Content · Field Guide
Google does not penalise AI-written content. It penalises what AI makes trivially easy to produce — and in March 2026 it enforced that distinction hard enough to remove some sites entirely.
The question “does Google penalise AI content” has a clear answer that almost nobody finds satisfying, because it’s a technicality that turns out to be the whole thing. No, it doesn’t. Google’s position, consistent since early 2023, is that appropriate use of AI is not against its guidelines and that quality and originality are what matter regardless of how content was produced.
What Google penalises is scaled content abuse: producing many pages primarily to manipulate search rankings rather than to help users. That policy, introduced in the March 2024 spam update alongside site reputation abuse and expired domain abuse and taking effect on 5 May 2024, is deliberately method-agnostic. Human-written spam and AI-generated spam face identical treatment.
The reason AI-generated sites dominate the penalty lists isn’t that Google detects AI. It’s that AI removed the cost constraint that used to limit how much thin content anyone could produce. Publishing five hundred worthless pages used to require a budget. Now it requires an afternoon.
What Happened in March 2026 Enforcement, not new policy
Google released a spam update on 24 March 2026 that completed in a little over nineteen hours — unusually fast — applying globally across all languages. It didn’t introduce new rules. It enforced the existing scaled content abuse policy considerably more aggressively than before.
Reported patterns among affected sites cluster tightly:
Mass generation. Sites publishing dozens to hundreds of articles daily with identical structure, no editorial review, and no original contribution. Many had gone from zero to thousands of pages within weeks — a growth curve that is itself a detectable signal.
Template-with-variable-substitution. The classic programmatic play: “Best [service] in [city]” across hundreds of locations where only the place name and a few variables change. This pattern predates AI by a decade and AI made it cheap enough to do at ten times the previous scale.
Reverse-engineered AI-citation content. Pages structured specifically to match what AI answer systems appeared to favour, produced at volume. Google has explicitly brought this under scaled content abuse.
Traffic declines reported among hit sites ran to 50–80%. And a further development landed on 15 May 2026: Google confirmed that its spam policies apply across all of Search including AI Overviews and AI Mode. Content violating spam policy is ineligible to appear in AI-generated summaries. For anyone who had treated the generative layer as a policy-free zone worth gaming separately, that assumption is now explicitly removed.
Where AI Genuinely Helps Uses that survive scrutiny
Having established what gets punished, here’s what actually works. These are the uses I’d defend without hesitation.
Research synthesis and structure. Google Search Central’s own generative AI guidance acknowledges AI as useful for researching topics and adding structure. Compressing six sources into an outline, identifying what a comprehensive treatment would need to cover, spotting the angle nobody has taken — this accelerates the thinking without substituting for it.
First drafts of things you already know. The highest-value use by a distance. If you possess genuine expertise and the constraint is transcription speed rather than knowledge, drafting from your own detailed notes and then rewriting produces good work fast. The output contains your knowledge because you supplied it.
Editing and tightening. Cutting a rambling draft, flagging unsupported claims, identifying where the argument skips a step. Models are notably better as editors than as authors, and this use is almost entirely upside.
Structured, verifiable transformation. Turning a spreadsheet of specifications into readable prose, formatting a set of data into a comparison table, generating schema markup. The facts come from your data and the model handles presentation.
Titles, meta descriptions and internal linking suggestions. Genuine time-savers with limited downside, provided you check the output against the actual page.
Translation and localisation drafts, reviewed by someone who speaks the language. The draft saves substantial time; the review is not optional.
Where It Actively Hurts Named failure modes
Publishing without domain review. The core failure. Models produce fluent, confident, plausible text containing errors that only someone who knows the subject will catch. In technical, medical, legal or financial content this is the difference between a useful page and a liability.
Volume as a strategy. If your plan is measured in articles per day rather than in what each article contains, you have described the thing the policy targets. The March 2026 enforcement was aimed precisely at this.
Content with nothing original in it. A model trained on the existing web produces, by default, a competent synthesis of what already ranks. That is definitionally not additive. The page has no reason to outrank its own sources, and Google’s stated intent for the 2024 update was reducing exactly this class of unoriginal content in results.
