AI Alt Text: How Alt Text Generators Work and When to Use Them

Writing alt text by hand is slow. A product catalog with 5,000 images needs 5,000 descriptions. An archive of ten years of blog posts may hold tens of thousands of images with empty alt attributes. This is the gap that AI alt text fills: a vision model looks at an image and writes a description for you.

This guide is the automation companion to our complete alt text writing guide. That guide covers what good alt text sounds like and how to write it. This guide covers how to generate it at scale, where generators fail, and how to build a workflow that produces usable results instead of spam.

What an AI Alt Text Generator Does

Every AI alt text generator uses the same core technology: a vision-language model. The model receives your image as input. It outputs a sentence that describes what it sees.

The pipeline looks like this:

1. Your CMS or tool uploads or references the image file.
2. The image goes to a vision model such as GPT-4V, Claude, or a smaller specialized model.
3. The model returns a text description, usually one sentence.
4. The tool writes that sentence into the alt attribute of your <img> tag.

The quality of step 3 depends almost entirely on the prompt. Ask a model to “describe this image” and you get a generic caption. Ask it to “write alt text under 125 characters for a screen reader user, focused on the informative content” and you get something much closer to what a human would write.

That distinction matters. Generic captions describe pictures. Good alt text serves a purpose on a page. We cover that purpose in depth in our screen reader accessibility guide.

Where Alt Text Generators Exist Today

You can get AI alt text from four kinds of sources. Each fits a different workflow.

CMS Built-ins

Most major platforms now ship auto alt text features:

Platform	Feature	How it works
WordPress	Plugins such as Alt Text AI	Generates alt on upload or in bulk for existing media
Shopify	Magic alt text	Generates descriptions for product images
Squarespace	AI image descriptions	Offers generated alt during upload
Wix	Auto alt text	Fills alt for media library items

CMS built-ins win on convenience. There is nothing to install beyond a plugin or toggle. They lose on control. Most give you one description per image with no way to tune length, focus, or tone. If the output misses your page context, your only option is to edit it by hand anyway.

Dedicated Alt Text Tools

A second tier of standalone tools focuses only on alt text. You upload images or point them at a URL, and they return descriptions through a web UI or API. These tools typically offer:

• Batch processing for entire media libraries
• Length and style controls
• Export back to your CMS
• Human review queues

They cost more than a free plugin, but they treat alt text as a workflow rather than a checkbox.

CDN and API Generation

Some image platforms generate alt text at the infrastructure level. Sirv’s AI Studio includes an Alt Text tool alongside background removal and other AI editing features. You connect it to sirv.studio and generate descriptions across your hosted image library without touching your CMS.

This approach has a structural advantage: the platform already knows your images. It can process thousands of files in bulk, and the alt text travels with the image wherever the CDN serves it. For teams running galleries through something like the Sirv Media Viewer, generated alt text flows into every zoomable, spinnable asset automatically.

DIY With Vision APIs

If you have development resources, calling a vision API directly gives maximum control. Here is a working prompt pattern for GPT-4V-class models:

import base64
import openai

def generate_alt_text(image_path: str, page_context: str) -> str:
    with open(image_path, "rb") as f:
        image_data = base64.b64encode(f.read()).decode()

    response = openai.chat.completions.create(
        model="gpt-4o",
        messages=[
            {
                "role": "user",
                "content": [
                    {
                        "type": "text",
                        "text": (
                            "Write alt text for this image. Rules:\n"
                            "- Under 125 characters\n"
                            "- Describe the informative content, not the aesthetics\n"
                            "- No phrases like 'image of' or 'picture of'\n"
                            f"- Page context: {page_context}\n"
                            "- Return only the alt text, nothing else"
                        ),
                    },
                    {
                        "type": "image_url",
                        "image_url": {
                            "url": f"data:image/jpeg;base64,{image_data}"
                        },
                    },
                ],
            }
        ],
        max_tokens=100,
    )
    return response.choices[0].message.content.strip()

The page_context parameter is the part most tools skip. We explain why it matters next.

The Accuracy Reality Check

AI alt text fails in predictable ways. Know these before you trust any generator.

Vision Models Hallucinate Details

Vision models sometimes state things that are not in the image. Common examples:

• Inventing brand names for unbranded products
• Guessing emotions or relationships between people
• Naming specific locations, models, or species with false confidence
• Describing objects that appear only in training data patterns

A hallucinated detail in alt text is worse than missing alt text. A screen reader user hears wrong information and trusts it. Never publish generated alt text for images where factual precision matters — news photos, medical diagrams, legal evidence, technical documentation.

