Every 2026 AEO trend roundup says roughly the same thing: content is going multi-format, and your videos, podcasts and webinars are the next citation surface. The audience half of that claim is measurable and true — the numbers are below. The engine half is where the roundups get vague, because what OpenAI, Google and Perplexity actually publish about their own crawlers tells a less exciting, more useful story: for citation purposes, AI engines are still overwhelmingly text-first, and the citable layer of any video or audio asset is its transcript and the page around it.
This guide covers what the engines’ own documentation does and doesn’t say about non-text formats, why the transcript page is the asset that earns citations, which schema is worth shipping (and which is hygiene at best), and an honest effort-versus-payoff comparison for a small B2B team. It builds on our foundational guide to answer engine optimisation for SaaS; this post is about getting recorded media into that game.
At a glance: No major AI engine documents watching or listening to your media to decide what to cite. OpenAI’s and Perplexity’s crawler documentation describes fetching web pages and says nothing about video or audio ingestion; Google’s video guidance leans on metadata rather than content analysis. A video or podcast episode therefore becomes citable when its transcript lives as real, crawlable HTML on your own domain, wrapped in accurate VideoObject schema and structured like any strong AEO page. Publish the media for humans; publish the transcript page for the engines.
What OpenAI, Google and Perplexity actually document
Start with the primary sources, because this is an area where secondary commentary routinely overstates what the engines do.
OpenAI: crawler docs describe pages, not media
OpenAI’s crawler documentation lists OAI-SearchBot — “used to surface websites in search results in ChatGPT’s search features” — alongside GPTBot (training) and ChatGPT-User (live user-initiated fetches). Nowhere does it describe ingesting video or audio from the open web, transcribing it, or citing its contents. The unit these bots work in is the web page. If your webinar’s insights exist only inside an MP4, they are invisible to the crawler that decides whether your site surfaces in ChatGPT search.
Google: metadata first, and a capable model is not a crawling promise
Google’s video SEO documentation is explicit about where signals come from: video structured data, video sitemaps, quality thumbnails and the text of the hosting page. The current best-practices page doesn’t even mention transcripts or captions as an indexing input. Meanwhile, Google’s Gemini API documentation shows the models genuinely can process video natively — sampling visuals at one frame per second, processing audio, handling videos up to an hour at default media resolution. That capability is real, but it’s what happens when a developer hands a file to the model; it is not documentation that Google’s search or answer pipelines watch your videos at web scale to choose citations. Conflating the two is the most common error in multi-format AEO advice.
Perplexity: search surfacing, nothing on formats
Perplexity documents two agents: PerplexityBot, which exists to “surface and link websites in search results on Perplexity” and is explicitly not used for training foundation models, and Perplexity-User, which fetches pages when a user’s question requires it. As with OpenAI, the documentation is silent on video and audio ingestion — the mechanism it describes for choosing what to link is web-page crawling.
| Engine | What its own documentation says | Practical implication for your media |
|---|---|---|
| OpenAI (ChatGPT search) | OAI-SearchBot surfaces websites in ChatGPT search; no mention of video/audio ingestion anywhere in the crawler docs | Only the HTML around and about your media can be retrieved and cited |
| Google (Search / AI surfaces) | Video guidance relies on structured data, sitemaps, thumbnails and page text; Gemini processes video natively via the API — a developer capability, not a documented citation pipeline | Ship accurate VideoObject markup and put the substance in crawlable text |
| Perplexity | PerplexityBot surfaces and links websites; docs say nothing about media formats | Same as above — the transcript page is what enters the retrieval pool |
Why the transcript is the citable layer
An AI engine deciding what to cite needs extractable passages: self-contained text that answers a question, on a page it can crawl and attribute. A recorded conversation is full of exactly that material — definitions, numbers, contrarian takes, worked examples — but locked in a format no citation pipeline documents reading. Transcribing it moves that material into the one format every pipeline reads.
Three qualifications, and they’re where most teams go wrong:
- Auto-captions are not a transcript strategy. YouTube’s own documentation says automatic captions “might misrepresent the spoken content due to mispronunciations, accents, dialects, or background noise”, and that “you should always review automatic captions and edit any parts that haven’t been properly transcribed”. If the platform generating the captions tells you to review them, don’t build your citation surface on them raw. Speech-to-text is now cheap enough that a corrected transcript costs an editing hour, not a budget line.
