Image and video SEO: how pictures and video get found in search, on YouTube and in AI answers

Images and video are found through more doors than text. The same product photo can appear in Google Images, in an image pack on the main results page, in Discover, as the thumbnail beside a text result, in a Google Lens match, in Bing’s image results and in an AI answer that shows pictures. A video can rank on YouTube, in Google’s Video mode, as a video result with key moments, and as a citation in an AI Overview. Each of these has its own rules, and most of the rules are documented. This guide works through them in the order you would fix them on a real site: where media appears, how an image is understood, the image work itself, video on your own pages, YouTube, accessibility, AI labelling, AI answers and measurement.

It is written for two readers at once. An experienced SEO should find the exact requirements, the markup and the failure reasons here, with sources. A business owner should be able to see which parts matter for their site and which they can ignore. If you sell products, read it alongside the ecommerce SEO guide, which covers product images in Merchant Center and product feeds. Most of the crawling, rendering and speed questions here are part of technical SEO, and the markup sits within entities and structured data. If you work with an AI coding agent, the free media skill runs the checks in this guide on a live site: alt text and formats, VideoObject markup, media sitemaps, and captions and transcripts.

Start with a map of the surfaces, because the work differs for each. On Google, videos can appear on the main results page, in Video mode, in Google Images and in Discover.¹ Images appear in Google Images, as thumbnails beside text results, and in Discover, where a large image does much of the work of getting the click.² Google’s image crawler, Googlebot-Image, fetches images for Google Images, Discover, Google Video and every feature in Google Search where images, logos and favicons are shown.³

Visual search is the fastest-moving of these. In October 2024 Google reported that Lens was used for nearly 20 billion visual searches every month, and that Lens queries were one of the fastest growing query types on Search.⁴ Search Console now separates this traffic: the Performance report has a "Web: multimodal" search type for searches made with images, such as with a phone camera, alongside "Web: text-based", "Image", "Video" and "News".⁵ For a shop, that means a photo of a product taken in the street can lead to your product page, provided your images are indexed and your page clearly describes what is in them.

Bing runs its own image and video verticals, and its webmaster guidelines now tie media to Copilot as well. In Bing’s words, images and video "should reinforce the primary text on the page and should not be the sole source of information required to understand the topic", and clear, accessible multimodal content improves grounding accuracy and citation quality across Bing and Copilot.⁶ Bing Webmaster Tools reports Images and Videos as separate traffic sources in its Search Performance report.⁷

YouTube is the other big search engine for video, and it shows up inside Google and inside AI answers too. Two vendor studies, measured differently, put YouTube at the top of the domains cited in Google’s AI Overviews. Ahrefs’ Brand Radar data for September 2026, covering over 3 million US queries, gave YouTube a 22.9% "mention share", which Ahrefs defines as a domain’s citations as a percentage of the summed citations of the top 50 most-cited sources.⁸ BrightEdge, using its own platform data from May 2024 to September 2025, reported a 29.5% figure for YouTube in AI Overviews and an average citation share of 20% across the AI platforms it tracked, with Vimeo and TikTok at 0.1% each.⁹ The BrightEdge page describes its 29.5% figure in two different ways, as a citation share and as the share of AI Overviews citing YouTube, so treat the exact number loosely. Both studies point the same way: if your subject is something people learn by watching, a YouTube presence is part of search visibility.

There is a commercial caveat that affects how much effort to spend. Image and video surfaces often keep the user inside the platform. SparkToro’s 2024 zero-click study, built on Datos clickstream data from tens of millions of US and EU panellists between September 2022 and May 2024, found that almost 30% of clicks from Google searches went to platforms Google owns, including YouTube and Google Images.¹⁰ A click to Google Images is not a visit to your site. Image traffic is worth most where the image itself sells something: products, places, recipes, interiors, designs, anything people choose by looking.

An image search result has two halves: the image file, and the page it sits on. Search engines need to fetch both, and most of what they understand about the image comes from the page. Google extracts information about an image’s subject from the content of the page, including captions and image titles, and it is worth placing images near relevant text on pages that are about the image’s subject.¹¹

Use the img element, and make sure the src is there when the page renders

Google finds images in the src attribute of the img element, including when the img sits inside a picture element. It does not index CSS images.¹¹ That one rule explains a common audit finding: a hero image set with background-image in CSS looks the same to a visitor but does not exist as an image for search. If an image matters for search, put it in an img element.

Lazy loading is fine as long as the real URL ends up in the src when the image is in the viewport. Google Search does not scroll or click, so lazy loading has to work without user actions: the browser’s built-in lazy loading, IntersectionObserver or a library that loads on visibility all qualify.¹² To check, inspect the page with the URL Inspection tool in Search Console and look at the rendered HTML. If your image or video URLs appear in the src attribute of the img or video elements there, the setup works.¹² A script that keeps the real URL in data-src until a scroll event fires will fail this test. Checking what crawlers see on a JavaScript site covers rendered HTML checks in more depth.

