A store owner with 4,000 product photos and no alt text recently asked me a fair question: “Can’t I just let AI write all of it?” Many content tools, plugins and site builders now offer a one-click “generate alt text” button. It feels like the problem is solved.
It is partly solved. AI is very good at describing what is in a picture. But alt text is not a picture description. It is a short piece of text that does the same job the image does on that page, for someone who can’t see it, and for Google. AI usually can’t see the page, the product, or the reason you put the image there.
Let me explain what AI-generated alt text gets right, where it goes wrong, and a practical review workflow that gives you most of the speed without the quality problems. I will also cover how to process images in bulk and which accessibility rules you still need to meet.
Why so many teams are handing alt text to AI
The size of the gap explains the appeal. The WebAIM Million 2026 report, based on February 2026 data from the top one million home pages, found an average of 66.6 images per home page. It found that 16.2% of home page images had no alt text at all, about 10.8 per page, and that 53.1% of home pages had at least one image missing alt text. Of the images that did have alt text, 10.8% used questionable or repetitive wording.
Those are home pages, which usually get the most care. Product catalogs, blog archives and old landing pages are often worse. Writing good alt text by hand for thousands of images takes real time, so it never gets done.
That is the honest case for AI:
- It closes the “nothing at all” gap fast. Even a plain description is better than a screen reader announcing a file name like “IMG_4471.jpg.”
- It is consistent. It doesn’t get tired at image 900.
- It gives writers a first draft. Editing a sentence is faster than writing one from nothing.
- It can run inside your workflow. Many tools generate a suggestion at upload time, so new images don’t add to the backlog.
If you are new to the basics, start with the lesson on image SEO and my full guide to alt text for SEO. This article focuses on the AI part.
What AI image descriptions usually get right
Modern vision models are good at the literal layer of an image. In general, they handle these well:
- Objects and people: “a woman holding a coffee cup,” “a red brick house with a black door.”
- Colors, materials and settings: “a navy linen shirt on a white background.”
- Visible text in the image: reading a sign, a button label or a heading in a banner.
- Clean, short sentences without “image of” at the start, if you tell them not to use it.
This lines up with how Google itself works. Google’s image documentation says it uses computer vision together with alt text and page content to understand images. Here is its example of good alt text:
“<img src="puppy.jpg" alt="Dalmatian puppy playing fetch"/>”
— Google Search Central, Image SEO best practices
That is the kind of sentence AI can produce well: short, specific and accurate. For a generic photo with an obvious subject, AI output often needs little editing.
Where AI alt text goes wrong
The problems start when the meaning of an image depends on something the AI can’t see. Silktide summed up the issue well in a May 2026 article:
“AI tools tend to optimize for visual accuracy when what you need is contextual meaning.”
— Silktide, The downsides of AI alt text
Silktide gives the example of a protest photo described as people outdoors holding rectangular objects with text. Every word is true, and the point of the photo is lost. Here are the error types I see most often, with an example of each:
| Error type | AI output (example) | Better alt text |
|---|---|---|
| Accurate but pointless | “A man smiling in an office” | “Dr. Priya Shah, lead dentist at our Leeds clinic” |
| Wrong product details | “Black running shoe” | “Trailmax 3 running shoe in charcoal, side view showing grip sole” |
| Describes a chart, not its message | “A bar chart with blue and orange bars” | “Bar chart: leads rose from 40 to 95 per month after the redesign” |
| Misses the function of a linked image | “Magnifying glass icon” | “Search” |
| Describes decoration that should be silent | “Abstract wavy blue lines” | Empty alt (alt="") |
| Guesses or invents | “A happy family in their new kitchen” | “Kitchen renovation in Harrogate, after: oak cabinets and quartz worktops” |
| Repeats the caption or nearby text | Same sentence as the caption below | Short alt, or empty if the caption fully covers it |
A few of these deserve a closer look.
Linked images and buttons
When an image is a link or a button, its alt text should say what happens when you click, not what the icon looks like. The W3C calls these functional images. This matters for SEO too: Google says it uses the alt attribute of a linked image as the link’s anchor text. WebAIM found that 45% of images missing alt text were linked images, so this is a common gap. AI tools that only see the image file will describe the picture and miss the purpose.
