Image Metadata & Alt Text at Catalog Scale

Alt text used to be a compliance line item — something you filled in so a screen reader had something to say and so Google Images had a caption to index. That's still true. But alt text and image metadata now feed a second audience: AI shopping agents and AI search crawlers that can't fully "see" a product photo the way a human does. When Shopify's Brand Agents or an off-site assistant like Perplexity evaluates a product, the alt text and image metadata are part of what it reads to confirm the image matches the claim. Get this wrong across a 5,000-SKU catalog and you have thousands of small, silent gaps in what your agent can vouch for.

Do alt tags matter for AI shopping?

Yes, and the reason is more specific than general SEO advice suggests. Multimodal AI models can look at a product photo, but they still weight text signals heavily when forming a confident answer — alt text, file names, surrounding captions, and structured data all corroborate or contradict what the model infers from pixels. When an assistant is deciding whether to recommend a product, cite it in an AI Overview, or answer a Brand Agent shopper's question ("does this come in a color that isn't black?"), a well-formed alt tag is one more piece of corroborating evidence. A missing or generic one (image1.jpg, alt="product photo") is a gap the agent has to either guess across or decline to answer.

This is the same logic behind structured data for AI shopping: machines don't get credit for facts they can't read reliably. A photo that visually shows a stainless-steel finish is worth less to an agent than a photo captioned "stainless steel finish, side angle" — because the caption removes ambiguity the model would otherwise have to resolve on its own, and ambiguity is exactly what makes an agent hedge or stay silent.

What good shopping alt text actually contains

Generic SEO alt text advice says "describe the image, include your keyword." That produces alt text like "Men's Blue Running Shoes - Buy Running Shoes Online - Brand Name." It's stuffed, it reads badly to a screen reader, and it gives an AI agent nothing useful to corroborate — the keyword repetition looks like manipulation, not description.

Shopping-context alt text does three things in one short sentence: names the product specifically, states the attribute that's actually visible in that image (not every attribute the product has), and gives context when the image shows use or scale. It reads like a caption a careful human would write, because that's exactly the bar — write it the way you'd describe the photo out loud to someone who can't see it and is deciding whether to buy.

Shopping-context alt text

Alt text written to answer a shopper's implicit question about that specific image — what product, which attribute, what context — rather than to maximize keyword density for search ranking.

Two rules keep this consistent at scale. First, never repeat the full product title verbatim across every image of the same product — vary the angle/attribute so each image adds information instead of duplicating the last one. Second, never assert a fact in alt text that isn't visible in the image itself; if the fabric composition isn't visible, don't put it in the alt tag even if it's true — that's what structured attributes and the product description are for.

How to write alt text at scale: template by category and image role

You cannot hand-write alt text for 8,000 images one at a time and hit a deadline. The move is the same one that works for product attribute schemas: stop treating each image as unique and start treating it as an instance of a category × image role pattern. Most catalogs only have five or six recurring image roles per category.

1. 1

   Group images by role, not by product

   Across your catalog, sort every image into a small set of roles: hero, alternate angle, detail/close-up, in-use/lifestyle, size or scale reference, packaging. Most catalogs need no more than six to eight roles total.

2. 2

   Write one template per category × role combination

   For "apparel > hoodies > hero shot," the template might be {color} {material} hoodie, front view. For "kitchen > cookware > detail shot," it might be {material} finish detail on {product_type}. This is a manageable number of templates — usually 30 to 80 for a mid-size catalog — not thousands.

3. 3

   Populate templates from existing structured attributes

   If your product attribute schema already has color, material, and product type as fields, the templates fill themselves. This is the biggest argument for doing attribute enrichment before or alongside image metadata — the same fields do double duty.

4. 4

   Batch-generate, then spot-check by category

   Run the templates across the catalog programmatically or with an AI pass that fills in the visible-specific slot per image. Spot-check a sample from each category rather than reviewing every single output.

5. 5

   Route exceptions to human review

   Flag anything the template can't handle cleanly — patterned or multi-color items, images with ambiguous framing, or hero SKUs where wording quality actually matters to conversion.

What image metadata does AI shopping use

Alt text is the signal most people think of first, but it's one of several. Image metadata splits into three layers, and most catalogs only ever populate the first.

HTML alt attribute

The alt attribute on the <img> tag. This is what screen readers announce, what Google Images indexes, and what's most directly visible to a crawling agent parsing your page's HTML. It's the highest-priority layer to get right because it's the cheapest for any crawler or agent to access — no file download required.

IPTC and EXIF metadata embedded in the file

IPTC (International Press Telecommunications Council) fields — title, description, keywords, creator — are written directly into the image file itself, separate from anything on the webpage. EXIF carries more technical, camera-level data. IPTC matters here because it travels with the file: if your images get pulled into a marketplace feed, a Google Merchant Center listing, or scraped and re-hosted, the IPTC metadata often survives even when the surrounding HTML doesn't. Most DAMs (digital asset managers) and many PIMs can batch-write IPTC title and description fields from the same template logic used for alt text — it's close to free once the templates exist.

File naming and structured data

Descriptive file names (navy-fleece-hoodie-front.jpg rather than IMG_4821.jpg) are a minor but real signal, and they're nearly free to template since they can use the same slots as alt text. Beyond the file itself, ImageObject markup in your page's structured data — caption, and association with the parent Product — gives agents a machine-readable confirmation of which image belongs to which SKU and what it depicts, independent of alt text entirely.

Where automation is safe, and where it isn't

Templated generation gets you 80-90% of a catalog to a good baseline fast, and that's genuinely the right call for the bulk of a large catalog — a correctly-templated alt tag is a large improvement over a missing one, and perfect is not a reasonable bar for SKU number four thousand. But a few situations deserve a human pass before publishing, because the failure mode of a wrong automated caption is worse than the failure mode of a merely generic one.

• Hero SKUs and top-traffic products. These get the most agent queries and the most human eyes — wording quality has outsized conversion impact, so review beats speed here.
• Multi-color, patterned, or textured products. Templates assume one clean attribute per slot. A plaid-and-solid colorblock jacket breaks a simple {color} {material} template; an automated pass will either oversimplify or guess.
• Anything adjacent to a compliance or safety claim. Alt text that implies material composition, certification, or fit accuracy needs a human check — a wrong caption here isn't just a bad customer experience, it's a corroboration problem if an agent repeats it as fact.
• Images with people, especially size/fit context. Automated captioning of people (build, size worn) is exactly where models are least reliable and where getting it wrong reads worst to a shopper.

Governance: keeping alt text correct as the catalog grows

A one-time bulk pass decays the same way attribute enrichment does the moment new products stop following the template. The fix is the same fix as everywhere else in catalog enrichment: make the template a required step in product intake, not a cleanup project you run again in a year.

• Require the category's alt text template and image roles to be assigned before a new product can go live, the same way you'd require a required attribute field.
• Store the visible-specific slot as a lightweight editable field next to each image in your PIM or DAM, so future edits don't require re-running a whole generation pass.
• Re-audit alt text coverage on the same cadence you audit attribute coverage — it's the same kind of gap and it should live on the same checklist.