Supported formats and crawl access

Google Images supports BMP, GIF, JPEG, PNG, WebP, SVG and AVIF, and you can also inline images as data URIs.¹¹ The image file has to be fetchable by Googlebot-Image. If your images sit on a CDN subdomain, that host’s robots.txt applies, not your main site’s, and it is common to find a CDN robots.txt that disallows everything. Images hosted on another domain can still be listed in your image sitemap, as long as you verify both domains in Search Console.¹³

Controlling how images appear, or keeping them out

Three controls cover most needs. The noimageindex robots rule tells Google not to index images on the page. The max-image-preview rule sets the largest preview Google can show for the page: none, standard or large.¹⁴ Large previews matter for Discover, where large images are enabled by max-image-preview:large or by AMP.² If you want to stop Google Images from showing your full-size image inline, you can check the HTTP referrer and, when the request comes from a Google domain, reply with a 200 or a 204 and no content.¹¹

<meta name="robots" content="max-image-preview:large">

X-Robots-Tag: noimageindex

Use noimageindex sparingly. Blocking images from search rarely protects them from copying, and it removes them from Lens and Discover as well. For photographers and image libraries, the licence metadata in the next chapters is usually the better route.

Alt text exists for people who cannot see the image: screen reader users, people with images turned off, and anyone on a connection where the image fails. It is also the most important attribute for giving search engines information about an image, and the two goals point the same way.¹¹ The guidance from the accessibility side and the search side agrees on almost everything. Google uses alt text together with computer vision and the content of the page to work out what an image is about, and alt text on a linked image works as anchor text.¹¹ Describe what the image shows, briefly, in the context of the page. Google’s own example sets "Dalmatian puppy playing fetch" against a stuffed string of puppy keywords.¹¹

When to write alt text and when to leave it empty

Not every image needs a description. WCAG success criterion 1.1.1, at level A, requires a text alternative for non-text content that serves the equivalent purpose, and makes an exception for decoration.¹⁵ An empty alt attribute (alt="") tells assistive technology to skip the image, which is right for decoration and spacer images. Leaving the attribute out altogether is worse: some screen readers then announce the file name instead.¹⁶ The W3C’s alt decision tree works through the cases in order.¹⁷

• Does the image contain text? If the same text is on the page as real text, or the text is only a visual effect, use an empty alt. If the text appears nowhere else, put it in the alt.¹⁷
• Is the image a link or a button, and would the control be hard to understand without it? Then the alt describes the destination or the action: "Basket", "Search", "Download the price list (PDF)".¹⁷
• Does it add meaning to the page? A photo or simple graphic gets a short description. A chart or diagram gets its information elsewhere on the page as well, because a sentence of alt text cannot hold a dataset.¹⁷
• Is it decorative, or redundant with nearby text? Use an empty alt.¹⁷

The web is still behind on the basics. In the HTTP Archive’s 2024 Web Almanac, which parsed more than 10 million pages, 45% of img elements had no alt text, and the accessibility chapter found that many of the rest only contained filenames or other meaningless short strings.¹⁸ For most sites the alt text audit is the cheapest image improvement there is.

Filenames

A filename can give Google "very light clues" about what an image shows.¹¹ "my-new-black-kitten.jpg" is better than "IMG00023.JPG", and if you localise a site into other languages, translate the filenames too, following the URL encoding rules for non-Latin characters.¹¹ I would not rename an existing library of indexed images just for this: changing an image URL means search engines have to find and index the new one, and the gain is small. Get filenames right in the upload process for new images.

Captions, nearby text and the page title

Captions are read by more people than body copy, and they sit right next to the image, which is the position that matters for understanding it.¹¹ A caption should add information the image cannot carry alone: who or what it shows, where, when, and why it is on the page. The figure and figcaption elements make the relationship explicit in HTML. Schema.org also has an embeddedTextCaption property on ImageObject for text that is part of the image itself, such as the words on a meme.¹⁹

Social sharing images have their own alt field. The Open Graph protocol defines og:image:alt as a description of what is in the image (not a caption), and says a page that specifies og:image should specify og:image:alt too.²⁰

Images are often the largest contributor to page size, which can make pages slow and expensive to load.¹¹ They are also usually the element that decides Largest Contentful Paint. In the HTTP Archive’s 2025 Web Almanac, 76% of mobile pages and 85.3% of desktop pages had an image as their LCP element, and 62% of mobile sites had a good LCP.²¹ Image work is therefore performance work, and both affect how a page does in search.

