Your Traffic Dropped. It’s Probably Not a Penalty.

Google’s scaled content abuse policy doesn’t grade your writing. It reads the shape of your publishing. Here’s where the line sits and how to diagnose a drop.

The most popular defense of AI content is also the first one to fail. It goes: my content is good, so I’m fine.

I understand why people believe it. Google spent a decade telling us to write helpful things, and quality slowly became the answer to every question anyone asked. So when the scaled content abuse policy landed, people reached for the tool they had.

Then you read the policy and the word quality is barely in it. Google defines scaled content abuse as generating many pages for the primary purpose of manipulating search rankings rather than helping users. Then it adds five words that end the argument before it starts: no matter how it’s created.

Your defense is answering a question nobody asked. You’re arguing about the writing. Google’s looking at the assembly line.

What counts as scaled content abuse?

Scaled content abuse is Google’s spam policy covering pages produced in bulk primarily to move rankings rather than help people. Google’s own wording is that the practice centers on large amounts of unoriginal content with little value to users, no matter how it’s created. The policy is method-agnostic. It applies to AI output, human writing, scraped feeds, and automated translation equally. Volume plus manipulative intent is the violation. Authorship is not.

Google’s published examples make the shape clear. Using generative tools to produce many pages without adding value. Scraping feeds or search results to spin up pages, including through synonymizing or translating. Stitching together content from other sites without adding anything of your own. Notice what’s missing from that list. Nothing about grammar. Nothing about depth. Nothing about whether a human touched it.

The policy isn’t new either, which is the part people miss. It’s a rename. Before March 2024, this same rule was called spammy auto-generated content, and Google has been acting against machine-spun pages since long before anyone could prompt a model. AI didn’t create this category. It just made the category cheap enough that everyone could enter it at once.

The policy grades the process, not the prose

Here’s the test that clears up most of the confusion.

Take two sites. Both publish 400 pages in a month. Both read well. On the first, a person picked each topic, checked what was already ranking, wrote or edited each page, gave it a title they thought about, and hit publish 400 separate times because that’s what the business needed.

On the second, someone fed a keyword list into a pipeline and 400 pages appeared.

Same word count. Similar reading experience. Completely different risk, because the policy is written around intent and mechanism, and mechanism leaves marks the second site can’t scrub.

This is why “but the content is genuinely good” doesn’t work as a defense. It’s a quality argument in a process case. If a page could have been produced by feeding a list into a machine and letting it run, the writing quality doesn’t retroactively change how it got made.

That’s also the good news, and it’s the part that gets lost in the panic. Google said in February 2023 that its focus is on rewarding high-quality content however it’s produced, and it has not walked that back through any update since. Using AI to structure a piece, to draft a section, to catch the three points you forgot, to speed up research you then verify: none of that is the violation. I use AI daily, and I’d say the same thing if I didn’t. The violation is scale replacing judgment.

Detection looks more like network monitoring than a reading test

Before I moved into SEO, I spent years as a network engineer. That background changed how I read this policy, and it’s the analogy I keep coming back to when people ask me how Google could possibly know.

Intrusion detection on a network doesn’t work by reading the contents of your packets and deciding whether they’re written well. It works on shape. Volume over time. Timing intervals. Repeated structures in headers. Traffic from one source behaving in a way that individual traffic doesn’t. A single packet tells you almost nothing. Ten thousand packets with a shared signature tell you almost everything, and they tell you fast.

Publishing at scale generates the same kind of signature, and it doesn’t live in your paragraphs. It lives in the metadata around them. Titles built from the same formula with one variable swapped. URLs constructed by a script. Publishing velocity that jumps from four pages a month to four hundred. Internal links that all point the same direction because a template put them there. Pages that appeared in one batch and were never touched again.

To be clear about what’s verified and what isn’t: Google publishes the policy, not the heuristics. The signature list above is my inference from the policy language and from what enforcement visibly hits, not a leaked spec. But the logic holds regardless of the specifics, because the alternative is worse. Google can’t afford to evaluate quality page by page across a web where anyone can generate ten thousand articles in an afternoon. Reading everything doesn’t scale. Pattern detection does. It’s the only approach that survives the volume.

Which leads to the second thing engineers learn about signature detection, and it’s the uncomfortable one. It’s blunt. It catches things that share the signature but not the intent. Any system tuned to catch a pattern at scale will take some collateral, and the collateral is usually the person who did nothing wrong but happened to look like the person who did.