Fabricated specifics. Invented statistics, misattributed quotes, studies that don’t exist, prices that were never accurate. This is the single most damaging output failure because it survives casual editing — a fabricated figure reads exactly like a real one.
Uniform structure at scale. Pages sharing identical skeletons across a large set are a detectable pattern independent of the quality of any individual page.
“A model trained on what already ranks will, by default, produce a competent summary of what already ranks. That page has no argument for outranking its own sources.”
The Two Other Policies That Catch People Introduced alongside scaled content abuse
The March 2024 update introduced three policies, not one, and the other two catch a different population.
Site reputation abuse — often called parasite SEO — covers publishing third-party content on a trusted domain with little or no first-party oversight, in order to exploit that domain’s ranking signals. The classic form is a coupon section, a sponsored directory or a partner content area bolted onto a reputable publisher and detached from the site’s actual purpose.
Google’s May 2026 clarification made this explicitly applicable to AI Overviews and AI Mode as well. The exposure is broader than most large sites realise: if you host review aggregators, lead-generation content or partner-supplied pages on your own domain without genuine editorial involvement, that’s the described pattern regardless of whether anyone intended it as a tactic. Large, high-authority domains carry the most risk here precisely because their authority is what makes the arrangement commercially attractive to the third party.
Expired domain abuse covers buying a lapsed domain primarily to exploit whatever ranking signals its history carries. The line between this and a legitimate acquisition is intent and continuity: acquiring a domain because you’re continuing or absorbing what it did is normal business, and acquiring it to put unrelated content on top of its historical authority is the violation.
Worth knowing if you’re on the buying side of the expired domain market: the same characteristics that make a domain attractive to a link seller are the ones this policy targets.
How Detection Appears to Work Pattern, not prose analysis
A widespread assumption is that Google runs a classifier over text deciding whether a machine wrote it. The observable evidence points somewhere different, and understanding this changes what you’d do about it.
The signals that appear to matter operate at site level rather than page level. Publishing velocity and its shape — a domain going from nothing to thousands of pages in weeks is a pattern no organic editorial operation produces. Structural uniformity across a page set. The ratio of pages to demonstrable expertise or authorship. Engagement behaviour once users arrive. Whether the content contains anything not already present in the sources it summarises.
None of those require identifying the writing tool. All of them describe scaled thin content whether a person or a model produced it, which is exactly what the method-agnostic framing of the policy says.
The practical implication: efforts to make AI text “undetectable” are aimed at a mechanism that doesn’t appear to be doing the work. A site publishing four hundred structurally identical pages a month is describable from its sitemap alone.
AI Detection Tools: Don’t Use Them A brief, firm digression
A market exists for tools claiming to identify AI-written text. I’d advise against building any process on them, for two reasons.
First, they’re unreliable in both directions. False positives on human writing are common — particularly on non-native-speaker English and on formal or technical prose — and false negatives on lightly-edited model output are trivially achieved. Building an editorial policy on a classifier with that error profile means rejecting good human work and accepting bad machine work.
Second, and more fundamentally, the question is wrong. Google is not asking whether a page was AI-written; it’s assessing whether the page is useful and original. A detector that answers the first question tells you nothing about the second. A team that passes a detection check and publishes something worthless has optimised for a proxy nobody is measuring.
Review for accuracy, originality and usefulness. Those are the criteria that apply, and they require a person who knows the subject rather than a classifier.
The E-E-A-T Layer What actually signals a real operation
Google’s quality framework — Experience, Expertise, Authoritativeness, Trustworthiness — is often treated as vague advice. In the context of scaled content enforcement it becomes concrete, because it describes the signals that distinguish a genuine publication from a content farm.
A named author who demonstrably exists. Not a stock photograph and an invented biography. A person with a professional footprint elsewhere, whose stated expertise is verifiable and relevant to what they wrote. Sites hit hardest in enforcement waves are consistently those where authorship is decorative or absent.
First-hand experience, stated explicitly. The E that was added to the framework is the one AI cannot supply. “I ran this for eight months and here’s what broke” is not producible by a model summarising the web, and it’s the clearest available differentiator.
Something checkable that’s yours. Original data, a test you ran, a screenshot of a real account, pricing you verified this week. A single verifiable original element does more than a thousand additional words.