Choosing a format

The modern formats compress much better than the old ones. MDN’s image format guide puts lossy WebP at 25 to 35% smaller than JPEG at similar visual quality, lossless WebP at about 26% smaller than PNG, and lossy AVIF at around 50% smaller than JPEG.²² Both are supported in Chrome, Edge, Firefox, Opera and Safari, but MDN still advises a fallback for older browsers through the picture element.²² There are trade-offs. AVIF does not support progressive rendering, so the whole file has to arrive before anything shows, which can make WebP the better choice for very large images.²² Screenshots, diagrams, logos and line art suit lossless encoding, and diagrams and icons that have to draw sharply at every size belong in SVG.²²

Adoption is still slow. Of the image requests the 2024 Web Almanac counted, 32.4% were JPEG, 28.4% PNG, 16.8% GIF, 12% WebP, 6.4% SVG and 1% AVIF, although AVIF use had grown almost fourfold in two years.¹⁸

Responsive images that search engines can still read

Responsive images let the browser pick a file sized for the screen. The srcset attribute offers a list of candidate files with their widths, and sizes tells the browser how wide the image will be displayed.¹⁶ The picture element lets you offer different formats or crops, and the HTML standard requires an img element inside it as the fallback.²³ Google supports both, and asks for a fallback URL in the src attribute of the img.¹¹ The width and height attributes let the browser work out the aspect ratio before the image loads, which reserves the space and prevents layout shift.¹⁶

<picture>
  <source type="image/avif"
    srcset="/img/oak-bedside-table-800.avif 800w, /img/oak-bedside-table-1600.avif 1600w"
    sizes="(max-width: 800px) 100vw, 800px">
  <source type="image/webp"
    srcset="/img/oak-bedside-table-800.webp 800w, /img/oak-bedside-table-1600.webp 1600w"
    sizes="(max-width: 800px) 100vw, 800px">
  <img src="/img/oak-bedside-table-1600.jpg"
    alt="Oak bedside table with two drawers and brass handles"
    width="1600" height="1200"
    fetchpriority="high">
</picture>

The 2024 Web Almanac found srcset on 42% of mobile pages, and found that many sizes attributes were wrong enough to make the browser pick a larger file than it needed: on desktop, one in five.¹⁸ A sizes value that says 100vw for an image that displays at a third of the screen wastes the whole point of srcset. Check it against the layout at a few breakpoints.

Lazy loading, fetchpriority and the LCP image

loading="lazy" defers an image until it nears the viewport, which saves bandwidth on long pages.¹⁶ Applied to the main image at the top of the page, it delays the very thing the user is waiting for. The 2025 Web Almanac found that about 16 to 17% of pages still lazy-loaded their LCP image, and that 62% of mobile pages failed to set dimensions on at least one image.²¹ The fix is simple: never lazy-load the image that is likely to be the LCP element, and mark it with fetchpriority="high" so the browser fetches it early.²⁴ In 2025, 17% of mobile pages with an LCP image used fetchpriority="high" on it.²¹

If you want an agent to do the diagnosis, the performance skill works from field data first and breaks LCP into its sub-parts, which is where image problems show up.

Image sitemaps

An image sitemap helps Google discover images it might otherwise miss, such as images loaded by JavaScript. The format has been cut back to two tags: image:image, which wraps each image, and image:loc, its URL. The image:caption, image:geo_location, image:title and image:license tags have been removed from Google’s documentation, and each url entry can list up to 1,000 images.¹³ If an older plugin still writes captions and titles into your image sitemap, they do no harm, but they do nothing either, so put that information on the page.

<?xml version="1.0" encoding="UTF-8"?>
<urlset xmlns="http://www.sitemaps.org/schemas/sitemap/0.9"
        xmlns:image="http://www.google.com/schemas/sitemap-image/1.1">
  <url>
    <loc>https://www.example.com/bedside-tables/oak-two-drawer</loc>
    <image:image>
      <image:loc>https://cdn.example.com/img/oak-bedside-table-1600.jpg</image:loc>
    </image:image>
    <image:image>
      <image:loc>https://cdn.example.com/img/oak-bedside-table-drawer-open-1600.jpg</image:loc>
    </image:image>
  </url>
</urlset>

For most sites, an image sitemap is a backstop. If your images are in img elements with real src URLs on crawlable, indexable pages, Google will find them through the pages. The sites that benefit most are those with images behind JavaScript galleries, on a separate CDN host, or in such numbers that crawling every page takes a long time.