- The transcript must live on your domain as real HTML. A transcript trapped in a player widget, loaded by JavaScript on demand, or attached as a PDF is somewhere between hard and impossible for crawlers to use — and even when fetched, it earns the citation for whoever hosts it. Publish it as a normal page: headings, paragraphs, speaker labels.
- A raw transcript is the floor, not the finish. Spoken language is repetitive and unanchored (“like I said before, that thing we discussed…”). The high-payoff version is an edited transcript page with an answer capsule up top, question-shaped headings, tightened quotes and numbers stated in complete sentences — the same discipline as our guide to structuring content for agentic parsing, applied to speech.
If you want AI agents handling the follow-up when those assets generate enquiries, you can Apply For Partnership.
The schema that matters — and the schema that doesn’t (yet)
VideoObject is the workhorse. Google’s structured-data documentation makes three properties required — name, thumbnailUrl and uploadDate — and recommends description, duration and contentUrl (a direct link to the file, which Google prefers) among others. Markup makes video pages eligible for video rich results, and two extensions handle timestamps: Clip declares labelled segments manually; SeekToAction tells Google your deep-link URL pattern so it can identify key moments itself. If a recording answers several distinct questions, labelled clips turn one vague asset into five addressable answers.
The transcript property exists — use it. Schema.org defines transcript on AudioObject and VideoObject. Google doesn’t document rewarding it, so treat it as cheap machine-readable insurance alongside — never instead of — the visible transcript page.
PodcastEpisode is real but unrewarded. Schema.org defines PodcastEpisode and PodcastSeries, but they sit in the layer of newer types still gathering implementation feedback, and Google’s structured data feature gallery — the definitive list of what Search actually supports — contains no podcast feature at all (its only audio-adjacent entry, Speakable, marks up news text for read-aloud, not your media). Video is in that gallery; podcasts are not. Ship PodcastEpisode markup if it’s low-effort in your CMS, but expect no rich result. Your podcast’s citation surface is the episode’s transcript page, same as video.
Effort versus payoff for a small B2B team
Assume a two-person marketing function producing one recorded asset per fortnight. Here’s the honest hierarchy — deliberately free of the invented “X% more citations” figures this niche is littered with, because no engine publishes format-level citation data.
| Approach | Incremental effort | Citation mechanism | Realistic AEO payoff |
|---|---|---|---|
| Video on YouTube, description only | Low | YouTube page text; auto-captions YouTube itself says to review | Low — visibility accrues to youtube.com, not your domain |
| Video + raw transcript dump on your site | Low–medium | Crawlable text on your domain | Modest — passages exist but are repetitive and poorly anchored |
| Video + edited transcript page with VideoObject, headings, answer capsule | Medium | Extractable passages + video rich-result eligibility | The best ratio on this table — one recording, one strong citable page |
| Podcast + show notes + edited transcript per episode | Medium | Transcript page (no Google podcast rich result exists) | Same mechanism as video; payoff scales with how quotable the guests are |
| Webinar rewritten as a standalone article, recording embedded | High | Purpose-written AEO page plus supporting media | Highest ceiling — the article competes as first-class content |
| Audio/video with no text layer at all | Lowest | None documented by any engine | Effectively zero |
Why bother at all? Because the audience genuinely moved
None of the scepticism above means skipping multi-format content — the consumption data is unambiguous. Edison Research’s Infinite Dial 2025, the longest-running annual survey of US digital media behaviour, found 73% of Americans aged 12+ have consumed a podcast in either audio or video format (an estimated 210 million people), 55% are now monthly consumers, and 51% have watched a podcast. Strikingly, Edison Podcast Metrics found YouTube is the service used most often for podcast listening, used by 33% of US weekly podcast listeners. The boundary between “video” and “podcast” is dissolving in the audience’s hands.
So: make the media for the humans — it builds trust, reach and first-hand expertise — and make the transcript page for the machines. The mistake is doing the first without the second and wondering why the citations never come. And when a vendor promises multi-format citation lifts, apply the standard from how to read AEO benchmark reports: if a percentage doesn’t trace to a named study you can open, it doesn’t exist.
