Every trading day, onvista broadcasts a live market show. Together with Birthe Stuijts at onvista, we spent 107 days testing what happens to AI Search visibility when you stop treating that broadcast as one asset and start treating it as raw material.
The pipeline was simple. Cut the day’s live show into short single-stock videos, publish each one to YouTube, and pair most of them with a written article on onvista.de. Between January and August that produced 610 videos and 571 articles from a show that would otherwise have shipped as 140 uploads.
It worked. Campaign citations per day grew to 3.4x the pre-launch baseline, and onvista’s share of every German-language YouTube citation in AI answers roughly doubled.
Then there is the part we did not plan for. Partway through, onvista started keeping only three of each day’s clips publicly listed on the channel. The rest stayed unlisted: live, working on a direct link, invisible on the channel page. That gave us a control group nobody designed. 118 unlisted videos, zero AI citations, across 107 days and all 7 engines.
Here is the data.
Citations per day indexed to 31 May 2026 = 100%. Video rose to 460%, written articles fell to 10%.
Key Findings (TL;DR)
- Slicing works. Clips cited 3.2x more often than the full show (27.2% versus 8.6%). 7.8x the citations per video.
- Unlist your YouTube videos = zero citations. During the experiment 118 were unlisted from the public YouTube channel. Unlisted videos do not get AI citations.
- The two formats do not compete. Video: Google AI Overviews 56.1%, AI Mode 40.0%. Articles: Copilot 43.0%, ChatGPT 39.2%. Perplexity, Copilot, Claude and Gemini cited 0 of 610 videos.
- Video cost the articles nothing. Articles grew 23% where video cannot reach, flat where video surged 180%.
- Written articles decay if you stop publishing. Article citations track new-article output almost directly (correlation +0.63) and stay active a median of 11 days. Video keeps earning from the back catalogue for months.
- Video peaked during the three-week publishing pause. Zero new clips, 88.2% of all citations. Catalogue, not calendar.
- Changing titles mattered. One word was changed from “Aktie” to “Analyse”: 24.1% to 34.4% cited, +101% per day live.
- Vanity metrics like views predict no AI visibility. Median views: 888 cited, 947 never cited.
Why This Matters for AI Search
Most media brands already produce long video. Webinars, livestreams, panels, daily shows. That footage sits as a single upload, an hour long, one title, one description, one URL for an AI engine to reason about.
The repurposing question is whether cutting it up multiplies your AI Search surface area or just multiplies your workload. It is a real cost question, because slicing and publishing 610 assets is not free.
This experiment answers it with a clean production pipeline, a fixed prompt set, and a natural control group in the middle. And because the unlisted videos were still fully functional on a direct link, it isolates something narrow and useful: how much of AI Search visibility depends purely on an asset being publicly discoverable, separate from its quality, its topic and its audience.
Scope of Study
- 107 days measured, 12 May to 26 August 2026
- 125 tracked prompts with language: German, region: Germany
- 7 AI Search engines: ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot, Gemini, Claude
- 610 YouTube videos + 571 written articles created, published 5 January to 17 August 2026
- 140 broadcast days, producing 140 full daily shows and 470 single-stock clips
- 2,319 campaign citations captured, 1,617 to video and 702 to written articles
- Cross-checked against a full enumeration of 21,136 youtube.com citations and a scrape of all 3,743 public videos on the onvista channel
One citation means one tracked prompt citing one URL, on one day, in one engine. A video cited by two engines on the same day counts twice. We report percentages throughout rather than raw totals, because the prompt set is a sample of the German finance query space and not a census of it.
Methodology: Automating Video + Written Articles
The AI-automated repurposing pipeline, from live broadcast to clip, article and daily citation tracking.
From 1 June the entire production side ran automatically. An AI process cut each broadcast into single-stock clips, generated every title and description, published each clip to YouTube, and drafted the paired written article. Output moved from clips on 36% of broadcast days at a median of 3 per day, to clips on 80% of broadcast days at a median of 10. Nothing in the measurement was automated by onvista: every asset was measured individually rather than by domain. For each of the 1,181 video and article URLs we pulled a daily citation history broken down by engine, then summed.
That last step matters more than it sounds. AI engines cite the same YouTube video under several different URLs: bare, with a &t= timestamp, with a &vl= language parameter, with &utm_source=chatgpt.com appended. The 9 July Nordex clip appears under four separate URLs. Treating those as one video is the difference between counting 1,617 video citations and undercounting by roughly 60%.
