Updated August 5, 2026 with a fourth month of data and a new chapter: what Bing did with the same content, and why ranking in Bing did not get us into ChatGPT.
Can you shortcut your way into AI Search by publishing thousands of AI-generated blog posts with a single click? We tested it. We built two brand-new websites, filled each with roughly 1,000 fully AI-generated articles in less than a day, and tracked what Google and seven AI Search platforms did with them for nearly four months.
The short answer: Google silently killed both sites without a single warning, while Bing didn’t. Most AI Search platforms did not. Nearly four months after Google removed it, the deadest of the two sites is earning more AI citations every single month, and it is now cited more than two real capybara websites we tracked as competitors.
This is our most failed experiment so far, in two ways. The tactic failed, and part of the experiment design failed too. We are publishing both, because the data tells a more interesting story than a simple “AI slop does not work.”
Both test sites were deindexed by Google without any warning. Google Search Console showed zero manual actions on either domain. No penalty notice, no message. Traffic just stopped.
The first website died roughly after 2 weeks. capybara-world.com peaked at 2,426 daily Google impressions on April 5, 2026 and collapsed overnight on April 9, from 1,629 impressions to 15.
The control site survived 3 months, then died the same way. capybaraverse.com grew to over 21,000 weekly Google impressions by mid May, then collapsed on June 26, from 859 daily impressions to 85, then to single digits.
AI citations went up, not down, after Google deleted the sites from its index. capybara-world.com went from 100 citations in April to 1,666 in the first 27 days of July, a 17x increase in the period when it was invisible on Google.
Microsoft Copilot absorbed the AI slop. Google and ChatGPT did not. Of capybara-world.com’s 3,934 citations, 2,718 (69%) came from Microsoft Copilot alone. ChatGPT, Gemini, Google AI Overviews and Google AI Mode cited it essentially zero times.
ChatGPT crawled the sites hardest and cited them least. ChatGPT-User logged 4,808 page visits across both domains, more than every other AI agent combined, and produced almost no citations.
New in the August update: Bing rewarded what Google executed. Bing never deindexed either site, grew their impressions every single month to a combined 119,935 in July, and Microsoft’s own AI report credits capybara-world.com with 42,817 citations. Ranking in Bing still produced exactly zero ChatGPT citations, and we explain why below.
Why This Matters for AI Search
One-click AI content generation tools now promise exactly this workflow: generate hundreds or thousands of posts, publish, and wait for traffic. The pitch sounds plausible because early results look real. Both of our sites got indexed, ranked, and earned clicks within days.
The question nobody answers with data is what happens next, and whether Google and AI Search react the same way. Based on this experiment, they do not, and the split is not where most people assume it is. It does not run between “Google” and “AI Search.” It runs between individual AI Search platforms, and the gap between them is enormous.
Definition: What we mean by scaled AI content. Roughly 1,000 blog posts per site, fully AI-generated with an off-the-shelf AI content tool that connected to WordPress, published in less than one day, with no human editing beyond the initial setup.
Methodology: Two Capybara Websites, One Variable
We registered two fresh domains with no history, both in the same harmless niche: capybaras.
Version A
Version B
Domain
capybaraverse.com
capybara-world.com
Content
~1,000 AI-generated blog posts
~1,000 AI-generated blog posts
Difference
No author signals
Author boxes with detailed bios on every post
Publishing
One click, live in under a day
One click, live in under a day
Launch
Early April 2026
Early April 2026
The variable we wanted to isolate was authorship. Version B displayed an author box under every article, presenting a fictional wildlife biologist with 15+ years of field experience. The hypothesis: author signals strengthen perceived E-E-A-T and help scaled AI content survive longer.
Version B’s articles were instructed to be complete rewrites, not copies of Version A. The site design and structure were near-identical.
Version A (capybaraverse.com) and Version B (capybara-world.com) with the added author box.
We tracked both domains in three places: Google Search Console for organic performance, server logs for AI bot activity via OtterlyAI’s Agentic Analytics, and an OtterlyAI Brand Report for AI citations across ChatGPT, Microsoft Copilot, Claude, Perplexity, Gemini, Google AI Overviews and Google AI Mode.
