generative AI "to generate many pages without adding value for users" 2. When Google tightened the rule in March 2024 it said the policy applies "whether automation or humans are involved" 3 — the same standard a human content farm has always been held to.

The test Google applies is not "was this written by AI" but "does this page add something the searcher cannot get elsewhere."

The measurement data points the same way. Ahrefs checked roughly 600,000 top-ranking pages and found the correlation between a page's AI share and its Google position was 0.011 — effectively zero — while 86.5% of top-ranking pages contained some AI content 5. A later Ahrefs study of about 331,000 pages found 5.3% of top-three results were fully AI-written, with no hard cutoff in the data that would indicate a binary AI classifier 6.

Google's generative-AI guidance adds the one condition that actually matters in practice: "It is critical to manually fact-check and review all AI-generated content for accuracy and trustworthiness before publishing" 4. That is a quality instruction, not a prohibition.

Heavily AI-written pages do underperform on average, and we do not dispute that. Pages with low or moderate AI content earn two to three times the impressions of pages with high AI content, and 86% of articles ranking in Google are human-written. The most likely explanation is that mass-produced pages tend to be thinner and less original, not that an AI detector is down-ranking them. The pattern tracks quality, and quality is what we control.

Three separate datasets point the same way. Ahrefs measured impressions falling two to three times for high-AI pages, with indexation dropping from about 49% to 40% at the very high end 6. Graphite found 82% of ChatGPT and Perplexity citations go to human-written articles 7. A Semrush study reported by Search Engine Land found human-written pages roughly eight times more likely to hold position one 8.

The same studies describe a gradient, not a switch. Ahrefs: "no obvious hard cutoffs suggesting a binary AI classifier" 6. Graphite explicitly did not test heavily edited AI-assisted content and notes it "may be an effective strategy" 7. Search Engine Land cautions that the studies rely on AI detectors, which are inconsistent 8.

So the defensible reading is: raw, unedited, volume-first AI output underperforms because it is usually lower quality. That is a statement about production standards, not about the model that typed the first draft. Our position is that the fix is process — original inputs, real expertise, human review, one page worth citing — not pretending the tool is the problem.

Which is why the checks matter more than the writing. Every page we publish is written from the client's own brand kit and approved documents, and then has to clear a compliance review against their product truths, a fact-check that deletes any number, date or quote it cannot trace to an exact source sentence, and a deduplication gate against pages the site already has. A page that loses those checks does not publish. That is a narrower claim than "our AI writes better," and it is the one we can actually prove.

No AI vendor publishes a policy that blocks or devalues AI-written content. Access to answer engines is governed by crawler permissions, not content origin: OpenAI documents that sites which block OAI-SearchBot will not appear in ChatGPT search answers, and Perplexity documents PerplexityBot as the crawler that surfaces and links sites. The real risk is indirect — if a Google penalty removes a site from the index, its AI Overviews and ChatGPT citations usually go with it.

Only two permissions actually decide this. OAI-SearchBot must be allowed for a site to appear in ChatGPT search answers 12, and PerplexityBot is the crawler Perplexity documents as the one that surfaces and links sites 13. Neither page says anything about how the content was produced.

The risk that is real is spillover. In a 2026 case study, a site with roughly 850,000 AI-generated local-news URLs took a manual action, then lost AI Overviews and AI Mode visibility, and its ChatGPT citations dropped heavily once it was out of Google's index 14. Google eligibility is the gateway to most AI citation, so Google quality hygiene is answer-engine hygiene.

Practically: make sure GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot and Google-Extended are not blocked at the firewall, CDN, or age gate. We check this during onboarding because a blocked crawler is a silent zero.

AI detectors are not reliable enough to be a ranking mechanism, and there is no evidence Google uses one. OpenAI withdrew its own classifier in July 2023 citing its low accuracy — it caught 26% of AI-written text while flagging human text as AI 9% of the time. Peer-reviewed testing found detectors misclassified non-native English essays as AI more than 60% of the time, and that the tools are neither accurate nor reliable.

OpenAI retired its classifier five months after launch, "due to its low rate of accuracy" 9. Independent testing is worse. Liang et al. ran seven detectors over non-native English TOEFL essays and found an average 61.22% misclassified as AI, with at least one detector flagging 97.8% of them 10. Weber-Wulff et al. tested fourteen tools including commercial ones and concluded they "are neither accurate nor reliable" 11.

Google has never stated that it scores pages for AI authorship. Its public position is that its spam systems look at whether content is helpful and whether it was produced to manipulate rankings 12. A detector that is wrong about a third of the time is not a basis for a penalty system, and Google has not built one.