Visibility came from onvista’s own channel scrape: 3,743 publicly listed videos as of 27 August. Any video in the production sheet absent from that list is treated as unlisted. That method independently reproduced onvista’s stated rule, which is the strongest validation available: from 13 July onward, every broadcast day has exactly four publicly listed videos, one full show plus three clips. It also confirmed nothing was hidden retroactively.
Engine coverage caveat. Copilot and Claude first return citations in this workspace on 1 and 3 June, and Gemini only on 25 July. Where we report growth over the full window, we use the four engines live throughout.
Results: Does Slicing Used Videos Work?
This is the finding onvista set out to test, and it is the clearest result in the dataset.
Take the same content, the same presenters, the same day’s market analysis. Published as one full broadcast, 8.6% of those uploads ever got cited by an AI engine, at 0.56 citations each. Published as short single-stock clips, 27.2% got cited, at 4.36 citations each.
That is 3.2x the hit rate and 7.8x the citation volume per asset, from re-cutting footage that already existed. Put differently, roughly 1 in 4 clips (27.2%) earned a citation against fewer than 1 in 11 full shows (8.6%).
The mechanism is topic specificity. A one-hour show covering twelve stocks is a weak answer to “Nordex Aktie aktuell Kurs News Analysen”. A five-minute clip titled “09.07.26 Nordex Analyse” is a strong one. AI engines are matching a question to an asset, and a narrower asset wins a narrower question.
If you already produce long video, the cutdowns are the highest-return asset you are not publishing. You do not need new footage, only one clip per topic, each titled for the question it answers.
Before the individual findings, here is the entire production run in one picture.
Of 610 campaign videos, 17.7% were listed and cited, 63.0% were listed but never cited, and 19.3% were unlisted and therefore never cited.
Yes, only 17.7% of the 610 videos were ever cited, and that number needs its two separate causes spelled out, because they are not the same problem.
The first is visibility. 19.3% of the run was unlisted, and as section 3 shows, an unlisted video cannot be cited at all. That share is recoverable with one setting.
The second is topic choice, and it explains most of the 63.0% that were listed but never cited. 249 of the 610 videos cover a stock that none of the 125 tracked prompts asks about, and those were cited 1.6% of the time. Narrow to listed single-stock clips on a stock the prompts do ask about, and close to half (45.8%) were cited, at 7.44 citations each. Topic selection sets the ceiling. The format decides how much of that ceiling you reach.
1. AI Visibility Over Time, Split by AI Search Engines
- Only ChatGPT, Google AI Mode and Google AI Overviews cited the videos.
- Both Google AI Mode and Google AI Overviews grew the most as a result of the automated video + article program.
- Perplexity surprisingly cited only a few of the videos.
- Claude and Gemini mainly only cited the written content.
Share of each day’s campaign citations by engine.
Videos only: Results over time
Only ChatGPT, Google AI Mode and Google AI Overviews cited the videos, with Google AI Mode and Overviews benefiting the most.
Articles only: Results over time
Here we can see that ChatGPT and Copilot benefited the most from the written pieces.
2. Video and Written Cover Completely Different Engines
We expected overlap between the two formats. There is almost none.
Share of each format’s citations by AI Search engine: video is 56.1% Google AI Overviews and 40.0% Google AI Mode, articles are 43.0% Copilot and 39.2% ChatGPT.
onvista’s videos were cited by three engines. Google AI Overviews took more than half (56.1%), Google AI Mode 40.0%, ChatGPT the remaining 3.9%. Perplexity, Copilot, Claude and Gemini cited none of the 610 videos, not once.
96% of video citations came from Google surfaces, with all seven tracked engines listed.
The articles are close to a mirror image. Copilot 43.0%, ChatGPT 39.2%, Claude 6.0%, Google AI Mode 4.1%, Perplexity 3.7%, Google AI Overviews 2.7%, Gemini 1.3%.
82% of article citations came from Copilot and ChatGPT, with all seven tracked engines listed.
Across the 571 topics where onvista published both a video and an article, only about 1 in 10 (10.5%) were cited on both sides. Roughly 1 in 3 (31.5%) earned at least one citation somewhere, against 17.5% for the video alone. Publishing both is close to doubling your odds of being cited at all.