Visible accountability. Real contact details, a genuine organisation behind the site, corrections when you get something wrong. Trust is described as the most important component of the framework, and it is largely conveyed by structural signals rather than by prose.
Editorial coherence. A site about one thing, expanding sensibly, rather than a domain covering finance, pets and travel because those keywords had volume.
None of this prohibits using AI. It describes the wrapper that makes AI-assisted content defensible — and notably, it’s the same wrapper that made human-written content credible before any of this arose.
A Workflow That Holds the Line
| Stage | AI role | Human role | Risk if skipped |
|---|---|---|---|
| Topic selection | Suggest angles, find gaps | Decide, based on audience knowledge | Publishing what exists already |
| Research | Synthesise sources, build outline | Verify every source is real and says what’s claimed | Fabricated citations |
| Original input | None | Supply experience, data, opinion, testing | The page has no reason to exist |
| Drafting | Draft from your notes and structure | Provide the substance being drafted from | Generic synthesis of the top ten |
| Fact checking | Flag unsupported claims | Verify every number, date, name, price | Confident errors published at scale |
| Editing | Tighten, flag weak arguments | Rewrite in a voice that is yours | Uniform, recognisably templated prose |
| Publishing | Metadata, schema, internal link ideas | Final read; accountability for what’s said | Everything above becomes theoretical |
The row that matters most is the third one. If you cannot state what this page contains that isn’t already in the search results — original testing, proprietary data, professional experience, a genuine argument, a synthesis nobody has made — then no amount of editing rescues it. That question is the whole quality bar, and it’s answerable before you write a word.
Auditing Content You Already Published Before an update finds it
If you have published AI-assisted content at any volume, the useful exercise is finding your weakest pages before an enforcement wave does.
Pull every URL and cross-reference against Search Console impressions over six months. Pages with meaningful crawl cost and negligible impressions are your candidate set — Google has seen them, indexed them, and decided they’re not worth showing.
Then apply one question to each: what does this contain that isn’t in the pages currently ranking for its target query? Read it as a sceptical stranger. If the answer is nothing, you have three options and no fourth. Improve it substantially by adding something original. Merge it into a stronger page on the same topic. Or remove it.
Removal is the one people resist, and it’s frequently correct. A site’s quality assessment appears to be affected by its aggregate character rather than only by its best work, which means a large body of unremarkable pages can suppress the pages you actually care about. Pruning is not lost effort — it’s recovering the ranking potential of everything that remains.
Do this proactively rather than after a drop. Recovering from an algorithmic reassessment takes months and depends on subsequent updates. Avoiding one takes an afternoon with a spreadsheet.
If You’ve Already Been Hit A recovery path
First, establish which kind of decline you’re experiencing. A manual action appears in Search Console as an explicit notification following human review. An algorithmic decline has no notification — rankings simply fall as systems reassess the site. These require different responses and people routinely conflate them.
For either, the work is similar:
Identify the thin page sets. Usually obvious: templated groups, pages with near-identical structure, anything published in bulk. Cross-reference your crawl against Search Console impressions and find the pages with crawl cost and no traffic.
Consolidate or remove. Merge overlapping thin pages into fewer substantial ones. Deindex or delete what can’t be rescued. Removing content feels wrong and is frequently the correct move — a site’s quality assessment is affected by its weakest sections, not only its best.
Enrich what survives. Add what was missing: original data, real examples, actual expertise, an author who exists and is identifiable.
Fix internal linking after the consolidation, so the surviving pages inherit the structure rather than pointing at deleted URLs.
Submit a reconsideration request — but only for manual actions. There is no reconsideration process for algorithmic declines; recovery there comes with subsequent updates once the underlying assessment changes, which takes months rather than weeks.
Set expectations honestly: recovery from scaled content enforcement is slow, and for sites whose entire model was volume, there may not be a recovery so much as a rebuild.
Policy positions verified against Google Search Central spam policies and generative AI guidance; update dates and enforcement patterns from published reporting on the March 2024 spam policy introduction, the March 2026 spam update, and Google’s 15 May 2026 confirmation regarding AI Overviews and AI Mode. Traffic-impact figures are reported observations rather than Google-published data. This article contains no affiliate links.