Decorative images
Some images add nothing for a screen reader user: background shapes, dividers, generic mood photos. These should have an empty alt attribute so screen readers skip them. Most AI tools will happily describe them anyway, adding noise. My guide on when to leave alt text empty covers how to decide.
People and sensitive details
AI may guess at age, ethnicity, gender or emotion. On most business pages, those guesses add nothing and can be wrong or hurtful. I suggest instructing the tool not to infer personal traits and to name people only when you supply the name.
Feeding the AI the context it can’t see
Most of the errors above have one cause: missing context. Research backs this up. In a 2022 Stanford study covered by Stanford HAI, sighted and blind or low-vision participants rated image descriptions written for different Wikipedia articles. Quality ratings closely tracked how relevant the description was to the article. The researchers’ conclusion was clear:
“the alt-text description should reflect the particular context to help clarify its meaning.”
— Stanford HAI, Creating a More Accessible Internet: Context Matters
So the fix is to give the AI that context. If your tool lets you write a prompt or send extra fields, include these:
- Page title and purpose, for example “Product page for Trailmax 3 running shoe.”
- Product data: name, color, variant, material, and the angle of the shot if known.
- Role of the image: main product photo, detail shot, lifestyle photo, chart, logo, icon, link or button.
- Nearby text, such as the caption or the paragraph next to it, so the AI doesn’t repeat it.
- Rules: under about 125 characters as a working limit, no “image of,” no personal traits, no keywords that aren’t visible or true.
A simple prompt template might look like this:
Write alt text for an image on this page.
Page: {page_title} – {page_purpose}
Image role: {role}
Product data: {name}, {color}, {variant}
Nearby caption: {caption}
Rules: one sentence, under 125 characters, describe what matters
for this page, do not start with "image of", do not guess
people's age, ethnicity or feelings, return "DECORATIVE" if the
image adds no information.The 125-character figure is a common rule of thumb, not an official limit. The W3C doesn’t set a length. It asks that the text alternative convey the meaning of the image, and Ahrefs puts the practical side well:
“Be concise. Lengthy alt text is annoying for those using screen readers.”
— Ahrefs, Alt Text for SEO: How to Optimize Your Images
For product-specific patterns you can feed into the prompt, see my product image alt text templates. If your images have captions, read image captions and surrounding text so your alt text and captions work together instead of repeating each other.
A review workflow that keeps humans where they matter
Reviewing every AI suggestion by hand defeats the purpose. Reviewing none of them is risky. The middle path is to sort images by risk and review the risky ones properly. Here is how I would split them:
| Image type | Risk if AI is wrong | Review level |
|---|---|---|
| Main product photos | High: wrong color or model misleads buyers | Check every one against product data |
| Charts, infographics, diagrams | High: AI describes the shape, not the data | Human writes it; AI draft optional |
| Linked images, icons, buttons | High: wrong function breaks navigation | Human writes it from the link target |
| Team, people and location photos | Medium: needs names and places AI can’t know | Human adds names; spot-check the rest |
| Blog illustrations and lifestyle photos | Low to medium | Spot-check 1 in 10 |
| Decorative images | Low, but adds noise if described | Confirm empty alt; no description |
And here is the step-by-step process I suggest for a team:
- Inventory. Crawl the site and export every image with its page URL, current alt text, whether it is inside a link, and nearby caption.
- Prioritize. Start with pages that get organic traffic or sales, and linked images. Ahrefs also suggests focusing on pages that already earn search traffic rather than fixing every image at once.
- Classify. Tag each image by role using the table above.
- Generate. Run AI with context only on the roles where it helps.
- Review. Apply the review level for each role. Reviewers fix, not just approve.
- Publish and log. Store who approved what and when.
- Test. Listen to a few key pages with a screen reader such as VoiceOver or NVDA. Ten minutes of listening finds problems that spreadsheets hide.
Running AI alt text across thousands of images
Bulk work needs a little more care, because one bad rule repeats itself thousands of times. A few practical points:
- Don’t overwrite good human alt text. Only fill blanks, file-name alt text (like “DSC_0042”) and obvious junk. Keep a backup column of the old values.
- Use the same image URL everywhere. Google recommends referencing an image by the same URL across pages. If the same photo appears in several places with different purposes, set alt text per placement, not just in the media library.
