The awkward fact underneath most guidance on this subject: systems generating answers mostly don't look at your images. They read the text around them.
Captions, alt attributes, the paragraph above, the filename, the structured data. That's what describes a visual to anything assembling an answer — which means most of what's sold as "optimising media for AI" is text SEO applied to media, and saying so plainly makes the actual work much clearer.
It also explains a common frustration: a page full of good imagery that contributes nothing to how the page is understood, because none of it was ever described.
Your images are doing three different jobs
Bundled together as "image SEO", and they need different work.
| The outcome you want | What actually drives it |
|---|---|
| Ranking in image search | The visual itself, filename, alt, caption, page relevance |
| Helping the page rank | Relevance to the content, and not slowing the page down |
| Being usable in an AI answer | Almost entirely the text describing it |
| Being the reason someone clicks | Thumbnail quality and distinctiveness in the result |
| Surviving image blocking | Alt text written to stand in for the image |
Most teams do the technical work in row two — compression, formats, lazy loading — and skip rows one, three and five entirely, then conclude that image SEO doesn't do much. The technical work matters, but it's the part that stops images hurting you rather than the part that makes them earn anything.
Alt text is doing three jobs at once
And most alt text is written for one of them, if any.
It's read by assistive technology. It's displayed when the image fails or is blocked. And it's one of the main things telling a search or answer engine what the image contains. Those three uses point in the same direction, which is convenient — write it for a person who can't see the image and you've served all three.
The test that produces good alt text Read the page aloud, substituting your alt text where each image sits. If a sentence stops making sense, or you find yourself saying "image", the alt text isn't doing its job for anyone.
Three rules cover most cases. Describe what the image communicates rather than what it depicts. Give decorative images an empty alt attribute so screen readers skip them rather than announcing a filename. And where an image contains text, repeat that text, since nothing else carries it.
This is the same principle applied in email templates that render everywhere and it sits inside the wider obligations in accessibility requirements marketers should know — one piece of work, three returns, which is unusually good value.
What doesn't work: stuffing keywords. Alt text reading "blue running shoes trainers footwear buy online cheap" fails as a description, fails for accessibility, and is transparently manipulative to anything reading it.
The caption is more valuable than the alt attribute
Underused, because it feels like decoration.
Captions are visible page text sitting immediately adjacent to the image, and they're read by everyone and everything. Where alt text is a fallback, a caption is content — it can carry the detail, the context, the number, the thing the image is evidence of.
Two practical uses. On a product page, the caption is where you put the detail the photo can't state — dimensions, material, what's shown at what scale. And on any page making a claim, a captioned image is how you attach evidence to the claim in a form a reader and a machine both register.
Worth writing captions on the images that matter and skipping them on the ones that don't, rather than applying them uniformly. A caption on a decorative header photo is noise.
Rankmath
Most popular WordPress SEO plugin with powerful on-page optimization features built in
Best for: WordPress SEO Plugin
Video: the transcript is the asset
The single highest-return change available to most sites publishing video.
Search engines and answer engines read text. A video embedded on a page without a transcript is close to invisible to everything except the player — you've published a substantial piece of content that nothing can index, quote or cite.
Publishing the full transcript as readable page text changes that completely. It's indexable, it's quotable in an answer, it's accessible, and it's searchable by the visitor who wants one specific part. Frequently it attracts more search traffic than the video itself.
Three details that separate a useful transcript from a dump. Clean it up — remove filler, fix the automated mis-hearings, especially of product and place names, which is where machine transcription reliably fails. Add headings so it reads as a document rather than a wall. And keep it on the page, not behind an expander that hides it from readers who'd have used it.
Chapters and timestamps are the other underused element: they let both a viewer and a search engine address a specific moment, which turns one long video into several answerable segments.
Where to host video, and what you're really deciding
The choice gets framed as technical. It's a distribution decision.
Hosting on a large video platform gives you access to that platform's own audience and internal search, which is substantial — and increasingly visible in results, as covered in where Google is sending clicks now. But the visibility accrues to the platform page, and a viewer who finds you there is on their property, not yours.
Embedding with the transcript on your own page keeps the indexable text on your domain, which is what gets cited and what carries a link to your other pages.
The answer for most businesses is both, and not as a compromise: publish on the platform for reach, and maintain your own page carrying the full transcript, the context and the internal links. They serve different queries and the duplication costs nothing.
Structured data: what it does and doesn't do
Brief, because expectations here run ahead of reality.
Video structured data helps a page become eligible for video results and gives a search engine explicit metadata — duration, upload date, thumbnail, key moments. Image structured data supports specific features like product and recipe results.
What it doesn't do is make mediocre media rank. It's a way of stating clearly what's already on the page, not a ranking mechanism, and marking up something that isn't genuinely there risks the feature rather than winning it.
Worth adding where you publish video regularly or sell products, and worth skipping where you don't — the same triage logic as the reports covered in the Search Console reports that matter, where the Video indexing report is the one to check if you've done this work.
Speed is still the biggest technical factor
Unglamorous, and it remains the thing most likely to be actively costing you.
Images are usually the heaviest element on a page, which makes them the main lever on loading performance and therefore on the metrics discussed in Core Web Vitals and page speed.
- Modern formats — AVIF or WebP with a fallback — typically cut file size substantially at equivalent quality.
- Serve at the size displayed. A 3000px image rendered at 600px wastes most of what you sent, and it's the most common single fault.