A realistic per-recording workflow
- Record once, transcribe immediately, then spend one editing hour: fix names, numbers and mis-hearings; cut filler.
- Publish the transcript as a real page on your domain — answer capsule at the top, question-shaped h2s, speaker labels, the media embedded above the fold.
- Ship the schema: VideoObject with the three required properties plus description, duration and contentUrl; Clip markup for the strongest segments; the transcript property; Article/FAQPage where they fit.
- Extract the derivatives: any complete answer given on the recording becomes an FAQ entry on the relevant hub page; any strong claim becomes a quotable, sourced sentence in a related article.
- Interlink — the transcript page should cite and be cited by your text cluster on the same topic, with brand facts consistent across all of it.
This is exactly the shape of work an autonomous pipeline handles well: transform, structure, distribute, repeat. It’s the same philosophy Zian’s AI sales agents are built on — systematic follow-through on the work humans skip — applied to content operations. If that resonates, Apply For Partnership.
Frequently asked questions
Do AI engines actually watch videos or listen to audio when choosing citations?
Not according to their own documentation. OpenAI’s and Perplexity’s crawler docs describe fetching web pages and say nothing about media ingestion; Google’s video guidance relies on metadata and page text. The models can process media when handed a file — Google’s Gemini API documentation describes native video understanding at 1 frame per second — but a model capability is not a documented citation pipeline. Plan on the text layer doing the citation work.
Are YouTube’s automatic captions good enough to serve as my transcript?
No. YouTube’s own help documentation warns that automatic captions are machine-generated, that quality varies, that they may misrepresent speech due to accents, dialects or background noise, and that you should review and edit them. Use them as a starting draft, then publish a corrected transcript as HTML on your own domain.
Is there podcast schema that Google rewards?
Schema.org defines PodcastEpisode and PodcastSeries, but Google’s structured data feature gallery lists no podcast rich result — Video is supported; podcasts are not. Mark episodes up if it’s cheap, but invest your effort in the episode’s transcript page.
Should a transcript be verbatim or edited?
Edited, clearly labelled as a transcript, and faithful to what was said. Fix transcription errors, remove filler, add headings and speaker labels — but never alter the substance of quotes. An engine citing your page is trusting your fidelity; so is anyone quoted on it.
Does embedding a video make my page more likely to be cited?
Not directly by any documented mechanism. The embed makes the page eligible for video rich results (with VideoObject markup) and improves the human experience. What earns citations is the extractable text on the page — which is why the edited-transcript-plus-schema pattern beats a bare embed every time.
Source-ownership table
Every external claim in this post, its owner, and the URL it was verified at:
| Claim / figure | Owner | Verified at |
|---|---|---|
| OAI-SearchBot purpose quote; no media-ingestion documentation | OpenAI | developers.openai.com/api/docs/bots |
| PerplexityBot / Perplexity-User roles; no media-format documentation | Perplexity | docs.perplexity.ai/guides/bots |
| Video indexing best practices rely on structured data, sitemaps, thumbnails, page text | developers.google.com — video best practices | |
| VideoObject required properties; Clip and SeekToAction | developers.google.com — video structured data | |
| Search feature gallery includes Video; no podcast feature (closest audio-adjacent entry is Speakable, read-aloud news markup) | developers.google.com — feature gallery | |
| Gemini native video processing; 1 FPS sampling; up to 1 hour at default media resolution | ai.google.dev — video understanding | |
| Automatic captions machine-generated, quality varies, review advised (exact quotes above) | YouTube (Google) | support.google.com/youtube/answer/6373554 |
| 73% of Americans 12+ have consumed a podcast (~210M); 55% monthly; 51% watched; YouTube top service at 33% of weekly listeners | Edison Research (The Infinite Dial 2025) | edisonresearch.com/the-infinite-dial-2025 |
| transcript property on AudioObject/VideoObject | Schema.org | schema.org/transcript |
| PodcastEpisode type (newer layer, gathering implementation feedback) | Schema.org | schema.org/PodcastEpisode |