Measurement scope: all citation data below comes from our test website’s brand report on OtterlyAI, April 1 to July 27, 2026, covering 482,394 tracked citations across seven AI Search platforms. capybara-world.com is the tracked brand domain; capybaraverse.com is tracked as a competitor domain in the same report.
How 2 failed experiments still ended up giving us useful insights
Full transparency: this was not a clean A/B test. Two sites with 1,000 posts each, in the same niche, launched in the same window, ended up competing for the same queries. Search Console shows both domains surfacing for identical questions like “how many capybaras are left in the world?”
Failed experiment #1: Could not determine if the author variable had an impact: The sites cannibalized each other, despite explicitly instructing the AI content tool to generate completely different sets of content pieces. So the author variable cannot be cleanly isolated from the results.
Failed experiment #2: Is it better to have 1 large FAQ page or separate FAQs?: A second sub-test, comparing one giant FAQ page with 1,000+ questions against FAQs embedded at the end of every post, was compromised for the same reason. We will rerun that as a separate experiment.
What the setup still measures reliably: what Google and each AI Search platform do with 1,000 instantly published AI posts, twice, on two independent domains. That signal is unambiguous.
Headline Numbers: 4 Months of Search Console Data
Data below is from Google Search Console, Web search, April 1 to July 25, 2026.
Metric
capybaraverse.com (A)
capybara-world.com (B)
Total clicks
962
39
Total impressions
149,313
10,778
Average CTR
0.6%
0.4%
Average position
9.7
10.8
Pages with impressions
960
214
Peak day
May 7: 4,202 impressions
April 5: 2,426 impressions
Time until collapse
~12 weeks
~8 days of visibility
Google Search Console performance for both domains, April 1 to July 25, 2026.
Site B: dead in two weeks
capybara-world.com was indexed on April 3 and immediately started earning impressions: 1,742 on April 4, 2,426 on April 5, holding above 1,600 daily through April 8. On April 9 it fell off a cliff: 15 impressions. It never recovered. Average position for the leftover impressions drifted from around 10 to between 30 and 50, which is what a demoted domain looks like from the inside.
Site A: it works, until it doesn’t
capybaraverse.com looked like a success story for ten weeks. Impressions grew steadily to a peak of roughly 21,700 per week in mid May, ranking top 10 on average, earning real clicks on real questions. Then the same pattern: 859 impressions on June 25, 85 on June 26, 9 on June 27. The site has been flatlined since.
What this means for your strategy: the early traction phase is exactly what makes one-click content tools look like they work. Screenshots taken in week 6 of this experiment would have made a convincing case study for scaled AI content. The collapse arrives later, and it arrives all at once.
Google Never Warned Us: The Silent Deindexing
This is the part that surprised us most. Google Search Console shows “No issues detected” under Manual actions for both domains, to this day. There was no penalty notification, no email, no message.
Instead, we watched the index coverage report quietly shift: on capybaraverse.com, 478 pages sat in “not indexed” against 538 indexed. Far away from having all pages completely indexed early in April.
Search Console: Manual actions show “No issues detected” while pages drop out of the index.
You could call this a shadow ban. The more precise description: Google’s scaled content abuse systems act algorithmically, and algorithmic suppression does not generate a manual action. You get no signal you can appeal against. Your only warning is the traffic graph.
What this means for your strategy: do not count on Google to tell you that scaled AI content crossed a red line. By the time you can see the problem, the domain is already gone. On a test domain that costs you nothing. On your company’s primary domain, that risk is existential.
AI Crawlers Kept Coming Even After Google Left
While Google was deindexing both sites, AI crawlers behaved as if nothing had happened.
Across the experiment window we logged, per site:
AI agent
Website A visits
Website B visits
ChatGPT-User (On-Demand Fetcher)
2,799
2,009
GoogleOther / R&D Fetcher
17
64
Perplexity AI Crawler
43
1
Google NotebookLM Fetcher
4
0
Meta Web Indexer
3
0
Claude On-Demand Fetcher
0
2
Mistral On-Demand Fetcher
1
0
Total AI agent page visits
2.9K
2.1K
AI agent page visits over time for both domains. AI bots still crawl the websites to this day.
