Image SEO for a Multimodal Search World

Search can see now

For most of my career, image SEO was a footnote: write an alt attribute, compress the file, move on. That era is over. Search can see now. The crawlers reading your pages are no longer parsing text alone; they interpret pixels, extract objects, read text baked into a graphic, and match a photographed thing to a query typed on a phone. Image SEO has quietly become one of the highest-leverage technical disciplines you are probably neglecting, because it is now the difference between an image that surfaces in visual and multimodal results and one that stays invisible.

I have spent fifteen years moving numbers in large programs, and I will tell you where the money is hiding: on ecommerce and content sites, images are half the page weight, half the user's attention, and almost none of the optimization effort. When machines could only read text, that neglect was survivable. When machines can see, it is a leak.

Why images became a first-class ranking surface

Two shifts happened at once. Answer engines and multimodal models learned to look at an image and understand what is in it, and users learned to search with their camera instead of their keyboard. Point a phone at a lamp, a plant, a pair of boots, and the query is the picture. If your product image is slow, undescribed, and structurally anonymous, you are not in that result set. You never entered it.

A few realities worth internalizing:

• The machine reads three layers. It reads the file (format, size, dimensions), the markup around it (alt text, filename, captions, nearby copy), and the pixels themselves. Optimize one layer and ignore the other two and you are only a third of the way there.
• Visual search is intent, not novelty. Someone photographing an object is usually far down the funnel. They want to identify, compare, or buy. That is the same high-intent behavior driving voice and visual search beyond the text box, and it converts.
• Speed is a ranking input, not a nicety. An image the crawler times out on is an image that does not get seen, which is why the business case for speed runs straight through your media library.

The IMAGE checklist

Here is the framework I hand teams. I call it IMAGE, one letter per layer of work, and it is deliberately concrete so you can audit against it today.

I: Intent-named files. The filename is a ranking signal and a free one. bramble-terrier-hiking-harness.webp tells a machine what the picture is; IMG_4821.jpg tells it nothing. Name files for the thing depicted and the job the searcher is doing, before you upload them, because renaming later is a migration nobody schedules.

M: Modern formats, sized right. Serve next-generation formats with wide support, generate responsive sizes so a phone never downloads a desktop-sized asset, and set explicit width and height so the layout does not shift while the image loads. This is the single biggest lever on page weight, and page weight is what decides whether the crawler finishes the fetch.

A: Alt text that describes, not stuffs. Alt text exists to describe the image to someone who cannot see it. Write it that way. Say what is in the frame, plainly, in a full phrase. A machine that can see the image is now checking your alt text against the pixels, so a mismatch reads as a spam signal, not a keyword win. Describe the boots, the color, the setting. That is the keyword, honestly earned.

G: Grounded in structure. This is the layer most teams skip and the one that separates a picture from a product. Wrap your images in the right structured data so the machine knows a photo is a product, a recipe, an article's lead image, or a how-to step. Treating schema as the translation layer for machines is exactly how a raw image becomes an eligible result with a price, a rating, and a reason to click.

E: Environment on the page. No image stands alone. The machine reads the caption beneath it, the heading above it, and the body copy around it to confirm what the pixels suggest. Surround the image with the context that supports it. An undescribed product shot dropped into a wall of unrelated text is a machine's worst case; it sees an object it cannot place.

Run every important image on your site against those five letters. The failures cluster, and the fixes are almost always systemic, not one image at a time.

Where the leaks actually are

When I audit a large site for image performance, the damage rarely comes from the hero shots someone agonized over. It comes from the scale surfaces nobody owns:

• Product galleries with sequential filenames and empty alt attributes, multiplied across a catalog. Fix the template and you fix a hundred thousand images at once.
• Lazy loading done wrong, where the crawler never triggers the load and the image effectively does not exist. Lazy loading is correct; a lazy load the machine cannot resolve is not.
• CDN and parameter sprawl, where the same image lives at a dozen URLs and no single version accumulates authority. This is the visual-media version of the duplication problems that plague large catalogs.

That last one matters more for ecommerce than anywhere else, because the product image is the product. When you are competing to be the answer a generative engine surfaces, a clean, described, structured image is table stakes, which is the through-line in optimizing for generative search on ecommerce. The engine cannot recommend what it cannot confidently see.

A useful way to sequence the work: start where images carry commercial weight, then repeat at scale. Product listing and detail templates first, then category and collection pages, then editorial and blog media last. You are not hand-editing photos; you are fixing the systems that stamp out thousands of them. One template correction propagates further than a week of manual alt text, and it holds as new products flow in. Audit the machinery, not the museum pieces, and you turn a library that reads as noise into one the crawler can parse in a single pass.

What good looks like

A well-optimized image is legible to a machine at every layer at once. The filename states the subject. The format is modern and the size is right for the device. The alt text honestly describes the frame and matches the pixels. The structured data declares what kind of thing it is. The surrounding copy confirms all of it. When those agree, the machine is confident, and confidence is what gets surfaced.

None of this requires exotic tooling. It requires treating images as content that has to be understood, not decoration that has to be compressed. That mental shift is the whole game. Practically, the people who touch images most, the merchandisers, the editors, and the developers, need the standard baked into their workflow, not bolted on at the end. When intent-named files and honest alt text are the default at upload, and structured data ships with the template, the catalog stays legible without a heroic cleanup every quarter. The cheapest optimization is the one you never have to redo.

The takeaway

Image SEO stopped being a compression task and became a comprehension task. The crawler can see your images now, and it is judging whether the file, the markup, and the pixels tell one coherent story. Name your files for intent, serve them fast in modern formats, describe them honestly, ground them in structure, and surround them with context. Do that at the template level and you fix the whole catalog, not one photo at a time.

Keep reading: Canonical Tags and the Duplicate Content You Did Not Know You Had.

If you are sitting on a media library that is heavy, anonymous, and invisible to visual search, and you want it turned into a surface that earns, the channel is open by introduction. Bring your product catalog and your page-speed numbers, and we will find where the images are leaking.