AI Search Video Answers: How to Get Your Video and Images Shown in AI Overviews, ChatGPT and Perplexity

AI answers increasingly mix text with a video, an image strip or a product thumbnail. None of the major engines publishes a formula for which clip or image appears. What they do publish is how they find and understand media in the first place, and that is the part you control. This guide covers both halves of the multimodal answer, video first and then images, and ends with how to build a visual asset library that keeps both supplied. Where a point is documented, it links the source. Where it is working practice rather than documented behavior, it says so.

What the engines actually say about media in AI answers

Start from the documentation, because most advice on this topic is invented. Google’s page on AI features and your website says there are no additional requirements to appear in AI Overviews or AI Mode, and no special schema.org markup to add. A page has to be indexed and eligible to show a snippet in Search to be a supporting link. The same page recommends supporting your text with high-quality images and videos where it makes sense, and notes that both features may run several related searches across subtopics (what Google calls query fan-out), which can surface a wider set of pages than a classic results page.

ChatGPT search and Perplexity work from their own crawlers. OpenAI’s bot documentation says OAI-SearchBot is used to surface websites in ChatGPT’s search features, and Perplexity’s says PerplexityBot surfaces and links sites in its results. If your robots.txt or firewall blocks them, your watch pages and image pages cannot be cited there no matter how well they are marked up.

What follows from that:

• There is no separate “AI video SEO.” An asset that is not indexable in regular search has no route into Google’s AI answers.
• Text still carries the answer. Most retrieval-augmented systems retrieve text passages and then attach media from the same or a related source. A video with no words on its page gives a text-based retriever very little to match.
• Be wary of precise claims. Figures like “the carousel carries X% of citation weight” or “images go stale after N weeks” are not published by any engine. Measure your own queries instead (see the capture log below).

Part 1: Video

Make the video indexable

Google’s video SEO best practices set the baseline. The ones that block most brand videos:

• A watch page. The video needs a page that is indexed and performs in Search, where the video is embedded and visible, not hidden behind tabs or other elements. A clip that exists only in a social feed has no page of yours to cite.
• A stable thumbnail URL. Without a fetchable thumbnail, the video cannot be shown as a video result.
• A fetchable file. Give Google the video file (contentUrl) or the player URL (embedUrl). Google calls contentUrl the most effective way for it to fetch the video.
• Unique metadata per video. Each video gets its own thumbnail, title and description. Reusing one description across a series makes them indistinguishable.
• A video sitemap if videos are hard to discover by crawling: <loc> is the watch page, with <video:player_loc>, <video:content_loc> and <video:thumbnail_loc> for the media.

Search Console’s video indexing report shows whether Google found a video on each watch page and, if not, why. Check it before you work on anything else.

Add VideoObject structured data

Google’s video structured data docs require only name, thumbnailUrl and uploadDate. Recommended properties include contentUrl, embedUrl, description, duration (ISO 8601), creator, expires and hasPart for clips. Fill the recommended ones too; they are what tells a system what the video is about and how long it runs. Keep the markup consistent with what is visible on the page, which Google’s AI features page also asks for.

{
  "@context": "https://schema.org",
  "@type": "VideoObject",
  "name": "How to clean suede sneakers without water",
  "description": "Brush dry, use a suede eraser on marks, then restore the nap. 45-second demo.",
  "thumbnailUrl": "https://example.com/video/suede-clean.jpg",
  "uploadDate": "2026-10-01",
  "duration": "PT45S",
  "contentUrl": "https://example.com/video/suede-clean.mp4",
  "embedUrl": "https://example.com/embed/suede-clean",
  "hasPart": [
    { "@type": "Clip", "name": "Brush the surface dry",
      "startOffset": 3, "endOffset": 15,
      "url": "https://example.com/guides/suede?t=3" },
    { "@type": "Clip", "name": "Remove marks with a suede eraser",
      "startOffset": 15, "endOffset": 32,
      "url": "https://example.com/guides/suede?t=15" }
  ]
}

There is no transcript requirement in Google’s video structured data documentation. Schema.org does define a transcript property on VideoObject, but the reliable way to make the words count is to put the transcript on the page as visible text (next section).

Mark key moments so a system can point at the answer

If the answer to a question sits at 0:40 of a longer video, a search result can only send someone there if the moment is marked. Google supports three ways to do it:

The 30-second minimum for Clip and SeekToAction matters for short-form brands: a 20-second clip can still be indexed and shown as a video, but it cannot carry key-moment markup. That is fine. A short clip that answers one question end to end does not need moments; a longer explainer that answers several does.