August update: in the most recent 90-day window, ChatGPT-User alone logged 5,332 page visits across the two sites, still more than every other AI agent combined. OpenAI’s search-index crawler, OAI-SearchBot, visited just 80 times in the same window. Keep that split in mind for the new chapter below.
Notable detail: site B, deindexed since mid April, recorded 603 human referral visits against site A’s 272. The dead site sent more humans to its pages than the one that survived three months.
The AI Search Citation Data: The Dead Site Keeps Growing
Here is the finding that changed how we read this whole experiment. We expected AI citations to decay after Google removed the sites, following the ghost citation decay curves we measured in earlier research. They did the opposite.
Month
capybaraverse.com (A)
capybara-world.com (B)
April
85
100
May
467
845
June
920
1,323
July (1 to 27)
734
1,666
Total
2,206
3,934
AI Search Citations vs organic traffic
Monthly AI Search citations vs organic traffic, both domains, OtterlyAI Brand Report, United States.
capybara-world.com was deindexed by Google on April 9. Every month since, its AI citations have gone up: 100, then 845, then 1,323, then 1,666 in just 27 days of July. On a daily basis that is 3.3 citations per day in April against 61.7 per day in July, an 18x increase, entirely during the period when the site did not exist as far as Google was concerned.
Site A shows the first hint of a turn. Its citations peaked in June at 920 (30.7 per day) and slipped to 734 across 27 days of July (27.2 per day), an 11% drop in daily rate in the month after Google deindexed it. One month is not a trend, and we will keep measuring.
Top cited pages from the dead domain over the full period:
Page
AI Citations
Can Capybaras Kill You? Facts You Need to Know
220
How Many Capybaras Are Left in the World 2025?
201
Is the Capybara Endangered? Facts You Need to Know
185
Homepage (capybara-world.com)
158
Are Capybaras Edible? Facts You Need to Know
132
OtterlyAI Citations report: the deindexed capybara-world.com keeps earning AI citations.
For context on what those numbers mean competitively: capybara-world.com ranked 25th of all cited domains in this category with a 1% citation share. It was cited more than capybaramag.com (3,362) and capybarafacts.com (2,057), two real capybara websites tracked as competitors in the same report. A site Google deleted in April is beating real publishers in AI answers in July.
Not All AI Search Platforms Behave the Same
Splitting the same citations by platform explains almost everything, and it is where the experiment stops being a simple morality tale.
AI Search platform
capybaraverse.com (A)
capybara-world.com (B)
Microsoft Copilot
543
2,718
Claude
1,005
714
Perplexity
628
501
Google AI Mode
19
0
Google AI Overviews
11
0
ChatGPT
0
0
Gemini
0
0
Monthly AI citations by AI Search platform, both sites combined, April to July 2026.
Three things stand out.
Microsoft Copilot absorbed the AI slop wholesale. 69% of capybara-world.com’s citations came from Copilot alone. In Copilot’s slice of this report, the dead AI slop site holds an 8% citation share of every URL Copilot cited on capybara topics. That is a single one-click generated site claiming roughly one in twelve of Copilot’s sources in a category that includes Wikipedia, National Geographic and the San Diego Zoo. August update: in July, capybara-world.com became the single most cited domain in the whole category on Copilot, at 14% share, ahead of Wikipedia and Britannica.
ChatGPT crawled hardest and cited least. ChatGPT-User was by far the most aggressive agent on both domains, 4,808 page visits between them, more than every other AI crawler combined. ChatGPT’s citations of those pages: effectively zero, out of 120,902 ChatGPT citations tracked. Crawl volume is not a proxy for citation value, and anyone selling you AI bot hits as a GEO KPI should be asked about this chart.
Google’s suppression carried across its whole ecosystem. Google AI Overviews, Google AI Mode and Gemini cited capybara-world.com zero times and capybaraverse.com 30 times combined. Whatever classifier removed these sites from organic search removed them from Google’s AI surfaces too. That is a consistency worth noting: for Google properties, your Search penalty is your AI Search penalty.
