The Otterly 2,000 AI Blogs Experiment: Method, Findings and Limits

OtterlyAI, a company that sells AI visibility tracking, tested the volume shortcut in generative engine optimization: publish a large batch of AI-written articles and let AI search pick them up. It ran the test in public, updated the write-up on 2026-08-05 with a fourth month of data, and its CEO presented it at BrightonSEO San Diego in September 2026.

The question was simple. What do Google, Bing and seven AI engines do with roughly 2,000 AI-generated posts on brand-new domains? The answer split cleanly by index, which makes the test useful for anyone deciding where to spend publishing effort. For a wider view of the tactic itself, see what one test showed about mass-publishing AI content.

Otterly set up two new domains about capybaras and loaded each with roughly 1,000 AI-generated blog posts in one click.

• capybara-world.com showed an author box under every post, presenting a fictional wildlife biologist with more than 15 years of field experience.
• capybaraverse.com had no author at all.

The author box was the one variable meant to differ. Otterly then tracked four layers for about four months:

1. Google Search Console data.
2. Bing Webmaster Tools data.
3. AI citations across seven engines: ChatGPT, Microsoft Copilot, Claude, Perplexity, Gemini, Google AI Overviews and AI Mode.
4. AI crawler visits to both sites.

Citations came from Otterly's own prompt-based tracker, which matters for how the counts should be read (see the limits below).

Google removed both sites. Otterly reports zero manual actions in Search Console on either domain, so the removal was algorithmic and unannounced.

The order surprised people. capybara-world.com, the site with the fictional expert, lasted about 14 days. It peaked at 2,426 daily impressions on April 5, then fell from 1,629 to 15 on April 9. capybaraverse.com, with no author, lasted about 90 days. It built to more than 21,000 weekly impressions by mid-May, then fell from 859 to 85 daily impressions on June 26 and to single digits after that.

This matches Google's written policy. Google's spam policies define scaled content abuse as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users," and the first example is using generative AI to produce many pages without adding value. Google's AI content guidance adds that "appropriate use of AI or automation is not against our guidelines." The violation is volume without value, not the tool. Our digest of Google's scaled content abuse policy covers the wording in full.

Losing Google did not end AI visibility. For capybara-world.com it grew. Otterly tracked 100 citations in April and 1,666 in the first 27 days of July, about 17 times as many, while the site was gone from Google's organic results. Otterly did not publish May or June figures.

Of the 3,934 citations tracked for capybara-world.com, 2,718 (69%) came from Microsoft Copilot. Perplexity and Claude cited the sites too.

Two points follow. First, an engine cannot cite what its index no longer holds. Google states a page must be "indexed and eligible to be shown in Google Search with a snippet" to appear as a supporting link in AI Overviews or AI Mode. Second, being indexed is necessary but not enough: ranking in Bing produced no ChatGPT citations here. The comparison of which AI answers use the Bing index and which use Google's explains the split, and why Copilot can cite you when Google's AI features don't covers the common case.

One caveat on counts. Microsoft's AI Performance report credited capybara-world.com with 42,817 citations, more than ten times Otterly's 3,934. The time windows may differ, so this is not an exact ratio. It does show that any prompt-based tracker sees a sample.

The test is public and documented, which is more than most GEO advice offers. It is still one pair of sites.

• One site per arm. The two sites also differed in domain name and everything that comes with a separate domain. The author box cannot be separated from that noise.
• No human-written control. Without a site publishing 1,000 human-written posts at once, the test cannot say whether Google reacted to AI text or to a thousand pages appearing overnight on a domain with no history. Google's policy targets the second, however the content is created.
• The author box result is one data point. It fits Google's statement that E-E-A-T isn't a specific ranking factor. It does not show that author boxes hurt.
• A low-stakes topic. Capybara questions are informational and uncompetitive. The result says little about commercial categories with established sources.
• Prompt-dependent counts. Otterly offers its spreadsheet on request, but until the prompt list and raw counts are public, outsiders cannot separate engine behavior from prompt design.
• A vendor ran it. That does not make the data wrong. It is why replication matters.

The test adds weight to what already has evidence behind it. It does not create a new tactic.

1. Do not mass-publish thin pages to shortcut AI search. Google removed both sites silently, and Copilot citations did not replace Google. If you are weighing cadence, read whether publishing more often helps AI visibility.
2. Check index status engine by engine. Look at Google Search Console and Bing Webmaster Tools separately, because one being healthy tells you nothing about the other. The per-engine index status playbook walks through it.
3. Confirm crawlers can reach you. Access comes before indexing. A robots.txt that allows a crawler while your CDN returns 403 removes you as quietly as a deindex. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
4. Measure each engine on its own. One engine cited these sites thousands of times while four barely did. A blended score would have hidden that.

None of this guarantees citations. It removes the blocks that are in your control.

Does Google penalize AI-generated content? Not for being AI-generated. Google's guidance says appropriate use of AI or automation is not against its guidelines. It acts against scaled content abuse, meaning many pages produced to manipulate rankings without adding value for users. In Otterly's test, Google removed both 1,000-post AI sites from its index, and Otterly reports no manual action on either domain, so the removal was algorithmic and came without notice.

Can AI engines cite a page Google has deindexed? Yes, if they draw on a different index. In Otterly's test, Microsoft Copilot, which grounds its answers in Bing, kept citing the sites after Google removed them, and Perplexity and Claude cited them too. Google's own AI surfaces almost never did, which fits Google's rule that AI Overviews and AI Mode only link to pages that are indexed and snippet-eligible.

Does an author box protect AI-generated content? Nothing in this test suggests it does. The site with a fictional expert author box was deindexed after about 14 days, and the site without one lasted about 90. With only one site in each arm, that does not prove author boxes hurt. It is simply no evidence that they help, and Google says E-E-A-T isn't a specific ranking factor.

Why did ChatGPT barely cite the sites if they ranked in Bing? The test does not explain it, only that it happened: ChatGPT's crawlers kept visiting, yet citations were effectively zero. The practical point is that "ChatGPT uses Bing" is not a reliable strategy. Being in an index makes a page eligible, but each engine decides what it retrieves and cites, and nobody outside the engine controls that step.

Why do Otterly's and Microsoft's citation counts differ so much? Otterly's tracker logged 3,934 citations for capybara-world.com, while Microsoft's AI Performance report counted 42,817. A prompt tracker only counts answers to the prompts it asks, and the two time windows may not match. Read the gap as a reminder that any prompt sample is partial. Where an engine publishes its own count, that count sees more.