Telling search engines which image represents the page

A page can have dozens of images, and the one shown beside its search result or in Discover may not be the one you would choose. You can state a preference with the schema.org primaryImageOfPage property or with an og:image meta tag. Pick an image that is relevant and representative of the page, and avoid generic images such as a site logo, images with text in them, and images with extreme aspect ratios.¹¹ Schema.org defines primaryImageOfPage on WebPage as "the main image on the page", and ImageObject has a representativeOfPage flag for the same purpose.¹⁹ For Discover, the large image should be at least 1200 pixels wide, have more than 300,000 pixels in total and a 16:9 aspect ratio, and should not be a logo or a text-heavy graphic.²

<meta property="og:image" content="https://cdn.example.com/img/oak-bedside-table-1600.jpg">
<meta property="og:image:alt" content="Oak bedside table with two drawers and brass handles">

These are preferences. Search engines can still pick a different image if they judge it more relevant to the query, and I have not seen any published test that measures how often the preference is followed. Set it anyway, because the default choice is sometimes a banner or a logo.

Licence metadata and the Licensable badge

If you license images, Google Images can show a Licensable badge that links to the licence and to details of how to use the image.²⁵ You can supply the information as structured data or as IPTC photo metadata embedded in the file. In structured data, an ImageObject needs a license URL and an acquireLicensePage URL, plus at least one of creator, creditText or copyrightNotice.²⁵ In IPTC, the equivalent fields are Web Statement of Rights for the licence and Licensor URL for the licence page, with Creator, Credit Line and Copyright Notice for attribution.²⁵

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "ImageObject",
  "contentUrl": "https://cdn.example.com/photos/thames-at-dawn.jpg",
  "license": "https://www.example.com/image-licence",
  "acquireLicensePage": "https://www.example.com/licensing/thames-at-dawn",
  "creditText": "Example Photo Library",
  "creator": {
    "@type": "Person",
    "name": "Sam Taylor"
  },
  "copyrightNotice": "Example Photo Library"
}
</script>

IPTC metadata has the advantage of travelling with the file. The IPTC Photo Metadata Standard is the industry standard for administrative, descriptive and copyright information about images, used by photographers, news organisations and archives.²⁶ Its current version, 2025.1, defines Copyright Notice as the notice "for claiming the intellectual property for this photograph", and also includes accessibility fields: Alt Text (Accessibility), a brief description of the purpose and meaning of an image for assistive technology, and Extended Description (Accessibility) for more detail.²⁷ A content management system that reads the Alt Text field on upload can pre-fill the alt attribute, which saves time when a photographer has already written it. The catch is that many image pipelines strip metadata to save bytes. If you rely on embedded licence or source information, check the files your CDN actually serves.

Video on your own site follows stricter rules than images, and the rules tightened in 2023. In April 2023 Google stopped showing video thumbnails next to text results when the video was not the main content of the page. In December 2023 it extended the change to Video mode, so that clicking a result there only leads to pages where the video is the main content.²⁸ The post gave examples of pages where video is supplementary: a blog post where the video complements the text, a product details page with a complementary video, and a video category page that lists several videos of equal prominence.²⁸ It also warned site owners to expect fewer pages with indexed videos and fewer video impressions in Search Console.²⁸

What a watch page is

The current documentation uses the term watch page: a page where watching an individual video is the main reason the user is visiting. A video landing page, a TV episode player and a sports highlight page are examples. The video must be embedded on that page and must not be hidden behind other elements, and the watch page itself has to be indexed and performing well in Search before its video can be considered.¹ In practice, a watch page has the player near the top, visible without clicking anything, with the title, description and transcript of that one video around it.

This is a decision to make page by page. A product page with a demonstration clip is still a good product page, and it can rank as a text result. It will not be a video result, and the video indexing report will say so. If a video deserves its own visibility, give it its own page and link to that page from the product or article. If it is there to support the page, accept the report entry and move on.

Why videos are not indexed: the report reasons

Search Console’s video indexing report gives a reason for each video it found but did not index. The reasons map almost one to one onto fixes.²⁹

Removing or restricting a video

To take a video out of Google, return a 404 for the watch page, add a noindex robots meta tag, or set an expiration date in the structured data. To limit where it can appear, use regionsAllowed or ineligibleRegion in the markup, or video:restriction in a video sitemap.¹ The max-video-preview robots rule caps the length of the moving preview Google can show, where 0 allows only a static image and -1 sets no limit.¹⁴

Markup removes the guesswork about a video’s details. VideoObject structured data has three required properties: name, thumbnailUrl and uploadDate, with the date in ISO 8601 format. contentUrl or embedUrl are recommended, and providing at least one of them is strongly advised, along with description and duration.³⁰ The name and description should be unique for each video on the site.³⁰ Schema.org also defines a transcript property on VideoObject for the text of what is said.³¹