What this means for your GEO strategy: repurposing is not a choice between formats. The video buys you Google surfaces. The written piece buys you everything else. Publishing one and skipping the other leaves whole engines uncovered.
3. Discovery: Unlisted Videos Never Get Cited By AI Search
From 13 July onvista kept three clips per broadcast day publicly listed and set the rest to unlisted. The clips still played. Anyone with the link could watch them. They simply stopped appearing on the channel page, in YouTube search, and in the crawlable link graph.
Citations per day from onvista campaign videos split by YouTube visibility: the unlisted share never rises above zero across the whole period.
Across 107 days and all 7 engines, those 118 unlisted videos collected zero citations.
We wanted to be certain this was visibility and not topic selection or timing, so we built a matched comparison. Take only clips published on 13 July or later, only clips about one of the 100 stocks the tracked prompts actually ask about, and split by visibility. That gives 33 listed against 33 unlisted, from the same broadcast days, the same production process, the same topic universe.
Matched comparison: 33 listed clips reached 42.4% cited, 33 unlisted clips reached 0.0%.
42.4% versus 0.0%. The only variable that differs is one dropdown in YouTube Studio.
Unlisting is not a soft signal that reduces your odds. It removes the asset from the pool an AI Search engine can draw from at all.
What this means for your GEO strategy: if the concern is a cluttered channel page, use playlists, sections or ordering. Those keep the video discoverable. Unlisting is the one setting that guarantees zero.
4. Videos Did Not Cannibalize the Written Content
Since 1 June, video citations per day rose to 460% while article citations fell to 10%. The obvious reading is that the clips stole the citations from the articles. We tested it three ways. It does not hold.
The test that settles it. onvista’s clips only ever appear in Google surfaces. They have zero presence in Copilot and Claude. So take the articles whose paired clip did get cited, and see where those articles lost ground. They lost 48% in Google surfaces, where the clips compete. They lost exactly the same 48% in Copilot and Claude, where no onvista clip exists at all. An identical loss in engines the clips cannot reach is not displacement. Something else is affecting the articles everywhere at once.
The second test says the same thing from the other direction. In the Google surfaces where clips compete, article citations per day went from 3.07 to 3.08. Flat. In Copilot and Claude they grew 23%. Over the same weeks, clip citations in those Google surfaces grew 180%. If clips were pushing articles out, articles would have fallen where clips surged. They did not move.
The third test looked at first like evidence for cannibalisation. Articles whose clip got cited fell 52%, while articles whose clip was never cited rose 206%. That gap disappears once you compare articles that started from the same citation level. Among articles sitting on 2 or 3 citations, the ones with a cited clip fell 92% and the ones without fell 100%. The gap was simply the early top performers falling back to earth, and video had nothing to do with it.
So what did cause the drop? Supply. New article output fell from about 7 per day to under 1 during the August pause, a 95% collapse. Articles stop earning citations when you stop publishing them.
Median days from publish to first citation: 23 for video, 2 for articles. Median days since last citation: 9 for video, 41 for articles.
That is the evidence behind it. Daily article citations track new-article output almost directly, at a correlation of +0.63, and show no relationship at all with video citations, at -0.01. Articles also have a short shelf life: they stay active a median of 11 days, and the median article has now gone 41 days without a citation. For video that figure is 9 days.
What this means for your GEO strategy: written assets are a flow business and video is a stock business. Articles get more competition than videos and may need continuous publishing to hold citations. Video keeps earning from the catalogue. Judge them on different clocks, and do not read a drop in one as the cost of the other.
5. The Citations Kept Growing After The Campaign Paused
From 6 August, onvista published no new clips at all. The citations kept growing for Google AI Mode and AI Overviews.
Campaign citations per day in each phase, indexed to the pre-launch baseline = 100%.
Video citations peaked during that window, at 32.5 per day and 88.2% of all campaign citations, with zero new clips shipped. The existing public catalogue carried the entire result.
This is the most useful thing in the dataset for planning. AI Search is not rewarding your publishing cadence. It is drawing from the set of assets it can reach, and a clip published in June is still answering questions in August. Video citations took a median of 23 days to arrive after publication, so the work you do this month shows up next month regardless of what you ship in between.
6. Renaming the Video Titles Doubled Their Pickup
On 2 July the clip titles changed from the format “05.06.26 Almonty Aktie” to “09.07.26 Nordex Analyse”. Stock became Analysis.