What this means for your strategy: “Am I visible in AI Search?” is not one question. In this experiment the same content had an 8% citation share on one platform and 0% on four others. Measure per platform, or you will draw exactly the wrong conclusion from an average.
August Update: What Bing Did With the Same Content
New chapter, added August 5, 2026. Data: Bing Webmaster Tools exports for both domains, OtterlyAI Brand Report and Agents Analytics through August 5.
Everything above describes how Google treated 2,000 instant AI posts. We then pulled the same story from Bing Webmaster Tools, and the two search engines looked at identical content and reached opposite verdicts.
Bing rewarded what Google executed
Bing never deindexed either site. No collapse, no demotion, no cliff. Combined Bing impressions grew every single month: 3,882 in April, 44,855 in May, 73,712 in June, 119,935 in July. In the exact month Google served the two sites 751 impressions, Bing served them 160 times more. capybara-world.com, the site Google removed after 8 days, became Bing’s favorite: 91,104 July impressions and a peak day of 9,618.
Google and Bing impressions, both sites combined. Two search engines, identical content, opposite verdicts.
Bing’s own AI layer went further. Bing Webmaster Tools’ AI Performance report, which counts citations of your site across Microsoft’s AI experiences on real user queries, credits capybara-world.com with 42,817 citations from April to early August, growing from 314 in April to 19,072 in July. capybaraverse.com’s grounding-query report adds 5,574 more across 133 queries, reaching up to 42% citation share on individual questions.
The myth: “Rank in Bing and ChatGPT will cite you”
A popular GEO claim says ChatGPT search runs on Bing’s index, so Bing rankings translate into ChatGPT citations. Our sites are the perfect test: they rank well in Bing, Microsoft’s own AI cites them tens of thousands of times, and ChatGPT’s crawler knows them better than any other AI agent.
The result: 0 ChatGPT citations, out of 122,834 tracked ChatGPT answers.
Bing impressions climb, ChatGPT crawls thousands of times, ChatGPT citations stay at zero.
This is not because ChatGPT never saw the sites. ChatGPT-User fetched their pages 5,332 times in the last 90 days, more than every other AI agent combined. But OAI-SearchBot, the crawler that actually builds ChatGPT’s search index, came only 80 times. ChatGPT reads these sites constantly in conversations and admits them to its search citations never.
Copilot proves the index is not the problem
Microsoft Copilot runs on the same Bing index the myth points to, and Copilot cited the sites 3,232 times in four months, tracking Bing’s growth curve almost 1:1. Copilot does not operate a dedicated AI crawler at all; Agents Analytics logged zero Microsoft agents on either site, because Copilot’s grounding is fed by Bingbot, the ordinary search crawler.
Copilot citations track Bing impressions almost perfectly, while ChatGPT crawls more than any other agent and cites zero.
That is the mechanical answer to the myth. Copilot has no quality gate between Bing’s index and its answers, so Bing visibility flows straight through. ChatGPT interposes its own crawler, its own index and its own source selection on top of whatever Bing supplies, and that layer rejected this content completely. If Bing rank transferred automatically, Copilot and ChatGPT would look identical in this experiment. They are exact opposites.
Perplexity: the first platform where the decay began
Perplexity is the third data point, and it behaves like neither. It barely crawled the sites, 100 PerplexityBot visits all experiment, yet cited them 1,129 times. And it is the first platform showing the decay we originally predicted for all of them: Perplexity citations peaked at 444 in May and fell every month since, down 53% to 208 in July, while Bing impressions nearly tripled over the same stretch.
Perplexity citations decay while Bing impressions climb: the first platform to filter the content out.
One platform absorbs the content wholesale (Copilot), one never let it in (ChatGPT), and one let it in early and is now actively filtering it out (Perplexity). Same content, same window, three completely different retrieval systems.
What this means for your strategy: Bing Webmaster Tools rankings, and even Bing’s own AI Performance numbers, are not a proxy for ChatGPT visibility. Track citations per platform, on the platform itself. And treat “AI crawler visits” with suspicion as a KPI in both directions: ChatGPT crawled 5,332 times for zero citations, while Perplexity turned 100 crawls into 1,129.