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "VideoObject",
  "name": "How to fit a bedside table drawer runner",
  "description": "A four-minute walkthrough of removing, levelling and refitting a drawer runner on a two-drawer oak bedside table.",
  "thumbnailUrl": "https://cdn.example.com/video/drawer-runner/thumb-1280x720.jpg",
  "uploadDate": "2026-09-14T09:00:00+01:00",
  "duration": "PT4M12S",
  "contentUrl": "https://cdn.example.com/video/drawer-runner/drawer-runner.mp4",
  "embedUrl": "https://www.example.com/embed/drawer-runner",
  "transcript": "In this video we will take the drawer out, remove the old runner..."
}
</script>

Keep the markup consistent with the page and with everything else that describes the video. The title in the markup, the visible heading, the sitemap entry and the YouTube title (if the same video is also on YouTube) do not need to be identical, but they should describe the same thing. Mismatched records are a common reason for the wrong thumbnail or the wrong date appearing.

Thumbnails

Every video needs a valid thumbnail at a stable URL. Use a single, unique and stable thumbnail URL for each video, in BMP, GIF, JPEG, PNG, WebP, SVG or AVIF, at least 60 by 30 pixels, with at least 80% of the pixels having an alpha value greater than 250.¹ That last rule rules out mostly transparent PNGs. Thumbnail URLs that change on every request, as some video platforms generate with signed parameters, break indexing because the URL Google stored no longer works.

Video sitemaps and feeds

A video sitemap is the other way to describe videos, and it suits sites with a large library. Each video entry needs video:thumbnail_loc, video:title, video:description (up to 2,048 characters) and either video:content_loc or video:player_loc. Optional tags include duration, expiration date, publication date, restriction, platform, live, and up to 32 tags. video:category, video:gallery_loc, video:price and video:tvshow have been deprecated.³²

<?xml version="1.0" encoding="UTF-8"?>
<urlset xmlns="http://www.sitemaps.org/schemas/sitemap/0.9"
        xmlns:video="http://www.google.com/schemas/sitemap-video/1.1">
  <url>
    <loc>https://www.example.com/videos/fit-a-drawer-runner</loc>
    <video:video>
      <video:thumbnail_loc>https://cdn.example.com/video/drawer-runner/thumb-1280x720.jpg</video:thumbnail_loc>
      <video:title>How to fit a bedside table drawer runner</video:title>
      <video:description>A four-minute walkthrough of removing, levelling and refitting a drawer runner on a two-drawer oak bedside table.</video:description>
      <video:content_loc>https://cdn.example.com/video/drawer-runner/drawer-runner.mp4</video:content_loc>
      <video:duration>252</video:duration>
      <video:publication_date>2026-09-14T09:00:00+01:00</video:publication_date>
    </video:video>
  </url>
</urlset>

Bing takes a slightly different line. It supports the video extension to the sitemap protocol, but says you get the best results with a Media RSS (mRSS) feed, submitted through the Sitemaps feature in Bing Webmaster Tools. Each video in the feed must include a title, the play page URL, the publish date, a description, the thumbnail URL and the raw video URL or player URL.³³ If Bing video traffic matters to you, a feed is worth the extra build. Google accepts mRSS too, but its indexing report suggests switching to schema.org markup when mRSS processing fails.²⁹

For live streams, BroadcastEvent markup nested in the VideoObject can show a red LIVE badge in results. It takes isLiveBroadcast, a startDate and, once the stream has finished, an endDate.³⁰

Key moments let a video result link straight to a point in the video. There are three ways to get them. For videos on YouTube, timestamps in the description can mark the moments. For videos on your own site, Clip markup gives the exact start and end of each segment, and SeekToAction markup tells Google how your URLs link to a time in the video so it can identify the moments automatically.¹

Clip markup

Each Clip needs a name, a startOffset in seconds from the start of the video, and a url that opens the video at that point. The total video must be at least 30 seconds long, and no two clips on the same video can share a start time.³⁰

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "VideoObject",
  "name": "How to fit a bedside table drawer runner",
  "description": "A four-minute walkthrough of removing, levelling and refitting a drawer runner.",
  "thumbnailUrl": "https://cdn.example.com/video/drawer-runner/thumb-1280x720.jpg",
  "uploadDate": "2026-09-14T09:00:00+01:00",
  "duration": "PT4M12S",
  "contentUrl": "https://cdn.example.com/video/drawer-runner/drawer-runner.mp4",
  "hasPart": [
    {
      "@type": "Clip",
      "name": "Taking the drawer out",
      "startOffset": 0,
      "endOffset": 48,
      "url": "https://www.example.com/videos/fit-a-drawer-runner?t=0"
    },
    {
      "@type": "Clip",
      "name": "Removing the old runner",
      "startOffset": 48,
      "endOffset": 131,
      "url": "https://www.example.com/videos/fit-a-drawer-runner?t=48"
    },
    {
      "@type": "Clip",
      "name": "Levelling and fitting the new runner",
      "startOffset": 131,
      "endOffset": 252,
      "url": "https://www.example.com/videos/fit-a-drawer-runner?t=131"
    }
  ]
}
</script>