The crackdown people spent two years predicting already happened

If you’ve been waiting to see whether Google would do something serious about AI at scale, stop waiting. The answer arrived, and it arrived quietly.

On March 24, 2026, Google launched a spam update. It finished the next morning, in roughly 19 and a half hours. That’s the fastest confirmed spam update in the history of Google’s Search Status Dashboard. For comparison, the August 2025 spam update took 26 days and the March 2024 spam update took 14. Two days later, Google started a broad core update that ran until April 8. Then another spam update in late June, done in two days.

Google announced no new spam policies alongside any of it. That’s the detail worth sitting with. No new rules means the existing rules got enforced harder and faster. The rulebook didn’t change. The enforcement speed did.

And this is the second round, not the first. The March 2024 push was the loud one. Google projected it would cut low-quality, unoriginal content in results by 40 percent, and when the rollout finished in April 2024, the company reported the actual figure came in nearer 45 percent. That was the warning shot. What’s happened since is Google getting better at doing the same thing in less time.

There’s a psychology point buried in here that’s worth naming, because it explains why so many people kept scaling anyway. Survivorship bias. The operator still ranking with 20,000 generated pages is visible. He posts screenshots. The ninety operators who got flattened aren’t posting anything, because there’s nothing to post. So the strategy looks far more reliable than it is, and every month it keeps working is read as proof rather than as a countdown.

Second-order thinking is the antidote, and it’s not complicated. Twenty thousand pages looks incredible for two weeks. Ask what it looks like in month eight, and whether the domain you built it on is the same domain your business needs to still be standing on.

Most traffic drops in 2026 aren’t penalties

This is where I think practitioners are losing the most money, and it has nothing to do with spam.

Traffic falls. The first explanation that reaches for you is a penalty, because penalty stories are vivid and they’re everywhere. That’s the availability heuristic doing its work: the thing you can picture easily feels like the thing that’s most likely. Penalties are dramatic, so they get the airtime, so they become the default diagnosis for a problem that usually has a duller cause.

Occam’s razor is the better tool here. Check the boring explanations first.

Start with position, because it separates the two cases immediately. If your rankings held and your traffic fell, you didn’t get demoted. Something changed between the ranking and the click. If your rankings fell along with your traffic, you have a ranking problem, and a ranking problem is often fixable in a way that a click problem isn’t.

The most common click problem right now has a name. Pew Research analyzed nearly 69,000 Google searches from 900 US adults and found people clicked a traditional result 8 percent of the time when an AI summary appeared, against 15 percent when it didn’t. Clicks on the sources cited inside the summary ran at 1 percent. Google disputed the methodology, and that objection deserves to sit in the record, though several independent analyses through 2025 pointed the same direction. Whatever the exact figure, position one on a query that now returns an AI answer is not the asset it was two years ago.

That’s not a penalty. There’s no reconsideration request for it. Treating it like one means you’ll spend three months auditing content that was never the problem.

A real spam penalty tends to look different anyway. It’s a cliff, not a slope. It lands inside a window you can point at on a calendar. It usually hits site-wide or hits one clearly identifiable group of pages rather than thinning out everywhere at once. If your chart shows a gentle decline over five months, that’s almost certainly not enforcement.

How to tell a penalty from everything else

Here’s the sequence I’d run, in this order.

1. Check Search Console for a manual action first. It takes thirty seconds, and it’s the only definitive answer available to you. If there’s nothing there, you don’t have a manual penalty, full stop.
2. Compare average position against clicks over the same period. Position steady with clicks down points to SERP features or AI Overviews. Position down with clicks down points to ranking.
3. Line the drop up against confirmed update dates. Google’s Search Status Dashboard publishes start and end timestamps. If your cliff sits inside the March 24 to 25 or June 24 to 26, 2026 windows, that’s a real lead. If it doesn’t line up with anything, stop blaming updates.
4. Segment by page group. Blog versus landing pages, templated versus handwritten, old versus new. Enforcement usually concentrates. A drop spread evenly across every page type is more likely to be demand or seasonality.
5. Pull the queries that lost the most and search them yourself. Look at what’s occupying the space now. Sometimes the answer is visible in ten seconds, and it’s an AI Overview or a new competitor, not anything about you.

Only after those five would I start rewriting anything.

Volume isn’t the violation. Volume without judgment is.