Clips with the old wording were cited 24.1% of the time. Clips with the new wording, 34.4%. Normalized for how long each clip has been live, the newer format earned +101% more citations.
One honest caveat, and we want to be clear about it: onvista changed the written headline format on the same day, so we cannot fully separate the two changes, and July also brought a step change in Google AI Mode coverage. We treat the direction as solid and the exact size as indicative. It is worth an isolated retest.
7. Confirmed: Vanity Metrics like Views Do Not Predict AI Visibility
Share of listed videos ever cited by view bucket: 13.7% under 200 views, 26.9% at 200 to 499, 28.2% at 500 to 1,499, 19.2% above 1,500.
As seen in our YouTube AI Citation Study, vanity/popularity metrics like views, likes, total subscribers are not influencing AI visibility.
- The median lifetime views of cited videos: 888.
- Of never-cited videos: 947.
Effectively identical, and if anything pointing the wrong way. The most-watched bucket, videos above 1,500 views, had the lowest citation rate of any group (19.2%), because it is dominated by older evergreen uploads rather than dated single-stock clips.
Popularity metrics and citability turn out to be separate properties.
How This Compares to Our Previous YouTube AI Citation Study
In our YouTube Citation Study 2026, built on roughly 100 million citations, Perplexity was a meaningful consumer of YouTube content. Here, Perplexity cited zero of onvista’s 610 videos.
From the OtterlyAI YouTube Citation Study: Perplexity cites YouTube most, at 38.7% of YouTube citations across 6 AI Search platforms.
The general lesson is that platform-level benchmarks tell you where a format can work, not where it will work for you. In German-language stock research, Google AI Overviews and AI Mode were the only surfaces that treated a brand’s own video as a citable source at scale. Run the check on your own prompt set before you assume a benchmark transfers.
Was This GEO-Campaign a Success?
onvista’s share of all youtube.com citations in German AI answers rose from 7.0% to a peak of 14.9% for the whole channel, and from 3.6% to 14.1% for the videos created in this experiment.
onvista’s share of every German-language YouTube citation in AI answers went from 7.0% in mid-May to a peak of 14.9% in early August. The videos created in this experiment went from 3.6% to 14.1%, which means that by August the experiment accounted for nearly all of the channel’s AI citation presence.
Campaign assets rose from 7.7% to 14.2% of onvista’s total AI footprint, while onvista’s total footprint held between 2.0% and 3.6% of all tracked citations.
Measured against onvista’s whole AI footprint, every onvista.de citation plus every channel video citation, the campaign assets grew from 7.7% to 14.2%. Yes, it worked.
The honest bound on that: fewer than 1 in 20 (3.95%) of the channel’s 3,743 public videos are ever cited. Repurposing raised onvista’s slice considerably, and we would read the remaining gap as headroom rather than a limit.
Final Conclusion
Across 107 days, 1,181 assets and 7 AI Search engines, this GEO-experiment points to five things you can act on:
1. Repurposing long video/livestreams works, and the gain is large. Re-cutting an existing broadcast into single-stock clips produced 3.2x the citation hit rate and 7.8x the citations per asset. Campaign citations reached 3.4x the pre-launch baseline.
2. Discoverability is binary, not gradual. 118 unlisted videos earned zero citations. Matched on date and topic, listed clips hit 42.4% and unlisted hit 0.0%.
3. Video and written are additive, not competing. They covered almost disjoint engine sets, and article citations grew 23% in the engines video cannot reach while video surged 180% elsewhere. Publishing both roughly doubles the odds of any citation.
4. AI Search reads your catalogue, not your calendar. Video citations peaked during a three-week publishing pause, with a median 23-day lag from publish to first citation.
5. Titles and topic choice are the cheap wins. One word in the title moved citation rate by 10 points. Topic alignment moved it by 44.
Got a GEO experiment idea? Let’s collaborate
This study was a research collaboration with Birthe Stuijts at onvista, who ran the production pipeline and shared the full asset list and channel data that made asset-level measurement possible. Collaborations like this are how we test GEO properly: a real publishing operation, a real prompt set, and results we publish whichever way they come out.
OtterlyAI runs a public GEO Experimentation Tracker showing every experiment we have run and what actually moved AI Search. If you have a hypothesis worth testing on your own content, email rick.tousseyn@otterly.ai and we may run it with you and publish it with your name attributed.