SeekToAction markup

SeekToAction suits large libraries where writing clips by hand is not practical. You give a URL template with a {seek_to_second_number} placeholder and declare startOffset-input as "required name=seek_to_second_number", and Google picks the moments itself.³⁰ It only works if your player really does start at the time in the URL, so test the template in a browser first. Automatic identification is limited to certain languages: English, Spanish, Portuguese, Italian, Chinese, French, Japanese, German, Turkish, Korean, Dutch and Russian.³⁰

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "VideoObject",
  "name": "How to fit a bedside table drawer runner",
  "thumbnailUrl": "https://cdn.example.com/video/drawer-runner/thumb-1280x720.jpg",
  "uploadDate": "2026-09-14T09:00:00+01:00",
  "contentUrl": "https://cdn.example.com/video/drawer-runner/drawer-runner.mp4",
  "potentialAction": {
    "@type": "SeekToAction",
    "target": "https://www.example.com/videos/fit-a-drawer-runner?t={seek_to_second_number}",
    "startOffset-input": "required name=seek_to_second_number"
  }
}
</script>

YouTube chapters

On YouTube, chapters come from timestamps in the description. The first timestamp must be 00:00, there must be at least three timestamps in ascending order, and each chapter must be at least 10 seconds long.³⁴ YouTube can also generate chapters automatically for eligible videos, and manual chapters override the automatic ones.³⁴ Write the chapter titles as a viewer would search for the moment: "Removing the old runner" is more useful than "Part 2". The same timestamps can then serve as key moments in Google.¹

00:00 Taking the drawer out
00:48 Removing the old runner
02:11 Levelling and fitting the new runner
03:40 Checking the drawer runs smoothly

Most businesses that publish video put it on YouTube first, and many then embed it on their own site. That raises an early decision: which page should rank as the video result in Google, the YouTube watch page or yours? If the video is embedded on your own dedicated watch page with structured data, that page can be eligible for video features in its own right; Google still advises markup for watch pages that use a third-party player, and asks you to check that the host lets Google fetch the video file.¹ In practice the YouTube page often wins for the same video, because it is the more established watch page. I would not fight that for every video. Choose a small number of videos where a visit to your site is worth more than a view, give those a proper watch page on your site, and let YouTube carry the rest.

How YouTube search ranks videos

YouTube’s help centre names three main elements for search: relevance, engagement and quality, weighted differently by type of search. For relevance it looks at how well the title, tags, description and video content match the query. For engagement it looks at signals such as the watch time of a video for a particular query. For quality it tries to identify channels that show expertise, authoritativeness and trustworthiness on a topic. Results are also personalised by search and watch history where that is turned on.³⁵

Two details from YouTube’s own documentation cut through a lot of folklore. Tags "play a minimal role" in discovery and are mainly useful when the subject is commonly misspelt.³⁶ And the "video content" in the relevance list means what is said and shown, which is one reason accurate captions are worth the effort. Titles can be up to 100 characters and descriptions up to 5,000.³⁶ Put the words a viewer would search for early in the title, write the first lines of the description as a plain summary of what the video covers, and add chapters.

Recommendations: Home, Up Next and the Shorts feed

For most channels more views come from recommendations than from search, and recommendations follow different signals. YouTube’s guidance for creators says its surfaces are tuned to different viewer goals: search looks for relevance to keyword queries and at which videos drove the most engagement for a query, while the Shorts feed may lean on recency.³⁷ The system also learns from context such as time of day and device type.³⁷ Several points in the same guidance answer questions creators ask all the time:

• There is no ideal video length. YouTube advises the length the content needs, without filler, and points creators to audience retention curves to see where viewers drop off.³⁷
• One weak video does not penalise a channel. Recommendations respond to how viewers react to each video; longer-term channel performance can suffer if viewers consistently stop watching a channel’s videos when they are recommended.³⁷
• Changing a title or thumbnail can shift performance up or down, because viewers interact with the new version differently. The systems respond to those interactions, not to the act of editing. YouTube advises against changing what already works.³⁷
• Monetisation does not affect recommendations, and uploading as unlisted before making a video public should not significantly affect its performance.³⁷