There is no published cap, but there is a real ceiling set by site authority. Publishing far more than a domain can support produces pages Google will crawl but not index, which is the mechanism behind most "our content disappeared" stories. We cap output at 30 to 90 articles a month depending on domain rating and existing indexation, and we would rather publish 30 pages that earn impressions than 300 that do not.

Google does not publish a volume limit. What it does police is mass production of pages that do not help users 2, and its March 2024 update folded that into scaled content abuse regardless of whether a human or a tool did the producing 3.

The practical ceiling is authority. A domain with a low domain rating cannot get hundreds of new URLs crawled, indexed, and ranked quickly, so the extra pages sit as crawl debt and dilute internal linking. That is a crawl-budget and site-quality constraint, not a penalty.

Our planning range is 30 to 90 articles a month, set per client from domain rating, existing indexed page count, and how quickly previously published pages are picking up impressions. If a client pushes for more, the honest answer is usually no — the constraint is the site, not the writing capacity.

Yes, drafts are AI-produced, and we say so. What makes them different is what goes in and what is checked before publication. Each page is written from the client's own kit — brand voice, author persona, product truths, banned words and forbidden claims — plus approved excerpts from their own documents, a live Google SERP snapshot, and a scrape of the top three ranking competitor pages. Then it passes an automated compliance review against the client's product truths, a fact-check that requires an exact quoted source sentence for every number, date and quote, and a sealed-artifact QC gate that can block publication.

Drafts are AI-produced. We do not claim otherwise, and buyers who read a lot of content can usually tell, which is why pretending is a bad strategy.

The inputs are where the difference sits. Every article is written from the client's Brand Kit: brand voice and brand context, the author persona and writing examples, product truths from their offer capsules, and the rule list — banned words, banned claims, and exact forbidden claims. On top of that goes their Knowledge Library: PDFs and scanned PDFs, Word documents, text and Markdown, and YouTube transcripts, chunked and retrieved per article. Only documents the client has explicitly approved for public article use are eligible, and if a source is later revoked or changed, any unpublished article built on it is held.

The research is live, not remembered. Each article starts from a Google results snapshot of the top ten, a scrape of the top three ranking competitor pages, and customer language pulled from Reddit, review sites and forums. Research is written into a brief that has to reference the actual search results before drafting begins.

Then the checks, which are the part worth asking about. Banned words are enforced by code rather than requested in a prompt: they are stripped or blocked in the title, meta description, body, excerpt, FAQs and image alt text, including translations. Explicitly forbidden claims block publication outright. Every number, date, quote and named-source claim in the body has to trace to an exact quoted sentence from a source — if it cannot, the sentence is deleted rather than softened. And the client decides the last step: auto-publish, draft into their CMS, or review-and-approve, where approval is pinned to that exact version and any edit resets it.

Google's generative-AI guidance sets the bar we build to: "It is critical to manually fact-check and review all AI-generated content for accuracy and trustworthiness before publishing" 4. Where a client wants a human in the loop on every page, review-and-approve mode exists for exactly that. Publication can also stay in draft mode indefinitely while a client builds confidence in the voice.

Every article runs a fixed pipeline whose last four stages are finalize, compliance review, QC and publish. Along the way it is checked for brand voice and banned terms, forbidden claims, unsupported numbers and quotes, near-duplicates of pages already on the site, a minimum of three internal links, outbound links that must return HTTP 200 and are never competitors, broken HTML or invalid structured data, and a stale year in the title. The finished file is then hash-sealed, and publication requires passing compliance and QC records against those exact bytes. The database blocks edits after that.

The pipeline is fixed rather than bespoke per article, which is what makes the guarantees checkable. A draft moves through research, outline, writing, links, external citations, HTML assembly, finalization, compliance review, QC and publish.

Checks that can stop a page:

• A near-duplicate of a page the site already published. This gate fails closed — if the comparison cannot complete, the article is held rather than published.
• Fewer than three internal links to the client's own verified pages.
• An outbound citation that does not return HTTP 200, points at a competitor, or no longer supports the sentence it is attached to. Citations are re-fetched immediately before publishing.
• A leftover banned term or an exact forbidden claim.
• Broken HTML, invalid structured data, model refusal text leaking into the copy, or a page that would render under 150 words.
• A stale year in a freshness title.

Length is set by tier — roughly 1,200, 1,400 or 2,400 words — with a floor at 65% of target. A short draft gets one repair pass. We say that plainly because "no article ever ships short" would be a claim the system does not make.

The final step is what we would point a sceptical buyer at. The article file is hashed, and publishing requires a passing compliance attestation and a passing QC attestation recorded against that same hash. A database constraint then prevents the bytes from being edited. So the thing that got checked is the thing that went live — not a version of it.