Models Miss Context and Purpose

Alt text is not a photo caption. The same image needs different alt text on different pages:

Page	Same hero image	Correct alt text
Product page	Woman hiking in boots	“Trailhawk waterproof hiking boots, side view”
Blog post about ankle support	Woman hiking in boots	“Hiker on rocky trail showing boot ankle support”
Category page	Woman hiking in boots	“” (decorative on this layout)

A generator sees pixels. It cannot know which page the image sits on or why the page exists. This is the single biggest limitation of every auto alt text tool on the market. The DIY prompt above partially solves it because you pass page context in. No fully automatic tool can.

Models Miss Text in Complex Images

Charts, infographics, and screenshots break generators. WCAG expects charts to convey their data trend or takeaway, not a list of every bar. Models often produce either too little (“a bar chart”) or an unusable data dump. Charts and infographics need human writing — see the formula in our writing guide.

When Automation Is Fine

Auto alt text earns its place in these situations:

Large archives that need a baseline. A site with 40,000 images missing alt text will not get human descriptions this quarter. Generated baseline alt text beats empty alt attributes today. Mark low-confidence entries and improve them over time.

Product catalogs with structured data. If your PIM or inventory system already holds the product name, color, and category, combine those fields with a generated visual description. “Nike Air Max 90 in white, shown at three-quarter angle” comes from structured data plus a vision pass. Neither source alone gets there.

Decorative-adjacent imagery. Stock-style photos used for atmosphere need only generic descriptions or empty alt. Generators handle these well because precision does not matter.

First-pass drafts for human editors. Generation as a drafting step cuts writing time roughly in half in typical editorial workflows. The editor edits instead of starting from zero.

When Human Writing Wins

Keep humans on these image types:

• Informative images — diagrams, process illustrations, annotated screenshots
• Charts and data graphics — the alt must state the takeaway
• Branded hero messaging — the alt should carry the campaign message, not “smiling people in office”
• Images whose meaning depends on surrounding copy
• Any image where a wrong description causes harm

Our guide on common image accessibility mistakes lists the failure patterns that show up most in real audits. Auto-generated keyword stuffing appears near the top.

The SEO Angle: Generated Alt Text Can Hurt You

Search engines read alt text. Some site owners respond by prompting generators to stuff keywords into every description. This backfires.

Keyword-stuffed alt text is a long-recognized spam signal. Google’s guidance describes alt text stuffing as a negative experience for screen reader users and a reason to distrust a page. Our breakdown of image SEO mistakes covers this pattern in detail.

Rules that keep generated alt text safe:

1. Prompt for descriptions, never for keywords.
2. Reject any output above roughly 125 characters.
3. Ban phrases like “best cheap running shoes buy online” from your review checklist.
4. Treat the primary keyword as acceptable only when it genuinely names the image subject.

There is also an upside worth noting. Search engines increasingly parse images directly with their own vision models, and AI-driven search surfaces lean on accurate image understanding. Clean, honest alt text feeds those systems. Our guide on AI search image optimization explains how generative engines consume image signals.

The Recommended Workflow: Generate, Then Edit

The best results come from a hybrid loop. Do not publish raw generator output for important images. Do not write everything from scratch either.

1. Generate alt text for all images in bulk
2. Flag images by type:
   - Decorative      -> accept or set alt=""
   - Standard photos -> quick edit pass
   - Informative     -> full human rewrite
3. Review flagged items against page context
4. Publish, then re-audit quarterly

Choose tools that expose an editor UI for this loop. A generator that only writes straight to the live alt attribute removes your chance to catch hallucinations. Look for:

• Side-by-side image and editable description
• Confidence scores or flags for uncertain outputs
• Bulk export before anything goes live
• An audit log of what changed

If your current tool lacks an editing surface, export the generated text to a spreadsheet and edit there before importing back.