I want to be careful here, because “just don’t publish a lot” is bad advice and it isn’t what the policy says. Plenty of legitimate businesses need hundreds of pages. Programmatic pages built on real inventory, real data, or real differences between locations are not automatically a problem, and Google has never said they are.

Get The Guy From Internet’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

The line sits at whether a human made a decision about each page that a script couldn’t have made.

Take two versions of the same programmatic build to see where it lands. A booking site generating a page per city, where each page carries that city’s real inventory, actual prices, actual availability, and details that differ because the underlying reality differs, is publishing at scale on top of a genuine data set. Every page answers a question no other page on the site can answer. Now take the same city list where the only thing that changes between pages is the city name inside otherwise identical sentences, with no data behind it. Same technique. Same page count. One has something per page that couldn’t be copied from the page before it. The other doesn’t, and no amount of rewriting fixes that, because the missing thing isn’t language.

The practical version of that test is a question I’d ask about any page before it goes live: could I defend this page to a person who asked why it exists? Not defend the site. Defend the page. If the honest answer is “because the keyword existed,” the page is a liability regardless of how well it reads.

What a defensible publishing process looks like

The advice that usually follows an article like this is to write helpful content, which is true and useless. Here’s what I’d actually change if I were auditing a workflow today.

Break the sameness in the metadata, not just the body. Titles built on one repeating formula are the cheapest signature to generate and the cheapest to detect. If every title on your site follows “Best X for Y in Z,” that pattern is more visible than anything in your paragraphs. Write the title for the specific piece, then check whether it reads like the last twenty you published.

Put something in each page that only you could have put there. A number you pulled, a result you observed, a workflow you actually ran, a screenshot of something on your own screen. This is the single highest-value change available, and it does double duty: it’s what separates you from a pattern, and it’s also what AI engines quote when they cite a source. Everything else on your page exists somewhere else on the internet already.

Slow the ramp. A site that publishes six pages a month for a year and then publishes four hundred in March has drawn a line on a graph that anyone can see. If you genuinely need to scale up, scale up gradually and give the growth a shape that a real editorial operation would produce.

Stop treating publish as the finish line. Pages that appeared in one batch and were never touched again read differently, at the pattern level, than pages that get updated because someone is paying attention to them. Going back to revise a page for a real reason is one of the few signals that’s expensive to fake.

Audit what you already shipped before you ship anything else. Pull every URL, sort by clicks over the last six months, and look hard at the bottom. Pages with no clicks, no links, and nothing original on them aren’t neutral. They’re the part of your site that looks most like the thing Google is filtering.

None of that is about hiding your process from detection, and I want to be direct about why. Trying to disguise the signature is the same game as the scaled operators, run with less budget, and it’s a game where the rules change without notice, and your business is the thing on the table. The durable move is to have a process that doesn’t need disguising.

That’s the Lindy argument, and it’s the one I keep landing on. The tactics that work in search have a half-life measured in months. The fundamentals have been the same for twenty years: be a real source of something, publish things people are glad they found, and give the search engine no reason to look twice. That advice was correct before AI, it survived every update since, and it’s the only version I’d bet a client’s domain on.

Sunk cost is what makes this hard. Once you’ve paid for 3,000 pages, deleting 2,000 of them feels like setting money on fire. It isn’t. That money is already spent either way. The only live question is whether those pages help or hurt the domain from here, and thin pages that exist purely to catch queries have been on the wrong side of that question since 2022.

The slow route compounds and the fast route doesn’t. That’s the whole argument, and it’s boring, which is exactly why it keeps working while the clever approaches take turns getting caught.

What I’d flag as still unsettled

One claim circulates in SEO circles that I’m not comfortable stating as fact, and I’d rather flag it than quietly include it.

The claim is that Google runs a click-through rate test on newly published pages, and that failing it signals something is wrong with the page. The honest position is that this is contested. Google has said publicly and repeatedly for over a decade that CTR is not a ranking signal, on the grounds that click data is noisy and easy to manipulate. On the other side, testimony during the Department of Justice antitrust proceedings described a system using click data to adjust rankings, and the 2024 API documentation leak showed click signals inside the ranking pipeline.

So: click data almost certainly exists somewhere in the system as a re-ranking layer. The specific “new pages get a CTR audition” mechanism is not something Google has documented, and I wouldn’t build a strategy on it.

I’m including this because I think flagging the uncertain thing is more useful than presenting a clean story. Anyone who tells you they know exactly how Google’s spam detection works is telling you about their confidence, not about Google.

