Which Content Formats Do AI Search Engines Cite? Evidence From 610 Videos

AI search engines can cite both articles and videos, but the formats do not travel through every answer engine equally. A 107-day onvista publishing study found an unusually sharp split. Google AI Overviews and AI Mode produced nearly all observed video citations. Microsoft Copilot and ChatGPT produced most article citations.

The practical conclusion is not “make more video.” It is that format is a routing decision. Start with a valuable question, create a focused answer, make it publicly discoverable, and choose formats according to the engines you want to reach.

What did the onvista study measure?

German finance publisher onvista converted daily livestreams into topic-specific YouTube clips and written articles. Otterly.AI then measured the campaign from May 12 to August 26, 2026.

The proprietary study covered 107 days, 125 German-language prompts and seven AI engines. Those engines were ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot, Gemini and Claude. The production system created 610 videos and 571 articles between January 5 and August 17.

The 610-video total included 140 full shows and 470 single-stock clips. Otterly counted one citation when one engine cited one normalized URL for one prompt on one day. It reported 2,319 campaign citations: 1,617 to videos and 702 to articles.

These figures come from the campaign operator and its measurement provider. They are valuable first-party observations, but they have not been independently reproduced.

Which AI engines cited video and articles?

The strongest strategic signal was the format split. Of the observed video citations, 56.1% came from Google AI Overviews and 40.0% came from Google AI Mode. ChatGPT supplied the remaining 3.9%. No video citations were recorded from the other four measured engines.

Articles followed a different route. Copilot supplied 43.0% of article citations and ChatGPT supplied 39.2%. Claude, AI Mode, Perplexity, AI Overviews and Gemini divided the remaining 17.8%.

This does not prove an engine-wide preference for one format. The study involved one German finance publisher, one prompt set and one time window. It does show why a single “AI visibility” score can hide a useful operational truth: the same topic may need different assets for different retrieval systems.

Where does format affect the citation pipeline?

An answer engine cannot cite an asset merely because it exists. The asset has to pass through several distinct stages. First, a platform or search system must discover its URL. Next, it must be allowed and able to fetch the page or video metadata. The system then needs enough indexable information to connect that asset with a future question. Finally, a retrieval and answer-generation process decides whether the asset is useful enough to surface and cite.

This distinction prevents three common reporting errors:

1. Discovered is not indexed. A crawler visit does not prove that an asset entered a usable retrieval index.
2. Indexed is not retrieved. An eligible asset can remain absent for a particular prompt.
3. Retrieved is not cited. A system may use information without displaying that URL as a visible source.

Did short clips outperform full livestreams?

Yes, inside this campaign. Otterly reported that 27.2% of topic-specific clips received at least one citation, compared with 8.6% of full shows. That made clips 3.2 times more likely to be cited. Clips averaged 4.36 citations per asset, while full shows averaged 0.56, a 7.8-times difference.

Specificity is a plausible mechanism. A short clip can align its title, spoken content and watch page with one question or entity. A full show contains many subjects, so the relevant passage is harder to identify and retrieve. This explanation is consistent with the results, but the study did not randomly assign identical material to short and long formats.

Google recommends a dedicated, indexable watch page, a stable thumbnail and accessible video metadata. It also says eligibility does not guarantee appearance in a search feature. Those requirements support a simple distinction: an asset must be discoverable before a system can retrieve and cite it.

Does public discovery matter for AI citations?

The campaign provides a strong, though not randomized, discoverability comparison. From July 13, onvista left three clips per day public and marked the others unlisted. All 118 unlisted clips received zero measured AI citations.

Otterly also compared 33 public clips with 33 unlisted clips covering tracked stocks from the same production process. Public clips achieved a 42.4% citation rate; unlisted clips achieved 0%.

YouTube explains the likely discovery gap. Unlisted videos generally do not appear in YouTube Search or on a channel's Videos tab, although people with the link can still watch and share them. Therefore, zero citations in this dataset should not be turned into a universal rule that unlisted video can never surface elsewhere.

The matched comparison reduces some obvious differences, such as production process and time period. It does not remove every possible selection effect. The study does not say that publication status was randomly assigned, and the three public clips may have differed in unmeasured ways.

Does publishing an article and video improve coverage?

Across 571 paired topics, 31.5% received a citation to at least one format. Video citation coverage was 17.5%, while 10.5% of topics earned citations to both the article and video.

That 31.5% versus 17.5% comparison is descriptive. It does not prove that adding an article caused a 14-point lift, because topics were not randomly assigned to paired and video-only strategies. The useful inference is narrower: the two formats reached partly different citation surfaces, so pairing created more possible retrieval paths.

How strong is each conclusion?

A multi-format GEO workflow

A useful strategy starts with questions, not file types. Xtrusio's recommended decision path is:

High-value question → source moment → article plus focused clip → target engines → discoverability check → citation tracking

This approach extends an answer engine optimization workflow beyond article production. It also requires a measurement system that separates discovery from citation. The same principle applies when you measure GEO campaign performance: record evidence at asset and engine level before calculating an aggregate score.

What should publishers do now?

Publishers with webinars, interviews, podcasts or long videos already own raw material. The first step is to map the individual questions answered inside each source. Create a standalone clip only when the segment remains useful outside the full recording. Create a corresponding article when the written answer can add context, evidence or structure.

Keep both assets public and technically discoverable. Use descriptive titles that identify the entity and question. Link the article and video where that helps a person continue learning. Then track which engine cites which URL instead of assuming every format competes in the same pool.

Most importantly, test this model in your own category. Finance content, German prompts and Google's video integrations shaped the onvista results. A software, healthcare or local-services dataset may produce a different engine-format map.

Example: turn one webinar into an engine-specific test

Consider a 45-minute software webinar that answers six customer questions. Publishing the recording alone creates one broad asset. A multi-format test would identify the two questions with the clearest demand and strongest expert answers.

Then define the test before monitoring results. Track the same question across the relevant engines. Preserve the exact prompt, location, date, cited URL and answer screenshot. Compare the clip and article independently. A useful result is not merely “we received more citations.” It identifies which engine surfaced which format, how long discovery took, and whether the paired assets reached different prompts.

Allow enough time for discovery. The onvista study reported a median of two days from article publication to first citation and 23 days for video. That timing difference came from one campaign, but it illustrates why a seven-day review could incorrectly label video a failure. Use a fixed observation window and keep uncited assets in the denominator.

Sources

1. Otterly.AI and onvista multi-format citation experiment, published September 3, 2026.
2. YouTube visibility settings, explaining public, private and unlisted discovery behavior.
3. Google Search Central video SEO guidance, covering indexable watch pages, thumbnails, metadata and video eligibility.