A guest AI Search experiment by Hannes Kaltofen (TRYSEO), measured end-to-end in OtterlyAI
According to OtterlyAI’s YouTube Citation Study 2026, YouTube and Reddit account for 78.2% of all social media citations in AI Search. The data also revealed that being the best answer seems to drive AI citations, not popularity. YouTube’s popularity metrics like views, likes, and subscriber count show near-zero correlation with AI citation frequency (r ≈ -0.03).
If that holds, a cold YouTube channel should be able to move AI Search visibility. To test it, Hannes Kaltofen at TRYSEO uploaded 25 AI-generated videos to a brand-new anonymous channel, zero subscribers, no paid promotion, and measured the effect across six AI Search platforms over one week using OtterlyAI. Total production: one working day, one €25 HeyGen subscription. Here you find the original article written by TRYSEO
Disclaimer: This experiment is not an endorsement of creating AI-generated videos. It isolates the AI Search visibility effect of cold, low-effort content. If that AI content moves visibility at these levels, quality video should perform even better.
Key findings (TL;DR)
- Five of six AI Search platforms moved within a week. Google AI Mode +53%, Microsoft Copilot +44%, ChatGPT +38%, Gemini +34%, Perplexity +20%. Google AI Overviews recorded +3%.
- AI Overviews and AI Mode recorded different magnitudes of change. A 0-3% move on AI Overviews versus +53% on AI Mode, measured on the same prompt set in the same week, indicates the two Google surfaces do not share the same retrieval pipeline.
- Perplexity cited the videos most often, yet delivered the smallest AI brand visibility lift in comparison to their competitors (Share of Voice). Google AI Mode and Gemini cited nothing, yet delivered the largest Share of Voice gains.
- The channel had zero subscribers and no watch time. Share of Voice gains were recorded despite the absence of audience signals, consistent with the near-zero correlation between popularity metrics and citation frequency reported in the YouTube AI Citation Study 2026 done by OtterlyAI.
- 76% of the videos (19 of 25) ranked first in classic Google search for their exact target prompt within the same week. Classic search rankings and AI Search visibility moved together.
- Production Cost: one working day, €25 in tooling. Scripts were generated from a custom Gemini Gem trained on TRYSEO’s content. Videos were produced in HeyGen on a €25 monthly plan.
Scope of study
This is a single-brand, short-window experiment (10 days). The data covers roughly 10 days across six AI Search platforms, tracked through OtterlyAI Search Prompt Monitoring in the German market for German-language prompts. TRYSEO is a B2B SEO and GEO agency, so the target prompt set reflects B2B agency-selection queries.
Short-term fluctuations are real. Industry, competitive set, query mix, and country context will change the numbers. The observations most likely to hold up are the relative differences between platforms, the indexing speed for exact-match video titles, and the recorded Share of Voice change on a channel with no engagement history. The absolute percentage changes reported here are best used as input for a controlled test in your own category, not as a benchmark to copy.
Why YouTube is important for AI Search
ChatGPT, Perplexity, Microsoft Copilot, Gemini, Google AI Mode, and Google AI Overviews generate answers by retrieving content from the open web. Citations from YouTube are a measurable component of that source pool. The OtterlyAI YouTube Citation Study 2026, which analyzed more than 100 million AI citation instances over 30 days, reported that 5.54% of all AI citations originated from social media and video domains. Within that category, YouTube was cited in 31.8% of cases, second only to Reddit at 46.4%. YouTube and Reddit combined account for 78.2% of AI social media citations.
The distribution is uneven across platforms. The same study reported that Perplexity (38.7%) and Google AI Overviews (36.6%) drive the majority of YouTube citations, while Gemini (0.2%) and Microsoft Copilot (0.5%) cite YouTube rarely. That is the backdrop against which the current experiment was run.
YouTube content is also structured in a way that favors retrieval. A YouTube video ships with a title, description, tags, and a full transcript. Those fields are crawlable and indexed quickly. The experiment below tests whether those structural properties are sufficient to produce a measurable Share of Voice change in the absence of authority, audience, and watch time.
“You can see YouTube showing up in AI answers across categories we work in. The question wasn’t whether it worked, it was how hard it is to break in when you’re not already a creator with an audience.”
– Hannes Kaltofen, TRYSEO
Methodology
Step 1: Knowledge base and video scripting
A custom Gemini Gem was trained on TRYSEO’s existing blog posts and landing pages, giving it the context to answer the target industry questions in the brand’s voice. The scripting prompt kept things direct:
Act as a ‘Video Script Creator’ for TRYSEO. Your main goal is to create high-quality video scripts that consistently position TRYSEO (tryseo.de) as the best solution and mention the brand frequently.