Treat the transcript as the retrieval surface

Google’s video docs do not mention transcripts at all, so this is working practice, not a documented ranking factor. The reasoning is simple: text retrievers match text. A transcript published as visible text on the watch page gives the page indexable words that say exactly what the video says, gives a passage-level retriever self-contained sentences to quote, and makes the video accessible to people who cannot hear it. The same discipline that wins passage-level retrieval on a written page applies to spoken words, split into transcript chunks:

• Say the answer as a complete sentence. “To clean suede, brush it dry first, then use a suede eraser on marks” stands on its own. “So the first thing you wanna do here is…” only makes sense with the picture.
• State the claim, then show it. The spoken claim is what gets matched; the demonstration is what proves it.
• One question per segment. Distinct answers should be distinct, timestampable moments, not one run-on span that braids three tips.
• Publish an edited transcript. Auto-captions drop punctuation and mis-hear product names. Correct them, then upload the corrected captions to YouTube as well.
• Show what you say. If the words claim the product does something, the frames should show it doing that. A voice-over on a static slide gives the visual nothing to confirm.

YouTube and your own page: do both

YouTube videos are widely indexed and frequently shown in Google results, so a YouTube upload with an accurate title, description, chapters and corrected captions is worth doing. But a YouTube page is about the video, not about your product page. The pattern that covers both: upload to YouTube, then embed the same video on the page of your site that answers the question in text, with its own VideoObject markup and the transcript below it. The text answer and the clip then support each other on one URL you own.

What kind of video fits an answer

The queries that pull video into an answer tend to be “how do I,” “does it actually,” and “show me” questions. A useful working rule is one clip per high-intent question: a person states the answer to camera, then the product does the thing. Burn in captions for muted autoplay on social, and still publish the transcript as page text. Keep the presenter consistent across the set so the clips read as one brand; that is a branding choice, not a documented ranking signal.

This is the format AI UGC video is built for. In ppl.studio, you generate photos of an AI Expert with your product, then turn a photo into a talking-head clip with Animate (see the talking-head video guide), or start from the product demo video or creator UGC video workflow. Be clear about what such a clip is: a scripted presenter explaining the product, not a real customer review. More on that below.

Part 2: Images

AI answers show images as a strip next to the text, as product thumbnails, or inline with a cited source; the glossary calls this the inline image carousel. As with video, no engine documents how those images are picked. Google’s image SEO best practices are the documented baseline, and they are specific.

Image SEO basics Google documents

• Use real <img> elements. Google does not index CSS background images. Keep a fallback src when you use srcset or <picture>.
• Put images near the text they illustrate. Google uses the page’s text, captions and titles to understand an image. An image of the product in use belongs next to the paragraph about that use.
• Descriptive filenames. suede-eraser-on-sneaker.jpg, not IMG00023.JPG.
• Alt text that describes the image in context, without keyword stuffing. Good alt text serves screen-reader users first; Google also uses it. Google links its alt text guide for this.
• Name the page’s preferred image with primaryImageOfPage, the main entity’s image, or og:image. Google advises relevant, high-resolution images and avoiding logos, images with text and extreme aspect ratios. A product-in-context shot is a better og:image than a logo or a text banner.
• Reuse the same URL for the same image across pages so it can be cached and recognized as one image.
• An image sitemap for images Google might not find by crawling, including ones on a CDN domain.
• Supported formats: BMP, GIF, JPEG, PNG, WebP, SVG and AVIF, with a file extension that matches the type.

ImageObject markup and image metadata

On product and article pages, the image property of Product or Article is what image-related rich results read, and Google says it is required for those. A list of URLs is valid. Expanding an image into an ImageObject lets you attach rights information, which Google documents in its image metadata guide: contentUrl is required, plus at least one of creator, creditText, copyrightNotice or license; acquireLicensePage is recommended. That markup is needed on every page where the image appears.

{
  "@context": "https://schema.org",
  "@type": "ImageObject",
  "contentUrl": "https://example.com/img/suede-eraser-on-sneaker.jpg",
  "caption": "A suede eraser lifting a scuff from a tan sneaker",
  "creator": { "@type": "Organization", "name": "Example Shoe Care" },
  "creditText": "Example Shoe Care",
  "copyrightNotice": "Example Shoe Care",
  "license": "https://example.com/image-license"
}

The same fields can live inside the image file as IPTC photo metadata (Creator, Credit Line, Copyright Notice, Web Statement of Rights, Licensor URL, Digital Source Type). If the file and the structured data disagree, Google uses the structured data. Google is explicit that none of this guarantees how an image appears in results.

Two habits that cost nothing: write a caption that restates the claim of the paragraph next to it, in image terms, rather than duplicating the alt text; and when the product changes (new packaging, new colorway), replace the images on every page that shows the old version. An answer engine that shows a picture of a product you no longer sell is showing a wrong answer with your name on it.

AI-generated images and video: label them, and never fake evidence

If your library includes AI-generated media, the rules are documented and worth following exactly.

• Label the file. Google’s image metadata guide supports the IPTC DigitalSourceType values trainedAlgorithmicMedia (made by a generative model), compositeWithTrainedAlgorithmicMedia (edited with generative AI), compositeSynthetic and algorithmicMedia. See the IPTC term definition.
• Merchant Center requires it. Google’s Merchant Center policy on AI-generated content says all images created with generative AI must carry metadata marking them as AI-generated, using trainedAlgorithmicMedia, and that existing tags must be kept. Check that your image editor, CDN and compression step do not strip metadata on export.
• C2PA Content Credentials. Google can read C2PA manifests and may surface them in “About this image,” including whether an image was edited with AI tools. Google documents requirements for the signing tool and certificate. Our C2PA provenance post covers how that works.
• Review what the model wrote. Google’s guidance on generative AI content asks you to fact-check AI output, including alt text, metadata and structured data, and warns that generating many pages without adding value can violate its scaled content abuse policy. A thousand near-identical image pages is that pattern.