- Watch for CMS quirks. In WordPress, for example, alt text is typically copied into the post when an image is inserted, so changing it in the media library later may not update images already placed in posts. Check how your theme or builder outputs it.
- Run a sample first. Generate 50 images, review all 50, fix the prompt, then run the rest.
- Re-check after launch. Crawl again to confirm the alt text actually appears in the HTML.
Example: estimating effort for a 4,000-image store
This is a hypothetical example. Say a store has 4,000 images: 1,200 main product shots, 2,400 secondary product shots, 250 lifestyle photos, 100 icons and buttons, and 50 charts or size guides. Fully manual writing at about one minute per image is roughly 67 hours.
With AI plus risk-based review: main shots are checked at about 20 seconds each (around 7 hours), secondary shots spot-checked at 1 in 10 (around 1.5 hours), icons and charts written by hand (about 2.5 hours), lifestyle photos spot-checked (under an hour). That is roughly 11–12 hours of human time. The saving is real, and it still keeps a person on the images that matter most.
The accessibility rules AI output still has to pass
Alt text is first an accessibility requirement and only second an SEO signal. The relevant rule is WCAG Success Criterion 1.1.1, Non-text Content, which is Level A, the most basic level. It requires that non-text content has a text alternative that serves the equivalent purpose, with exceptions such as decoration (W3C, Understanding SC 1.1.1).
The key word is purpose. An AI description can be accurate and still fail the purpose test. The W3C’s alt decision tree is a good checklist for reviewers:
| Decision tree question | If yes | Can AI decide this alone? |
|---|---|---|
| Does the image contain text? | Include the text, unless it is also shown as real text nearby | Partly: it can read text, but not check what is nearby |
| Is it in a link or button? | Describe the destination or action | No, unless you pass in the link |
| Does it add meaning to the page? | Brief description; long description for complex images | Partly, with page context |
| Is it purely decorative? | Empty alt | Rarely; it tends to describe everything |
Also note what does not count. An automated checker will pass any non-empty alt attribute, including a useless one. Silktide warns that AI alt text can pass automated checks while failing real use. Passing a scanner is the floor, not the goal. For the wider link between accessibility and rankings, see accessibility and SEO, and for the related confusion around tooltip text, my article on link and image title attributes.
Questions teams ask about AI-written alt text
Does Google penalize AI-generated alt text?
Google doesn’t say it treats AI-written alt text differently. What it does warn against is keyword stuffing in alt text. Useful, accurate and relevant text is what matters, however it was drafted.
Is AI alt text better than none?
For most informative images, yes, a plain accurate description beats a missing one. The exceptions are linked images (where the description can mislead), decorative images (where silence is correct) and charts (where the description hides the data).
Should I add keywords to the AI prompt?
Give the AI the product name and page topic, because those are true facts about the image. Don’t ask it to “include SEO keywords.” That usually leads to stuffed, awkward alt text.
Can my CMS generate alt text automatically on upload?
Many can, through built-in features or plugins. Treat the result as a draft, especially for product photos and linked images, and make sure it doesn’t overwrite alt text a person already wrote.
How long should AI-generated alt text be?
Usually one short sentence. Many teams use about 125 characters as a working limit. Complex images like charts need a short alt plus a fuller description in the page text.
Getting your image backlog under control
In brief: AI is a great first-draft writer for alt text and a poor final judge. Use it to close the “missing” gap quickly, give it the page context it can’t see, sort images by risk, and keep people on product photos, linked images and charts. Then listen to your key pages with a screen reader.
If you run a large catalog and want help with an image audit or an alt text rollout, our ecommerce SEO team can plan it with you. Tell us about your site on the get a quote page.
Further reading on AI-generated alt text
- Keyword Relevance: How to Show Google Your Page Matches the Query
- SEO Over-Optimization: Warning Signs and How to Dial It Back
- Title Tags for Local and Service Pages
Sources for this AI alt text guide
- WebAIM – The WebAIM Million (2026 report)
- Stanford HAI – Creating a More Accessible Internet: Context Matters (2022)
- Silktide – The downsides of AI alt text (2026)
- Google Search Central – Image SEO best practices
- Google Search Central – Link best practices
- W3C – Understanding SC 1.1.1 Non-text Content
- W3C WAI – An alt Decision Tree
- W3C WAI – Functional Images
- Ahrefs – Alt Text for SEO: How to Optimize Your Images