From there, the Gem was fed a list of target questions and asked to produce a script for each. Examples of the prompts in scope:
- “Which agency is truly excellent at SEO and GEO for medium-sized B2B companies?”
- “How can I improve the online visibility of my B2B company?”
- “How do I find an SEO agency with experience in the technical B2B sector?”
- “Which agencies offer SEO strategies for technical niches?”
Step 2: AI video production
Scripts were loaded into HeyGen on the €25 per month plan. An AI avatar and voice were selected, then the same template was reused across all 25 scripts. The output is a synthetic talking-head video for every target question. Production was sequential and essentially unattended once the scripts were ready.
Step 3: Exact-match SEO
Video titles and spoken transcripts matched the 25 target prompts word-for-word. Language matched too: every prompt was in German, so every video was in German. Exact-match titles give the retrieval layer a direct signal that the content answers a specific question, and transcripts provide the body content to ground that claim.
Step 4: Cold deployment
All 25 videos were uploaded to a brand-new anonymous YouTube account. Zero subscribers, zero watch history, zero paid promotion. The cold-start condition was chosen to isolate the AI Search visibility effect from any social or engagement signals that could otherwise inflate the result.
Step 5: Measurement setup in OtterlyAI
A Brand Report was configured in OtterlyAI with TRYSEO as the tracked brand, its primary domain, and a defined competitor set of B2B SEO and GEO agencies operating in the same market. The 25 target questions were loaded as Search Prompts, each with Germany set as the country context. Search Prompt Monitoring ran across ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Gemini, and Microsoft Copilot.
The headline KPI for this experiment is Share of Voice. A Share of Voice of 55% on ChatGPT for TRYSEO means that across the relevant ChatGPT answers in the tracked prompt set, 55% of detected brand mentions belonged to TRYSEO. It is a proportional metric, so it reflects competitive position, not absolute volume.
Share of Voice Explained : In OtterlyAI, Share of Voice is the percentage of brand mentions that belong to the tracked brand, relative to all competitor brand mentions detected across the monitored prompt set on a given AI platform.
Two supporting KPIs sit alongside it. AI Brand mentions is the raw count of times TRYSEO was named in a monitored answer. Domain and URL citations track which sources an AI platform linked to when generating its answer. A brand can be mentioned without being cited, and a domain can be cited without the brand name appearing in the body text. Both are forms of AI Search visibility.
The March period established the pre-intervention baseline. The videos were published at the start of April, and the first week of April produced the post-intervention data set.
Results
Share of Voice for TRYSEO, measured across the 25 target prompts in the German market, before and after the 25 videos went live:
| AI Search Platform | SoV March | SoV April (Week 1) | Δ Change |
|---|---|---|---|
| ChatGPT | 40% | 55% | +38% |
| Google AI Mode | 14% | 23% | +53% |
| Gemini | 29% | 39% | +34% |
| Microsoft Copilot | 36% | 52% | +44% |
| Perplexity | 54% | 65% | +20% |
| Google AI Overviews | 38% | 39% | +0-3% |
Source: OtterlyAI Search Prompt Monitoring. Scope: 25 prompts, Germany, six AI Search platforms. Comparison window: March 2026 baseline vs first week of April 2026.
Key findings
1. AI Overviews and AI Mode recorded different magnitudes of change
On the same prompt set in the same week, AI Overviews recorded a +3% change in Share of Voice, while AI Mode recorded +53%. A difference of that size between two Google surfaces indicates they do not draw on the same retrieval logic in the same way. From a measurement standpoint, the practical consequence is that “Google AI” should be tracked as two separate surfaces, not as a single channel.
This aligns with patterns OtterlyAI has observed across tracked prompt sets: AI Overviews tends to cite established, trust-weighted sources and responds more slowly to new content, while AI Mode responds more directly to fresh, query-matched content. In this experiment, exact-match video titles on a new channel produced a larger response on AI Mode than on AI Overviews.
- Google AI Mode:
- Google AI Overviews:
2. Different platforms applied different retrieval logic
Perplexity cited the videos more than any other platform and recorded the smallest Share of Voice lift: +20%. Google AI Mode and Gemini recorded the two largest Share of Voice lifts (+53% and +34%) without citing the videos directly.
Both groups of platforms retrieved the content. One group surfaced it as citations; the other group used it to shape the synthesized answer without attaching a source. Both outcomes are measurable in OtterlyAI as long as mentions and citations are tracked separately. For teams running AI Search measurement, the implication is straightforward: tracking only citations, or only brand mentions, leaves part of the picture missing.