- Lazy load below the fold, load eagerly above it. Lazy loading your main image delays the thing the page is measured on.
- Set width and height attributes so the layout doesn't shift as images arrive.
- Responsive sources, so a phone isn't downloading a desktop asset.
None of this earns rankings. All of it stops you losing them, which is a different and equally important thing.
Original imagery is now a differentiator
For a practical reason rather than an aesthetic one.
Stock images appear on dozens of sites simultaneously, and generated imagery is now abundant enough that it distinguishes nothing. Neither evidences that you've handled the product, visited the place, or done the work.
An original photograph does. A picture of the actual installation, the actual room, the actual product in a real setting is genuinely unique, supports a claim the reader is assessing, and can't be reproduced from existing sources — which is the same durability argument made about data in publishing your own data, arriving in visual form.
It matters most on the pages where somebody is deciding: product pages, service pages, case studies. A generic hero photo on a blog post costs little; a stock photo standing in for the work you actually do costs a conversion — the argument in product page optimisation, and for retail specifically it feeds directly into the asset quality that drives shopping feed performance too. If imagery is the bottleneck across a catalogue, that's an e-commerce production problem before it's an SEO one.
What actually helps in AI answers
Pulling the thread together, since this is what the newer guidance claims to address and mostly doesn't.
If systems read the text around images rather than the images themselves, then media optimisation for AI answers means:
- Describe every meaningful image in page text — caption, adjacent paragraph, or both. An undescribed image contributes nothing.
- Publish transcripts, because that's how video becomes quotable.
- State the specifics in text, not only in the graphic. A chart with the key figure only inside the image is invisible; the same figure in a sentence beneath it is citable.
- Keep it accessible without JavaScript execution where you can, since crawlers vary in what they render.
That last set is the practical core of generative engine optimisation and being cited by machines, applied specifically to media — and it's consistent with what's actually working in AI Overviews.
The queries where the image is the answer
Worth ending here, because it's where visual work has the most durable value.
Some questions can't be answered in text. What does this look like in a small room, does this colour work with that one, how worn is too worn, what should this joint look like when it's done correctly. A generated summary can describe; it can't substitute for looking.
Those queries are the part of search that automation replicates least well, which is the same structural argument made in why discovery platforms are winning budget again, and it's visible in the traffic patterns in where AI search traffic actually lands.
Practically: if your category involves people looking before deciding, invest in the visual answer to the questions they're actually asking rather than in decorative imagery. A comparison photo, a scale reference, a before-and-after, a detail shot of the thing people worry about. Those earn links, get saved, and answer a query no summary closes.
The short version
Systems generating answers mostly read the text around your images rather than the images themselves, so most media optimisation for AI is a text exercise. Separate the three jobs your images do, because ranking in image search, supporting the page, and being usable in an answer need different work — and most teams only do the technical layer, which stops images hurting you rather than making them earn anything. Write alt text for a person who can't see the image and you serve accessibility, image search and image blocking at once. Use captions more, since they're visible content sitting beside the visual. For video, publish the cleaned-up transcript as page text, because that's the asset — the video is close to invisible without it. Host on a platform and on your own page; they answer different queries. Keep the technical hygiene, since speed stops you losing rankings. And use original photography where somebody is deciding, because stock evidences nothing.
The most valuable visual work is the answer that can't be written down — and that's the part of search a summary will never close.
Pages full of images that describe nothing?
We audit your media across alt text, captions, transcripts and load speed, then fix the commercial pages first.
Audit my media →Frequently asked questions
Can AI answer engines actually see your images?
Mostly they work from the text around an image rather than from the image itself. Captions, alt attributes, nearby paragraphs, filenames and structured data are what describe a visual to a system building an answer. That makes the majority of what gets sold as image optimisation for AI a text exercise applied to media — and it means an unlabelled image on an otherwise well-written page contributes almost nothing to how that page is understood.
What are the different jobs an image does for SEO?
Three, and they need different work. Ranking in image search depends on the visual itself plus surrounding signals. Supporting the host page's ranking depends mainly on relevance and on not slowing the page down. Being usable in an AI-generated answer depends almost entirely on the text describing the image. Treating these as one task is why media optimisation often produces no measurable result — effort goes into the wrong layer for the outcome wanted.
Do you need a transcript for video SEO?
For anything beyond the hosting platform's own search, the transcript is effectively the asset. Search engines and answer engines read text, so a video embedded without a transcript on the page is close to invisible outside the player. Publishing the transcript as readable page text — not only as a caption file — makes the content indexable, quotable and accessible at once, and it frequently attracts more search traffic than the video does.
Should you host video yourself or use YouTube?
It is really a decision about where you want discovery and citation to land. Hosting on a large video platform gives access to that platform's own audience and search, but the visibility accrues to the platform page. Self-hosting or embedding with the transcript on your own page keeps the indexable text on your domain. Many teams do both — platform for reach, own page with full transcript for search and citation — which is usually the right answer rather than a compromise.
Is original photography worth it compared with stock imagery?
Increasingly yes, for a practical reason rather than an aesthetic one. Stock images appear on many sites simultaneously and generated imagery is now abundant, so neither distinguishes a page or evidences first-hand experience. An original photograph of the actual product, place or process is genuinely unique, supports claims a reader is assessing, and is one of the few visual assets that cannot be reproduced from existing sources.