Thumbnails and titles on YouTube

A custom thumbnail needs a verified account. YouTube asks for JPG or PNG with a minimum width of 640 pixels and a 16:9 aspect ratio for standard videos, and thumbnails must follow the Community Guidelines.³⁸ YouTube Studio’s Test & compare feature can test up to three titles, three thumbnails, or both together on a long-form video. It judges the winner on watch time share, not click-through rate, on the reasoning that a title and thumbnail that set the right expectations reduce misclicks.³⁹ That design choice is worth copying in your own judgement: a thumbnail that wins clicks but loses viewers in the first seconds is not a win. Changing the title or thumbnail during a test stops it, and Shorts are not eligible.³⁹

Captions are the place where accessibility and search line up most clearly. A caption file is a timed text version of the speech and the non-speech audio a viewer needs, shown in the player. The W3C uses "captions" for the same language as the audio and "subtitles" for translations, and notes that captions and transcripts contain the same text, so one can be used to make the other.⁴⁰ Search engines and AI systems work mostly with text, so a transcript on the page gives them the content of the video in a form they can read and quote. Bing’s guidelines make the same point from the other side: media should reinforce the text on the page, never be the only place the information exists.⁶

What WCAG 2.2 requires for media

In the UK, public sector websites must meet WCAG 2.2 AA, with some exemptions, including prerecorded audio and video published before 23 September 2020 and live audio and video. Every UK service provider, public or private, also has a duty to make reasonable adjustments for disabled people under the Equality Act 2010 (the Disability Discrimination Act 1995 in Northern Ireland).⁴¹ Other countries have their own laws, so check the rules for each market you serve. For a business, uncaptioned marketing video is an accessibility gap before it is an SEO gap.

Automatic captions are a first draft

YouTube generates automatic captions with speech recognition, in the video’s default language only. You can also upload a caption file, type captions yourself, or upload a transcript and let YouTube sync the timings.⁴² The W3C is blunt about the automatic version: automatically generated captions do not meet user needs or accessibility requirements unless they are confirmed to be fully accurate, and they usually need significant editing.⁴⁰ Product names, place names and technical terms are where they fail most, and those are often the words people search for. Correct them, then publish the corrected file on YouTube and the transcript on your own watch page.

For video you host yourself, WebVTT is the most common caption format on the web, alongside SRT and TTML.⁴⁰ A WebVTT file is plain text: a WEBVTT header, then cues, each a start and end time followed by the text to show.⁴³ It is added to the player with a track element.

WEBVTT

00:00:00.000 --> 00:00:04.500
In this video we'll take the drawer out
and remove the old runner.

00:00:04.500 --> 00:00:07.000
[Drill whirs]

00:00:07.000 --> 00:00:11.200
Start by pulling the drawer fully open
and lifting it at the front.

<video controls width="1280" height="720"
       poster="/video/drawer-runner/thumb-1280x720.jpg">
  <source src="/video/drawer-runner/drawer-runner.mp4" type="video/mp4">
  <track kind="captions" src="/video/drawer-runner/drawer-runner.en.vtt"
         srclang="en" label="English" default>
</video>

Autoplaying video deserves a word here because it is so common. The 2024 Web Almanac found the autoplay attribute on 23% of video elements and the controls attribute on only 9%.¹⁸ Background video without controls can fail WCAG 2.2.2 if it moves for more than five seconds with no way to pause it, and 1.4.2 if it plays sound.¹⁵ Give it a pause control.

Generated images are now ordinary on commercial sites, and the platforms have started to ask for them to be labelled. There are two complementary standards, and the work for a site owner is mostly about not destroying the labels that tools already add.

IPTC Digital Source Type

The IPTC photo metadata standard has a Digital Source Type field that records how an image was made, using a controlled vocabulary.²⁷ The terms that matter here are trainedAlgorithmicMedia, labelled "Created using Generative AI"; compositeWithTrainedAlgorithmicMedia, labelled "Edited using Generative AI" and covering inpainting and outpainting; algorithmicMedia for images made by an algorithm without training data; and digitalCapture for images captured from real life with a camera.⁴⁴

Google Merchant Center makes this a requirement for shopping. All images created with generative AI must carry the IPTC DigitalSourceType value TrainedAlgorithmicMedia, and merchants must not remove that embedded metadata. AI-generated titles and descriptions go in separate structured_title and structured_description attributes.⁴⁵ If you sell through Merchant Center, the ecommerce guide covers the feed side.

# Read the field
exiftool -XMP-iptcExt:DigitalSourceType image.jpg

# Set it on a generated image
exiftool -XMP-iptcExt:DigitalSourceType="http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia" image.jpg

C2PA Content Credentials

C2PA is a provenance standard. A C2PA manifest, which its specification calls Content Credentials, bundles assertions about an asset with a digitally signed claim, and it can be bound to the content with a cryptographic hash (a hard binding) or with a fingerprint or invisible watermark (a soft binding). The c2pa.created action can record the digital source type using the IPTC vocabulary.⁴⁶ The specification is careful about what it does not do: it should not judge whether a set of provenance data is "good" or "bad", only whether the assertions in it can be validated.⁴⁶ Content Credentials record where an image came from and what was done to it. They do not certify that it is true.

