Alt text generator: what the tools actually produce, the four ways to run one, and the review step that decides the quality

An alt text generator is a vision model that looks at an image and writes a sentence describing what is in the frame. That is genuinely useful and it is not quite the same job as alt text, because alt text describes what an image is doing on a particular page. The model can see a brown boot. What it lacks is everything outside the frame: that the boot is the oxblood variant, that the page is a size guide, or that the picture is decorative and belongs skipped entirely. There are four ways to run one: a plugin or app inside your content system, an image CDN that generates on upload, a standalone web tool for a handful of images, or a script against a vision API for a library of thousands. All four produce output worth having, and all four need a review step on the pages that matter.

This page covers the tools and the workflow. For how to write the lines yourself, with a weak and a better version for each kind of image, see alt text examples.

What the tool is actually doing

A generator sends your image to a vision model and returns a caption. Knowing that explains everything the tools are good and bad at. The model has the pixels and nothing else: not your product catalogue, not the page the image sits on, not your naming conventions, and not the reason the image is there.

So a generator is strong at describing ordinary things in a photograph and weak at every decision that depends on context. The most important of those decisions is the one nobody expects a tool to get wrong: whether this image should have a description at all. A decorative texture behind a heading should carry an empty alt so assistive technology skips it. A generator will dutifully describe it as a blue gradient pattern, and every screen reader user then hears that on every page.

The four ways to run one

Route	Fits	The catch
A plugin or app in your content system	WordPress, Shopify and most hosted platforms. Writes straight into the alt field and usually has a bulk mode for the existing library	Pricing is normally per image or per credit, so a large library is a real cost. Check whether it overwrites alt text you have already written
An image CDN or asset manager with a generation feature	Ongoing work. New uploads get a description automatically, so the backlog stops growing	It fixes the future and not the past, and the output lives in the asset system rather than in your page markup unless it is wired up to write back
A standalone web tool	Ten images for a new page. Paste or upload, copy the line out, edit it, done	Nothing is bulk and nothing is saved. Fine as a drafting aid, useless as a process
A script against a vision API	Thousands of images, and anybody with a developer afternoon available	You own the prompt, the cost per image and the output, which is the point. You also own the review step, which is where the quality comes from

For most businesses the honest answer is a combination: the CDN or plugin route to stop the backlog growing, one scripted pass to clear what is already published, and hands on the keyboard for the twenty pages that matter most.

The prompt that makes the output usable

If you are running the API route, the prompt is the whole product. The difference between unusable and useful output is whether you pass the page context alongside the image. A working shape:

Write alt text for this image. It appears on a page titled "[page title]" about "[subject]". The nearby text reads: "[the sentence before the image]". Describe what the image shows in under 15 words, in plain English. Do not start with "image of" or "picture of". Do not list keywords. If the image appears purely decorative, return an empty string.

Four details in there carry most of the weight. The page title and nearby text give the model the context it otherwise lacks. The length instruction keeps it inside what screen readers read comfortably. The instruction about "image of" removes the most common fault in generated output. And the empty string instruction is the only way a tool can handle a decorative image correctly, which it will still get wrong sometimes.

Where generated alt text is reliably wrong

• Charts and graphs. The model describes a picture of a chart. It does not read your numbers reliably, and the numbers are the content. Write these by hand and put the figures in the page as text.
• Product variants. It will say brown, red or dark. It will not say oxblood, Burnt Sienna or your SKU, and on a gallery of variants that distinction is the only information the alt text needs to carry.
• Text inside an image. Prices, opening hours and offers set in a graphic get missed or garbled. Any text in an image has to be repeated in the alt text exactly.
• People. It describes a man in a blue shirt. The useful version names the person and their role.
• Decorative images. Covered above, and worth checking as a category before you publish a batch.
• Specialist images. Technical diagrams, medical images, legal documents, machinery. The description reads plausibly and lacks the one detail a reader in that field needs.
• Linked images. When the image is the only thing inside a link, the alt text has to describe the destination, and the destination lives in your markup rather than in the picture.

A workflow for a library of thousands

• Export the problem first. Crawl the site and pull the list of images with no alt attribute, with duplicate values, or with values over about 125 characters. That list is your scope, and it is usually smaller than people fear because templates repeat.
• Fix templates by hand before you generate anything. The header logo, the navigation icons, the social icons, the default blog thumbnail. Each is one decision that lands on every page.
• Generate a sample of thirty, then count. Read them and count how many you would publish unedited. That percentage is the only number that matters: high, and you can publish the batch and spot check. Low, and generate for the long tail only while you write the top pages yourself.
• Exclude the categories above from the automated pass. Charts, variant galleries and images containing text go on a manual list.
• Never let a tool overwrite existing alt text. Check this setting before the bulk run. A generator that replaces a human line with a generic one makes the site worse while reporting success.
• Add the field to your publishing checklist. Everything above is a one off clean up, and the only way to not need it again is the checklist. The rest of the image work belongs there too, and it is in the nine step image checklist.

Accessibility, stated plainly

A generated description is better than no description for somebody using a screen reader, and it is weaker than one written by a person who knows why the image is on the page. The accessibility standard asks for a text alternative that serves an equivalent purpose, and the word equivalent is doing real work there: it is a judgement about the function of the image, which is exactly the thing a model cannot see. Treat a generated pass as the floor you lift the whole site to, and hand write the pages where somebody is trying to make a decision: product pages, pricing, forms and anything with a chart on it.

Alt text generators: what people ask

Is generated alt text good enough for search?

For ordinary photographs, yes. Image search mainly needs an accurate description of the subject, and that is the one thing these tools do well. The gains that are left on the table are the specifics a model does not know, which are your product names and variants, and those are also the words buyers search for.

Is it good enough for an accessibility requirement?

It satisfies the mechanical part, which is that an alt attribute exists and describes the image. It does not settle the judgement part, which is whether the description serves the same purpose as seeing the picture. For a site under a formal accessibility commitment, use generation to clear the backlog and review the key journeys by hand.

Are there free options?

Yes. Several standalone web tools generate a description for one image at no cost, and some content system plugins include a free monthly allowance. Free tiers are sized for drafting rather than for a library, so they suit a new page rather than a migration.

Does WordPress generate alt text on its own?

No. WordPress gives you an alt text field in the media library and leaves it empty. What it does do, which catches people out, is reuse whatever is in that field everywhere the image is inserted, so one good line fixes many pages and one bad line spreads.

Can a generator handle a whole Shopify catalogue?

Apps exist that will run across the product images in bulk. Before you do, decide what each image type needs: the main product shot, the variant shots and the lifestyle shots are three different descriptions, and a single generated line applied to all of them loses the variant information that makes the gallery useful.

How long should the output be?

The same as any alt text: roughly five to fifteen words, and under about 125 characters. Generators often run long and wordy, so put the limit in the prompt or trim in review.

The same kind of model is reading your page

There is an irony worth sitting with. The model that writes your alt text and the model that answers a buyer's question are the same kind of thing, and the second one arrives at your page with no pixels at all. It gets the alt text, the caption and the words around the image, and from those it decides what your product is. So the line a generator produced, and you did or did not review, is the description of your own product that a machine will repeat to somebody who asked about your category. That makes the review step on your twenty most important pages worth an afternoon. AI Knows Us asks the assistants what they say about a business and reports the answers and the sources they came from.