So Did the Tactic Fail? Yes, But Not Symmetrically
The honest conclusion is not “AI slop does not work anywhere.” The data shows a sharply asymmetric outcome:
In Google, including its AI surfaces: scaled AI content is a domain-killing move. Both sites are gone from organic search, AI Overviews, AI Mode and Gemini, silently, with no path to appeal.
In Microsoft’s ecosystem: the content is not just tolerated but rewarded. Bing organic impressions grew every month, and on Copilot the dead site now out-cites Wikipedia in its category.
In Perplexity: the content got in early and is now being filtered out, month by month.
We predicted these citations would decay. Through July 27, on Copilot, they had not; for the site Google removed first they were accelerating, and we would rather publish that than quietly drop it. August update: the decay has now begun, on exactly one platform. Perplexity citations have fallen 53% from their May peak. Copilot keeps climbing. The prediction was directionally right and platform-wrong: decay is not a property of AI Search, it is a property of each platform’s filtering.
But the asymmetry is a trap, not an opportunity, for three reasons. The visibility is concentrated in a single platform you do not control and cannot predict. It is built on content that did not survive Perplexity’s filter and has already failed on Google’s. And the cost of entry is permanently burning a domain in the channel that still drives the majority of most websites’ traffic.
On a capybara test domain that trade cost us nothing. On a domain your business depends on, it is not a strategy. It is a countdown.
Recommendations for Marketers and SEOs
Never run scaled, unedited AI content on a domain you care about. Both of our test domains died in Google, permanently and without warning. The only reason this experiment was cheap is that capybaras were never our business.
Do not mistake early traction for validation. Ten weeks of growing impressions preceded the collapse of site A. Judge scaled content tactics on quarters, not weeks.
Watch your traffic graph, not your Manual actions tab. Algorithmic suppression arrives without notice. If you operate programmatic or AI-assisted content at scale, monitor indexed page counts weekly.
Never report AI visibility as a single number. This experiment produced an 8% citation share on Copilot and 0% on ChatGPT, Gemini, AI Overviews and AI Mode from identical content. Track citations per platform.
Do not treat AI crawler hits as a success metric. ChatGPT crawled these sites 5,332 times and cited them never. Perplexity crawled 100 times and cited them 1,129 times. Bot visits measure interest, not trust.
Do not treat Bing rankings as a ChatGPT strategy. Bing served these sites 119,935 impressions in July; ChatGPT cited them zero times in 122,834 tracked answers. The Bing-to-ChatGPT pipeline the myth promises does not exist as an automatic transfer.
Expect your content mistakes to outlive your cleanup. Whatever you publish can persist in AI Search memory long after deletion or deindexing. Publishing recklessly is a long-term liability, not a reversible test.
If you use AI for content, use it for content AI does not already know. Original data, first-hand experience, and proprietary insight survive. Rewording what the model already knows, 1,000 times, is exactly the pattern that got both sites removed.
What Comes Next
This GEO experiment keeps running in the background. Open questions we are tracking:
Do capybara-world.com’s Copilot citations ever peak, and how fast do they fall when they do?
Is site A’s 11% July dip the start of the decay curve we originally predicted, or noise?
Why does Copilot treat this content so differently from every other platform?
Does either domain ever return to Google’s index without intervention?
A clean rerun of the FAQ placement + author bio test on fresh, non-competing domains.
A follow-up: ABC-experiment: 100% AI vs 80% AI vs 0% AI content, co-published with Seeders.
This connects directly to our Dead Citations study of 20M+ AI answers, where 19% of cited URLs were no longer live. Deleted and demoted content lives on in AI Search memory. This experiment suggests it can do more than linger. It can grow.
Got a GEO experiment idea? Let’s collaborate
At OtterlyAI, we don’t write about AI Search from the sidelines. We test it. With the community, we run and update a public GEO Experimentation Tracker showing what we are running and what moved AI Search. If you are running your own GEO experiment, email rick.tousseyn@otterly.ai and we may feature it with your name attributed.
👉 Want to measure your AI Visibility? Start with OtterlyAI and our GEO Guide.