Google said in September 2024 that if an image contains C2PA metadata, people would be able to use the "About this image" feature in Search to see whether it was created or edited with AI tools.⁴⁷ YouTube may automatically label a video as AI-generated or altered if it contains C2PA metadata, and it requires creators to disclose realistic content made or meaningfully altered with AI, such as a real person appearing to say something they did not, altered footage of a real event, or a realistic scene that did not happen. The disclosure is a setting in YouTube Studio, and YouTube says it will not limit a video’s audience or its eligibility to earn money.⁴⁸

AI answers show images and cite videos, and almost none of the selection is documented. Google’s page on AI features says there are no additional requirements to appear in AI Overviews or AI Mode beyond being indexed and eligible to show with a snippet. Its list of SEO basics that still apply includes making sure important content is available in textual form and supporting that text with high-quality images and videos where relevant.⁴⁹ Bing’s guidelines, quoted earlier, say clear and accessible multimodal content improves grounding accuracy and citation quality in Bing and Copilot.⁶ Neither says how an image is chosen to sit beside an answer.

What can be said with evidence is narrower. The YouTube citation studies earlier in this guide show that video pages are cited heavily in AI Overviews, but they measure citations of YouTube URLs, not why those videos were chosen.⁸ My reading, which is judgement and not a measured finding, is that the text around a video does the work: the title, description, chapters and transcript give a retrieval system something to match against a question. A video with an accurate transcript and descriptive chapters is easier to cite than one with a vague title and no words. The same logic applies to your own pages. If a fact only exists inside an image or a video, assume an AI system will miss it. The AI search guide covers how assistants choose sources in more detail, and measuring visibility in AI search covers tracking it.

• Put the key information in text on the page, with images and video supporting it. This is the one point both Google and Bing document for AI features.⁴⁹
• Publish a corrected transcript on the watch page and keep the YouTube captions accurate, so the spoken content exists as text.
• Use chapters with descriptive titles. They mark the passages a question is most likely to match.
• Keep alt text and captions specific. An image described only as "hero image" gives a retrieval system nothing.
• Do not expect a dedicated report. AI Overviews and AI Mode traffic is counted inside the Web search type in Search Console, not separately.⁴⁹

Each surface reports in a different place, and none of them gives the whole picture. Set up the reports before you change anything, so you have a baseline to compare against.

Google Search Console

• Performance report, search type "Image": traffic from the Google Images tab.⁵ Compare it month on month, and filter by page to see which templates earn image clicks.
• Performance report, search type "Video": traffic from the Videos tab.⁵ The 2023 changes cut impressions here for pages where video was not the main content, so read older comparisons with that in mind.²⁸
• Performance report, search type "Web: multimodal": searches made with an image, such as Lens.⁵ For shops and visual products, this is the closest thing to a visual search report.
• Video indexing report: indexed and not indexed videos, with the reasons in the table above.²⁹
• Discover report, if your site appears there. Discover depends heavily on large images.²

Search Console has also started to report on content you publish on other platforms. Platform properties let a creator or publisher add a YouTube channel, or an Instagram, TikTok or X account, as its own property and see how that content performs on Google Search, Discover and News. The feature was rolling out gradually when I checked. It only covers performance in Google, not views on the platform itself, and a click that opens a video in Google’s own viewer still counts as a click.⁵⁰ For a business with a YouTube channel, this finally shows which videos people find through Google and with what queries.

Bing Webmaster Tools

Bing’s Search Performance report added Images and Videos as traffic sources, alongside Web, Chat, News and Knowledge Panel, from 24 March 2023. For videos, a click is only counted if it sends the user to your site.⁷

YouTube Analytics

YouTube Studio’s Reach tab splits traffic sources into internal sources, such as browse features, suggested videos, playlists and YouTube search, and external sources, meaning websites and apps that embed or link to the video, plus direct or unknown traffic. It also reports thumbnail impressions and the impressions click-through rate: how often viewers watched after seeing a thumbnail.⁵¹ Read search and recommendation traffic separately, because the levers differ: titles, descriptions and spoken content for search, the first seconds and audience retention for recommendations.

One honest limit: there is no reliable way to measure how often your images appear beside AI answers, or which of your videos are cited by assistants other than through sampling the answers yourself or using a third-party tracker. Treat any single number from those tools as a sample, and compare like with like over time.