---
title: Stop Writing Alt Text by Hand. Build a Vision Model Pipeline With a Human Gate.
description: Vision models can draft alt text for every image on your site, but only if you feed them page context, classify decorative images first, and gate risky outputs behind human review. Here is the pipeline we recommend.
author: LETSGROW Dev Team
date: 2026-11-12
category: AI Tools
tags: ["Accessibility", "Alt Text", "Vision Models", "AI Automation", "Image SEO"]
url: "https://letsgrow.dev/blog/ai-alt-text-pipeline-vision-model-accessibility"
---
Most marketing sites treat alt text as a checkbox an SEO plugin nags about. That is why so many of them ship thousands of images with an empty `alt`, a filename like `hero-final-v3.jpg`, or the same keyword-stuffed sentence pasted onto every product shot. Vision models have made fixing this cheap enough that there is no longer an excuse. But the teams getting value from it are not the ones clicking "auto-generate" in a CMS plugin. They are the ones treating alt text as a pipeline: context in, structured output out, a human gate in the middle, and a CI check that stops new debt from shipping.

Our position is simple. Let the model write the first draft of every alt attribute on your site. Never let it publish without context about the page, and never let it decide whether an image is decorative on its own.

## Why "Just Run It Through a Vision Model" Produces Bad Alt Text

A vision model looking at an image in isolation describes what it sees. That sounds right, and it is the core problem. Alt text is not a caption. It is a replacement for the image that conveys the same purpose in that specific spot on the page.

The same photo of a person typing on a laptop needs different alt text depending on where it lives. On a case study it might need to say who the customer is. On a blog hero it is probably decorative and should have an empty `alt=""`. Inside a product walkthrough it needs to name the screen and the action being shown. A model that only sees pixels will write "A person typing on a laptop at a wooden desk" in all three places, which is accurate and useless.

The other failure modes are predictable once you run a few hundred images through any model:

- **Hallucinated text in screenshots.** Models will confidently misread small UI labels, chart values, and numbers. On a pricing graphic or a dashboard screenshot, a wrong number in the alt text is worse than no alt text.
- **Over-description.** Models love to describe lighting, color palettes, and mood. Screen reader users have to listen to all of it.
- **Keyword stuffing by prompt.** If someone adds "include our target keyword" to the prompt, every image becomes a spam signal. Alt text helps image search because it describes images well, not because it repeats your head term.
- **Identifying people.** Do not ask a model to name people in photos. Pass names in from your CMS metadata when it matters.

## The Pipeline That Actually Works

Treat this like any other content operation. Here is the architecture we recommend for a site with a few hundred to a few thousand images.

**1. Inventory with page context.** Crawl your site (Screaming Frog, a Playwright script, or your CMS API) and export every image with its URL, current alt value, the page URL, the nearest heading, and the surrounding paragraph. Also capture whether the image is inside a link or button. A linked image needs alt text that describes the destination, not the picture.

**2. Classify before you describe.** Make the first model call a classification: is this image decorative, informative, functional (inside a link or button), or complex (charts, diagrams, data)? Use structured outputs with a fixed enum so you get a clean value every time. Decorative images get `alt=""` and skip the next step entirely. Complex images get routed to a human, because a chart needs a text alternative that carries the data, and that belongs in the page body, not a 125 character attribute.

**3. Generate with context and constraints.** For informative and functional images, send the image plus the page title, nearest heading, surrounding text, and link destination. Constrain the output: one sentence, no "image of" or "picture of" prefix, no visual styling unless it matters, and a hard character limit around 125. Ask the model to return a confidence field and a flag for any text it read from the image.

**4. Human review on a queue, not a spreadsheet.** Route anything with low confidence, any image containing text or numbers, and any image on a high-traffic template to a reviewer. Everything else can be spot-checked at a sample rate you choose. A reviewer moving through a queue that shows the image, the page, and the draft can approve dozens per minute.

**5. Write back through the CMS API.** Do not hand-edit. Write approved values back programmatically and log the old value, the new value, the model version, and who approved it. If you change prompts later, you will want to know which alt text came from which version.

::compare-table
title: "Three Ways to Fix Alt Text at Scale"
columns: ["Approach", "Speed", "Quality", "Best For"]
rows:
  - ["CMS plugin auto-generate", "Instant", "Low: no page context, no decorative detection", "Tiny sites with no review capacity"]
  - ["Vision model pipeline with review queue", "Hours to days", "High: context-aware, audited, versioned", "Marketing sites with hundreds to thousands of images"]
  - ["Fully manual writing", "Weeks to months", "High but inconsistent across writers", "Small sets of complex charts and diagrams"]
::

## Stop the Debt From Coming Back

A one-time cleanup is wasted if the next campaign launch adds 40 images with empty alt attributes. The fix is to move enforcement upstream.

Add an accessibility check to your build or deploy pipeline. axe-core, run through Playwright or your existing test runner, will fail on images missing an `alt` attribute entirely. That catches the worst case. Then add a lightweight custom rule that flags known bad patterns: alt text that matches the filename, alt text that starts with "image of," alt text over 150 characters, and identical alt text repeated across more than a handful of different images.

In your CMS, make alt text a required field with an explicit "decorative" toggle that sets an empty value on purpose. That single UI change does more than any audit, because it forces a decision at upload time instead of leaving a blank for someone to find later. You can also pre-fill the field with a model draft at upload, so editors are reviewing instead of writing from scratch.

::checklist
title: "Alt Text Pipeline Launch Checklist"
items:
  - "Export every image with page URL, current alt value, nearest heading, surrounding text, and link destination"
  - "Run a classification pass first: decorative, informative, functional, or complex"
  - "Set decorative images to an empty alt attribute instead of generating descriptions"
  - "Route charts, diagrams, and images containing numbers or UI text to a human reviewer"
  - "Generate drafts with page context, a 125 character limit, and a confidence score"
  - "Write approved values back through the CMS API and log model version and approver"
  - "Add axe-core plus a custom bad-pattern rule to CI so new images cannot ship without alt text"
  - "Make alt text a required CMS field with an explicit decorative toggle"
::

## Why This Is Worth Doing Now

Missing alt text has been a WCAG Level A failure under success criterion 1.1.1 for years, so it is not a new requirement. What has changed is the cost of compliance and the cost of ignoring it. The European Accessibility Act has applied to a wide range of consumer-facing digital services since June 2025, and accessibility lawsuits in the US keep targeting the same basic failures. Image alternatives are almost always on the list.

There is also a quieter upside. Descriptive, accurate alt text helps image search understand your visuals, and AI assistants that read pages as text get a better picture of what your product screenshots actually show. Accessibility work that also improves how machines read your site is about as close to free as marketing gets.

The takeaway: spend one sprint building the pipeline, not one quarter writing alt text by hand. Classify first, generate with context, review what is risky, and put a check in CI so you never have to do the cleanup twice.
