Key takeaways
- Google's March 2026 core update targeted scaled content abuse specifically, and niche sites with 500+ unreviewed AI pages saw 60-80% traffic losses
- The penalty is often page-group specific, not sitewide: one SaaS blog audit found 30-50% drops on AI-heavy sections but 0-10% drops on human-written pages on the same domain
- AI search engines are making the same call independently: ChatGPT's citation share of Reddit collapsed from 3.8% to 0.5% in a single day in August 2026, and product/landing pages are losing citation share to how-to content and documentation
- There is no reliable automated AI-detector Google uses; the real signals are behavioral (bounce rates, pogo-sticking, publishing velocity) and structural (no bylines, no primary sources, duplicate semantic patterns)
- Recovery from a scaled-content penalty averages around six months, and the fix that actually works is consolidation, not touch-ups
Why this is suddenly urgent
For a couple of years, "just add AI content, it'll be fine" was a reasonable bet for a lot of sites. It stopped being a reasonable bet in early 2026. Google's February 2026 Discover core update went after low-quality AI content surfacing in Discover specifically, and the March 27 to April 8 core update that followed made scaled content abuse the primary enforcement target of the year. Reddit threads from site owners described losing 80% of traffic almost overnight. SEO agencies tracking the fallout put typical drops for affected sites at 20-35%, with the worst-hit pages losing more than half their traffic.
Here's the part that trips people up: this isn't really about whether a machine wrote the words. Google's own scaled content abuse policy is explicit that it applies "no matter how [the content] is created," including human-written filler. The policy has always been method-agnostic. What changed is that Google got much better at spotting the pattern that AI production makes trivially easy to fall into: lots of pages, thin value, no real author, no real experience behind any of it.
And it's not only Google. AI search engines are quietly doing their own version of this cleanup. Reddit's share of ChatGPT Search citations dropped from a steady 3.8% to 0.5% in the space of about a week in August 2026, an 86% relative collapse in a single day, according to Promptwatch's citation data. That's ChatGPT actively pulling away from unmoderated, low-signal content. Meanwhile Promptwatch's citation-type tracking for August 2026 shows how-to content more than doubling its share of ChatGPT citations while thin landing pages fell from roughly 20% to under 12%. Two different systems, two different incentive structures, arriving at the same conclusion: generic, low-effort content is losing ground everywhere, not just in Google's organic results.
If you built content volume over the last two years using AI without a heavy editorial layer, it's worth spending an afternoon checking your site against the seven signs below. None of these require guesswork. You can check all of them with tools you likely already have.
The 7 warning signs
1. Zero information gain
Open ten of your AI-assisted pages. For each one, ask: if I deleted every sentence a model could have generated purely from its training data, what's left? If the honest answer is "nothing," that page is paraphrasing the existing web, not adding to it. Google's quality systems reportedly measure something close to this directly, comparing semantic similarity across your own pages and against what's already indexed. Pages that just restate consensus information with no unique data, case study, or original opinion get classified as redundant, and that classification can trigger a sitewide suppression, not just a penalty on the individual URL.
The fix isn't rewriting the sentence structure. It's adding something the model genuinely could not have known: a result from your own testing, a quote from someone on your team, a number from your own data.
2. High impressions, low engagement, and a rising "bounce to search" pattern
Pull up Search Console and cross-reference impressions against Google Analytics engagement time for your highest-volume pages. The specific pattern to look for is decent or even growing impressions paired with a falling click-through rate, combined with users who click through and then bounce straight back to the search results within seconds. That round trip, sometimes called pogo-sticking, is one of the clearest behavioral signals that a page technically ranks but doesn't satisfy anyone. Under 30 seconds of dwell time on a long-form article is a reasonable threshold for flagging a page as suspect.
Google also shipped a genuinely useful tool for this in June 2026: a dedicated Search Generative AI performance report inside Search Console, showing impressions and clicks specifically from AI Overviews and AI Mode, filterable by page, country, device, and date. If a page gets decent traditional impressions but essentially zero AI Overview presence, that's worth investigating on its own.
3. Generic expertise with no real E-E-A-T signals
Read a sample of your pages looking specifically for first-person language, original photography or screenshots, and any anecdote that couldn't have been lifted from a competitor's article. If every page reads like it was written by nobody in particular, about a topic nobody on your team has actually touched, that's the pattern automated quality systems and human quality raters are trained to flag as untrustworthy.
The concrete fix is mechanical: every article needs a named author with a real bio, ideally implemented with Person schema (name, credentials, a sameAs link to a LinkedIn profile) and an Organization schema for the site itself. A 100-150 word bio that states actual years of experience and credentials does more for trust signals than another round of AI polish on the body copy.