Bulk Audit: Find Every Missing Alt Attribute

Before you generate anything, measure the problem. This script crawls a sitemap and lists every image missing alt text:

import asyncio
from urllib.parse import urljoin
from xml.etree import ElementTree

import httpx
from selectolax.parser import HTMLParser

async def fetch(client, url):
    r = await client.get(url)
    return r.text if r.status_code == 200 else ""

def extract_images(html, base_url):
    tree = HTMLParser(html)
    out = []
    for img in tree.css("img"):
        src = img.attributes.get("src", "")
        alt = img.attributes.get("alt")
        out.append({
            "src": urljoin(base_url, src),
            "page": base_url,
            "has_alt": alt is not None,
            "alt_text": alt or "",
        })
    return out

async def audit(sitemap_url):
    async with httpx.AsyncClient(follow_redirects=True) as client:
        sm = await fetch(client, sitemap_url)
        urls = [
            loc.text
            for loc in ElementTree.fromstring(sm).iter("{*}loc")
        ]
        pages = await asyncio.gather(*[fetch(client, u) for u in urls])

    rows = []
    for url, html in zip(urls, pages):
        rows.extend(extract_images(html, url))

    missing = [r for r in rows if not r["has_alt"]]
    print(f"{len(rows)} images total, {len(missing)} missing alt")
    for r in missing[:50]:
        print(f"  {r['src']}  (on {r['page']})")

asyncio.run(audit("https://example.com/sitemap.xml"))

Run this first. The count tells you whether you need a generator strategy at all. Sites under a few hundred images may find that manual writing takes only a day or two — and produces better results than any automation.

For sites with thousands of gaps, sort the missing list by page traffic. Generate and fix alt text on high-traffic pages first. Low-value archive pages can wait or receive accepted raw output.

The Legal Stakes

Missing alt text is not just a quality issue. It is the most cited failure in web accessibility lawsuits.

WCAG success criterion 1.1.1 (Non-text Content) requires a text alternative for every meaningful image. WebAIM’s annual Million survey consistently finds missing alt text on the majority of home pages it checks. Plaintiffs’ law firms cite WCAG 1.1.1 failures in demand letters under the Americans with Disabilities Act, and similar claims arrive under the European Accessibility Act and comparable laws elsewhere.

Two points matter for your automation decision:

1. Empty alt on decorative images is compliant. alt="" tells screen readers to skip the image. Generating verbose descriptions for decorative images actually harms compliance.
2. Generated-but-wrong alt is a risk. A hallucinated description still technically satisfies 1.1.1, but it misinforms users. Courts care less; your users care more.

Automated scans such as axe or Lighthouse flag missing alt instantly. Fixing those flags at scale is exactly what generators are good at. Just route the important images through human review.

Tool Comparison

Tool type	Examples	Cost	Bulk support	Editing UI	Best for
CMS plugin	Alt Text AI (WordPress)	Free–$ monthly	Yes	Varies	WordPress sites wanting set-and-forget
Platform built-in	Shopify Magic, Squarespace AI	Included	Partial	Minimal	Stores already on the platform
CDN/API generation	Sirv AI Studio Alt Text	Usage-based	Yes, library-wide	Via dashboard	Teams hosting images on a CDN
Standalone SaaS	Dedicated alt text services	$–$$ monthly	Yes	Usually strong	Agencies and large media libraries
DIY vision API	GPT-4o, Claude	Per-call pennies	Your own scripts	Build your own	Dev teams needing custom prompts

Cost framing, approximate: DIY API calls run fractions of a cent per image. SaaS tools charge monthly fees scaled to volume. For a 10,000-image backlog, a DIY script costs a few dollars in API calls against hundreds in subscription pricing — if you have someone who can write and run the script.

Decision Summary

Use this table to pick your path:

Your situation	Recommended approach
Under 500 images missing alt	Write manually with the writing guide
Thousands of standard photos	Generate, then edit in priority order
Product catalog with structured data	Combine structured fields + generated visuals
Charts, infographics, diagrams	Human writing only
Decorative imagery everywhere	Set alt="" deliberately; skip generation
Regulated or news content	Human writing with fact-checking

Getting Started

Start small. Pick your 20 highest-traffic pages. Run the audit script on them. Generate alt text for their images with a strict prompt. Edit every result against the page context. Ship it.

Then expand in traffic order until the backlog clears. Keep the quarterly re-audit on the calendar, because new images arrive faster than old ones get fixed.

And when you write the final versions by hand, remember the rule that no generator replaces: alt text exists for a person who cannot see the image. Describe what they need to know. Everything else is decoration.

Ready to automate the rest of your image pipeline? Background removal, AI editing, and alt text generation live together in Sirv AI Studio — you can create a free account and test it against your own library.