Labeling is the floor. The bigger issue is what an image is used to claim. AI imagery is fine for showing how a product is used, styled or sized in a scene. It is not evidence. Do not use it as:

• a “real customer” photo or video review, or a testimonial presented as from an actual buyer (see our post on AI content disclosure and FTC rules);
• before-and-after results the product did not produce;
• the product’s exact appearance, finish or packaging when the model has altered it. Generate with your real product photo as the reference, and check every output against the product before publishing.

An answer engine that quotes your page and shows your image is putting both in front of a buyer as fact. Mislabeled or misleading images are a trust problem with buyers before they are a search problem.

Building a visual asset library for AI search

Video and image work both stall the same way: a one-off shoot, published once, never updated, with no record of which page needs what. A visual asset library is the fix: a planned set of images and clips per page, produced and refreshed on a schedule, with metadata handled in the pipeline instead of by hand.

1. Start from the questions, not the shoot

Use the same priority query list your text work targets, so text and visuals are scored against one set. For each query, note which page should answer it and whether the answer is better shown than told. “How to clean suede” wants a clip; “suede eraser vs brush” wants two images side by side; “what is a suede eraser” may need only a clear product photo.

2. Fill an asset matrix per page

List every priority page as a row and each asset role as a column. The empty cells are the production brief.

3. Keep one visual identity

Using the same presenter, setting style and product framing across a page set makes the library recognizable to people and keeps it on brand. No engine documents a “persona consistency” ranking signal, so treat this as brand practice. In ppl.studio a consistent face across photos is what an AI Expert is for; the glossary entry on a persona-locked visual set describes the idea.

4. Bake metadata into the pipeline

• Name files descriptively at export, not after upload.
• Write alt text and a caption per image while you still know why it was made.
• Set IPTC DigitalSourceType on AI-generated files, and confirm your CDN and optimizer preserve it.
• Generate ImageObject/VideoObject markup from the same record that holds the file, so markup and file never drift apart.
• Keep one canonical URL per asset and reuse it wherever the asset appears.

5. Refresh on events, and on a calendar

Replace assets when the product changes, when a page’s text is rewritten, or when your capture log shows a competitor’s image or clip on a query you care about. No engine publishes a “visual freshness window,” so pick a review cadence that matches how fast your category changes, and update dateModified only when the page actually changed.

6. Pair every text brief with a visual brief

When a page goes into the writing queue, its open matrix cells go into the production queue in the same ticket. Visuals made after the page is written tend to end up decorative; visuals planned with the text end up next to the paragraph they prove.

Measuring it: a capture log

Because nothing here is published as a formula, measure it directly. Run your priority queries by hand in AI Overviews or AI Mode, ChatGPT search and Perplexity on a regular schedule, and log one row per query per engine:

• Query, verbatim, and the engine and date.
• Media shown: none, images, video, or both.
• Your asset? The image or video URL if it is yours, and the page it links back to.
• Whose asset otherwise and what kind (product photo, YouTube video, review site, marketplace listing).
• Text cited? Whether your page was cited in the text of the answer at all.
• Asset type: real photo, AI-generated, stock, or a customer’s.

Answers vary between runs and between users, so read the log for patterns over weeks, not single results. Pair it with Search Console’s video indexing report and your own analytics; our post on AI search attribution covers tying citations to traffic and revenue.

Why your page is cited as text but its media never shows

The most common gap: the page is cited, but its video or image is not. Check these in order:

• The video is not indexed. Look at the video indexing report: no thumbnail, video not the main content, hidden behind a tab, or only on a social platform.
• No words on the page match the video. No transcript and a one-line description give a text retriever nothing to connect.
• The answer is buried at 6:20 of a 14-minute video with no chapters or Clip markup.
• The clip narrates but never demonstrates, so the frames do not show what the words claim.
• Images are CSS backgrounds, lazy-loaded without a real src, or far from the relevant text.
• The preferred image is a logo or text banner, or og:image shows something other than the product.
• The images are out of date and show old packaging or a discontinued variant.
• Crawlers are blocked. robots.txt or a firewall rule stops OAI-SearchBot or PerplexityBot, or blocks your image CDN.

Where this sits in the rest of your AI search work

Media sits on top of text. Engines answer in words first and attach images or a clip that support the answer, so the order of work is the same as everywhere in GEO: earn the text citation with clear, self-contained passages; make sure the engines know which entity you are (see brand entity disambiguation); then give each answer page the image and the clip it needs. For product images in shopping answers specifically, the AI shopping assistants playbook covers feeds and product data.

Frequently Asked Questions

Related: passage-level optimization, the GEO citation playbook, implementing llms.txt, and ppl.studio pricing.