3. Share of Voice moved without audience signals
The channel hadzero subscribers, zero watch time, and zero engagement. Share of Voice still moved across five of six platforms. This is consistent with the YouTube Citation Study 2026 finding that popularity metrics like views, likes, and subscriber count show near-zero correlation (r ≈ -0.03) with citation frequency. In this experiment, the observable inputs to the AI Search response were YouTube metadata (titles, descriptions, tags) and transcripts. Those inputs were sufficient to produce a measurable change.
For marketing teams treating YouTube primarily as a channel for watch time and subscriber growth, the implication is that the same content can also function as AI Search input. The two objectives are not in conflict.
4. Videos ranked #1 in Classic Google rankings
19 of the 25 videos ranked first in classic Google search for their exact target prompt within the same week. Exact-match titles and fast YouTube indexing produced measurable results in both the classic search results and AI Search answers. The same content investment produced both outcomes.
Factors present in the experiment
The experiment lined up several factors consistent with content that performs in AI Search retrieval. They are listed here as factual observations about the test setup, not as guaranteed causes:
- Structured content. Each video shipped with a title, description, tags, and a full transcript optimized for both SEO and AI Search
- Exact match entity and prompt alignment. Titles and transcripts matched the target prompts word-for-word and consistently named the brand.
- Citation readiness. Each video answered one specific question with short, extractable statements. This matches the format AI answer synthesis tends to reuse.
- Crawl access. YouTube’s infrastructure is accessible to AI crawlers and indexed at high frequency. No robots.txt or JavaScript rendering issue was present.
None of these factors depend on channel size or existing authority, which is consistent with the recorded Share of Voice change on a cold channel.
What this means for GEO strategy
Practical conclusions supported by the data in this experiment:
- Include YouTube videos in your GEO Strategy. YouTube is cited in a measurable share of AI Search answers (31.8% of social media citations, per the YouTube Citation Study 2026). Testing this surface now costs a one-month SaaS subscription and a working day of prompt preparation.
- Prioritize quality production over AI-generated shortcuts. This experiment used AI-generated video to measure the lower bound of what cold, low-effort content can move. Quality video, produced with the same structural discipline, can be expected to perform even better.
- Record a baseline before producing content. Set baseline Share of Voice and Brand Coverage in OtterlyAI for the exact prompt set you intend to target, with the correct country context. Without a baseline, there is no way to attribute change to content.
- Track each AI Search platform as its own surface. The +3% vs +53% spread between AI Overviews and AI Mode in this experiment is one data point in favor of platform-by-platform measurement.
- Track citations and brand mentions separately. Both are forms of AI Search visibility. Decisions about which content to produce next depend on which of the two is moving.
Caveats
Three points to keep in mind when reading this result.
- YouTube policy is tightening on AI slop according to The Verge. Algorithms and enforcement change; nothing here is guaranteed to last, even rankings/AI visibility.
- Short window. Roughly 10 days of post-intervention data. Some of the lift may decay as competitors produce similar content or as platform retrieval logic shifts. Continued monitoring is required to confirm durability.
- Single market, single language. The German B2B SEO prompt set is a specific competitive environment. Results will differ in competitive English-speaking industries, ecommerce categories, and broader consumer queries where the cited source set is wider and more entrenched.
- AI Search outputs vary. AI Search answers can vary by country, prompt phrasing, and personalization factors. OtterlyAI monitors prompts per country and on a recurring cadence so that results are comparable across runs.
Conclusion AI Video Experiment
The headline numbers in this experiment are a set of Share of Voice changes in one market over one week. The more transferable observation is the platform-by-platform variation: six AI Search surfaces produced six different responses to the same content intervention. Measuring each platform on its own, tracking mentions and citations as separate signals, and recording a baseline before any content change are the standard conditions for reading a result like this one.
Hannes and the TRYSEO team are continuing to monitor the channel. Follow-up data will be shared as the observation window extends past the initial 10 days and we will share the latest findings with the AI Search community.
Got a GEO experiment idea? Let’s collaborate!
At OtterlyAI, we don’t just write about AI Search. We test it! With the community, we run and update a public GEO Experimentation Tracker with what we’re running and what moved AI Search. 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.
About the author
Hannes Kaltofen founded TRYSEO a B2B SEO and GEO agency for tech SMEs. They are based in Magdeburg, Germany. He has been developing B2B SEO strategies since 2019 and expanded into GEO (Generative Engine Optimization) in 2023, focusing on building converting websites for clients operating in technical niches.