What we will not claim: that a person reads every article before it publishes. By default the system publishes to the client's CMS, and human review happens where the client has chosen review-and-approve, where a hold fires, or in their own drafts. The compliance reviewer and the fact-checker are automated. If a client wants per-article human sign-off, that is a setting, and it is worth being explicit about which mode you are buying.

It gets fixed, and it is expected rather than exceptional. Once a site publishes at volume, pages start competing for the same query, which is keyword cannibalization. It is a normal consequence of publishing a lot of content and has nothing to do with AI. We run cannibalization checks, consolidate or re-point the losers, and rework pages that pick up impressions but no clicks.

Cannibalization is the predictable side effect of a large content footprint: two pages target overlapping intent, Google picks one, and the other underperforms or splits signals. It happens to every site that publishes at scale, human-written or not, and it is a structure problem with a structure fix.

The fix is ordinary SEO work: identify the competing set, decide the canonical page per intent, consolidate or differentiate, re-point internal links, and re-check positions. Pages that rank but never get clicked usually have a title and meta problem, or they answer a question the searcher did not ask.

What we do not do is publish more volume to bury the problem. Volume is what created it.

Expect indexation and rising impressions in months one and two, measurable AEO traffic and prompt wins between months two and four, and compounding results from month five, with meaningful ROI usually around 180 days. AI citations lag publishing by weeks and move in steps rather than in weekly lines, so a monthly view is the honest way to read it.

New pages need to be crawled, indexed, and then scored against incumbents. That is measured in weeks, not days, and AI citations lag published content further still because the retrieval layer has to re-crawl and re-rank the source before it can quote it.

A realistic sequence: months one and two for indexation, rising impressions, and the first AI mentions; months two to four for measurable traffic and prompt wins; month five onward for compounding, with ROI typically around the 180-day mark. Reporting monthly is more honest than weekly because weekly AI-visibility movement is mostly noise.

The comparison that matters is against the client's own baseline, not against a competitor's absolute traffic. We agree the Search Console and analytics baseline before the first page ships.

If tracking AI visibility is the goal, yes. Citation share is a diagnostic, not an outcome. What pays is being the recommended answer for the questions that precede the deals you actually want, mapped backwards from those deals. A citation dashboard that climbs while calls stay flat is a vanity metric — and a 40% lift in AI citation rate can be entirely people who were already typing your brand name into ChatGPT. Brand demand, more calls, more revenue and more direct traffic are what a strategically built AEO programme produces.

The distinction that decides whether the work is worth doing is demand capture versus demand creation. Demand-capture visibility is showing up when a buyer already knows your brand or category and is comparing options — "is [company] worth it", "[company] vs [competitor]", "best [category] for [type of company]". It shortens sales cycles and lifts win rates on deals you were probably already in. Demand-creation visibility is showing up while a buyer is still working out the problem, before any vendor name is in play — "why is my organic traffic dropping", "how do I track brand mentions across LLMs". That creates pipeline that would not otherwise exist 17.

Both matter, they are not interchangeable, and they are not equally easy. Most teams buy the cheaper one — demand capture — then report the resulting visibility lift as though it covered both. Kaleigh Moore puts the trap plainly: "A 40% lift in AI citation rate sounds great until you realize it might all be coming from people who were already typing your company name into ChatGPT" 17.

That is why we plan backwards. Start from the business outcome — booked calls, revenue, branded and direct traffic — and map the clusters of buyer questions that precede those specific deals. Then measure whether the brand is the recommended answer inside those clusters, not whether a global share-of-voice number went up. A visibility programme that cannot name the revenue it is aimed at is a dashboard, not a strategy.

The external data supports the value of the channel while keeping it in proportion. AI platforms grew from 0.1% to 6.9% of sessions across 218 sites and 3.7 billion sessions, while organic search fell 24.8% 18. In B2B, AI search is still only 2.4% of search sessions — but accounts that arrived through an AI engine closed at 2.5x the rate of accounts from organic search, and AI's share of closed-won deals tripled in a year from 1.1% to 3.4% 19. Visitors from ChatGPT converted to leads at 4.7% against 1.9% for Google organic 18.

The caveats are real and we state them. Volume is small, so total conversions from organic still dwarf AI referrals. A 95-site benchmark found AI referrals converting 1.18x better than organic on average, but worse than organic on 23 of 63 sites, and at 0.62x in B2B product businesses specifically 20. AI traffic quality depends on the category and on whether the cited page actually answers a buying question.

On our own business, running AEO Engine on aeoengine.ai: 30% of booked calls now arrive through AI search. We assign $500 of value to a booked call, which puts that channel above $16,000 in the last month alone. That is the number we hold ourselves to — not a citation score.

The question is never "how do we rank in ChatGPT." It is "which conversations precede the deals we want, and are we the recommendation in those?"