The part that actually worries me

The people most exposed to this next round aren’t the ones running content farms. They know the risk they took, and they’ve priced it in.

It’s the small operator who published 200 pages over two years because that’s what the business needed, used AI to help with drafts, and happened to build titles from a consistent format because consistency seemed like good practice. Nothing that person did was against any rule. But blunt systems are blunt, and a signature is a signature.

If that’s you, the protection isn’t hiding your process. It’s making sure the thing a pattern-matching system finds when it looks closely is a record of decisions rather than a record of output. Original data. First-hand experience. Pages that visibly could not have been produced by anyone else.

So here’s what I’d like to know from the people reading this. Pull up your Search Console and look at March 24 through 25, then June 24 through 26. Did anything move? And if it did, are you sure you know why?

5. FAQ

Does Google penalize AI-generated content? No. Google’s February 2023 guidance states that its systems reward high-quality content however it’s produced, and appropriate use of AI or automation is not against its guidelines. What violates the spam policy is using automation to generate content primarily to manipulate rankings.

What counts as scaled content abuse? Generating many pages primarily to manipulate search rankings rather than help users. Google’s published examples include using generative tools to make many pages without adding value, scraping feeds or search results to spin up pages, automated translation without human review, and stitching together content from other sites without adding anything original.

Is there a page count that triggers scaled content abuse? Google has never published a threshold, and there almost certainly isn’t a fixed one. The policy is written around intent and pattern, not volume alone. A site publishing 500 genuinely differentiated pages is in a different position than a site publishing 50 templated ones.

Can I use AI to write blog posts and still rank? Yes, and plenty of sites do. The distinction that matters is whether a person is making decisions about each piece: what it covers, whether the facts are right, what it adds that nothing else does. AI as an assist is fine. AI as the whole assembly line is the exposure.

When was the March 2026 spam update? It launched March 24, 2026 and completed March 25, roughly 19 and a half hours later, making it the fastest confirmed spam update on Google’s Search Status Dashboard. A broad core update followed from March 27 to April 8, 2026.

Did Google announce new spam policies in 2026? No. Google described the March 2026 spam update as a normal spam update with no new policy categories. The rules stayed the same and the enforcement got faster.

How do I know if my traffic drop is a penalty? Check Search Console for a manual action first. Then compare average position against clicks. Position holding while clicks fall points to SERP changes such as AI Overviews rather than a penalty. Penalties usually appear as a cliff on a specific date, not a gradual slope.

Are AI Overviews reducing organic clicks? Pew Research found an 8 percent click rate on results pages with an AI summary versus 15 percent without, based on nearly 69,000 searches from 900 US adults. Google disputed the methodology. Multiple independent analyses through 2025 pointed in the same direction, though the exact magnitude varies by study and vertical.

Is programmatic SEO against Google’s rules? Not by itself. Programmatic pages built on genuine data, inventory, or meaningful differences between entries can be legitimate. The problem starts when the pages are near-identical with a variable swapped and no real value added per page.

How long does recovery from a spam update take? Google’s guidance is that algorithmic recovery requires its systems to observe compliance over a period of months, and there’s no reconsideration request for algorithmic actions. Manual actions are different: fix the issue, then submit a reconsideration request.

Should I delete thin pages? If a page exists only because a keyword existed, it’s more likely to be dragging than helping. The sunk cost of having paid for it isn’t a reason to keep it. Consolidate what can be merged, remove what can’t be defended, and keep what a real person would be glad to land on.

6. KEY TAKEAWAYS

• Scaled content abuse is about production method and intent, not writing quality. Google’s wording is “no matter how it’s created.”
• The policy isn’t new. It’s the old spammy auto-generated content rule, renamed in March 2024 and broadened.
• AI-assisted content is explicitly allowed. Google has said since February 2023 that it rewards quality however produced.
• The crackdown has already happened. March 2026 spam update in under 20 hours, core update through April 8, another spam update in late June 2026.
• No new policies were announced in 2026. Enforcement speed changed, not the rules.
• Detection works on pattern signatures in metadata, velocity, and structure, not on prose evaluation.
• Most traffic drops right now aren’t penalties. Check position against clicks before assuming enforcement.
• AI Overviews cut clicks roughly in half on affected queries per Pew Research, and there’s no reconsideration request for that.
• Penalties look like cliffs on specific dates. Slow declines usually aren’t enforcement.
• The defensible test for any page: could you explain to a person why this page exists, beyond the keyword?