4. Poor machine readability
This one costs you twice, once in traditional search and once in AI visibility. Walls of unbroken text, vague headings like "Introduction" and "Conclusion," and no structured data make it hard for any system, human or machine, to extract a clean answer from your page. If an LLM can't pull a snippet-ready answer out of your content, it stops citing you, and that's a direct hit to your presence in AI Overviews, AI Mode, and chat-based answers.
The fix is formatting discipline: headings that state the actual question being answered, short paragraphs, bulleted lists where a list genuinely helps, and schema markup that describes what the page actually is. If you want a second opinion on how your content stacks up specifically for AI citation, tools like Promptwatch will show you which of your pages are actually getting cited across ChatGPT, Gemini, Perplexity, and AI Overviews, and which ones AI crawlers visit but never cite.
5. Content flooding: hundreds of pages, a handful of them doing all the work
Check how traffic is distributed across your site. If a small fraction of your URLs drive nearly all the organic traffic while the rest sit at near-zero, you've built a content farm shape, even if that wasn't the intent. Google's enforcement in 2026 has increasingly treated that pattern as a signal of volume-over-value production.
The standard fix that's actually been shown to work in post-penalty recoveries is consolidation: identify pages competing for the same intent (sometimes called keyword cannibals), merge them into one genuinely comprehensive piece, and 301-redirect the old thin URLs into it to preserve whatever link equity they had. Trying to incrementally improve dozens of thin pages one at a time is reportedly far less effective than cutting the number of pages and making the survivors better.
6. Abnormal publishing velocity for your site's age and editorial capacity
If your publishing rate outran what your editorial team could plausibly have read end to end, that's a structural red flag independent of any individual page's quality. This is one of the clearest site-level patterns associated with the March 2026 enforcement wave: sites with 500+ AI pages published in a single year and no proportional increase in editorial staff took some of the hardest hits, in the 60-80% traffic loss range.
The rule that prevents most of this is simple and has nothing to do with AI detection: no page should be published that a human hasn't read start to finish. If your content calendar can't sustain that, the calendar is the problem, not the writing tool.
7. Your link and citation neighborhood is other thin content
Look at who links to you and, increasingly, who mentions you in places AI systems pull from. If your backlink profile is mostly other low-quality sites or content farms, and no authoritative source has ever cited or mentioned your work, that's a trust deficiency that both Google and AI answer engines pick up on. Mid-authority domains (roughly DR 46-75) captured nearly half of all ChatGPT citations in August 2026 according to Promptwatch's domain-rank analysis, while the very biggest sites (DR 91+) saw their citation share fall from around 7% to 3% over the same period. Authority alone doesn't buy citations; specificity and real-world credibility do.
A quick self-audit table
| Warning sign | Where to check | Red flag threshold |
|---|---|---|
| Zero information gain | Manual page review | Nothing left if you strip the generic parts |
| Bounce-to-search pattern | Search Console + GA4 | Under 30 sec dwell time, falling CTR |
| Missing E-E-A-T | Author bylines and bios | No named, credentialed author |
| Poor machine readability | Page structure, schema | Vague headings, no structured data |
| Content flooding | Traffic distribution by URL | A few pages drive 95%+ of traffic |
| Publishing velocity | Content calendar vs. team size | Output nobody could have reviewed |
| Thin link/citation neighborhood | Backlink and AI citation data | Links and mentions mostly from low-quality sites |
What actually works instead
None of this means AI-assisted writing is off the table. It means the review layer is the product now, not the drafting step. A few habits separate content that survives core updates from content that gets wiped out:
- Never publish a draft nobody has read end to end. This single rule prevents most of what got penalized in 2026.
- Put a real, verifiable name on every page, with a bio that links to genuine background.
- Add something the model could not know: an expert quote, a case outcome, your own data, a firsthand observation.
- Fact-check every number, date, and citation against a primary source.
- Audit by page template or URL group, not by picking three pages you already feel good about. Google evaluates at the group level, and you should too.
If you're trying to figure out whether your AI content program is quietly building an asset or a liability, and you'd rather get a second set of eyes on it than guess, 1001 SEO Media does exactly this kind of audit work: technical SEO, content review, and GEO strategy for sites that scaled content faster than they scaled editorial oversight.
For ongoing monitoring once you've cleaned things up, a GEO-focused visibility platform is worth the subscription. Promptwatch tracks which of your pages actually get cited across ChatGPT, Claude, Gemini, Perplexity, and Google's AI surfaces, shows you AI crawler activity down to the page level, and flags content gaps before a core update finds them for you. If you want to see how it stacks up against other options in the space, the GEO software directory at bestgeosoftware.com is a reasonable place to compare a few platforms side by side before committing to one.
The pattern across both Google's algorithm and AI answer engines in 2026 is consistent: volume without a real editorial backbone is getting more expensive, not less, every quarter. Fixing the seven signs above won't make your content rank by itself, but it removes the specific reasons a core update, or an AI model's source-quality filter, would have for ignoring you.