AI Alt Text Generator for WordPress: What It Can and Can’t Do
An AI alt text generator does one narrow thing: it looks at an image and writes a sentence about it. That sounds simple, and for a product photo on a white background it is. For a chart, a screenshot, or a photo that matters because of where it sits on the page, the sentence can be confidently wrong.
The useful question is where the model's description stops being enough, and what to do before it reaches your media library.
How a vision model turns a photo into alt text
The plugin sends the image — usually resized to keep the request small — plus a text prompt to a vision-capable model such as those behind OpenAI, OpenRouter, or Google Gemini. The prompt asks for a description of the image, often with a word limit and sometimes with context like the file name or product title.
The model returns a sentence. It does not know which page the image sits on or what the surrounding paragraph argues. It sees pixels and a short instruction, and writes the most probable description for those pixels.
That is why a model can describe a photo accurately and still write alt text that is wrong for your page.
What AI is genuinely good at
- Literal description. Objects, people, clothing, setting, colors, and composition. A model will notice the sand-colored linen shirt, the flat lay, and the white background without being told.
- Text inside images. Screenshots, slide decks, signs, and packaging usually get their visible text transcribed accurately, which is exactly what a screen reader user needs.
- Consistent phrasing at volume. Five hundred images described in the same register, instead of five hundred sentences written on five different afternoons.
- Language coverage. The same image can be described in the language your visitors read, without you writing it twice.
- Speed on the boring middle. The 2016 event photos nobody wants to open are exactly the work AI is good at.
Where it falls down
The failures are predictable once you know what the model cannot see.
- Page context. A photo of a laptop is a laptop. Whether it is "a laptop running the AltMint review screen" or "a developer's laptop covered in stickers" depends on the paragraph around it, and the model has not read the paragraph.
- Product facts. The model can describe a jacket. It cannot know it is the rain shell, that the color is called "storm", or that the fabric is recycled nylon. Invented specifics are the worst kind of alt text because they sound authoritative.
- Charts and diagrams. A model can say "a bar chart". It usually cannot say what the chart shows or which trend matters, and that is the only reason the image exists.
- Decorative images. The model will happily describe a background flourish that should have an empty
alt="". Adding a description to decoration makes a screen reader announce noise. - Intent and tone. Sarcasm, brand voice, and the difference between a product shot and a lifestyle shot are human judgments.
- Repetition. Similar images get near-identical descriptions. Twenty product photos can end up with twenty variations of one sentence, none naming the product.
| Job | AI handles it | Human still needed |
|---|---|---|
| Describing objects and scenes | Yes, reliably | Trim to the relevant detail |
| Reading text in a screenshot | Usually | Check names and numbers |
| Explaining a chart's message | Poorly | Write the takeaway yourself |
| Choosing decorative vs informative | No | Decide per image |
| Knowing the product or campaign | No | Add names and facts |
| Keeping 500 images consistent | Yes | Spot-check samples |
A concrete before and after
Take a photo used on a WooCommerce product page for a linen shirt. The model returns:
"A person wearing a light-colored shirt standing in a studio."
Accurate, and useless. It misses the product, the color, and the view. A reviewer who knows the page rewrites it as:
"Model wearing the sand linen camp-collar shirt, front view, against a plain studio backdrop."
The second version takes fifteen seconds and does three jobs: it tells a screen reader user what the image shows, it names the product to match the page, and it stays accurate when the price or season changes. That is the pattern for almost every suggestion — the model supplies the observation, the human supplies the nouns.
The review-first workflow
There are two ways to ship AI alt text. The first is auto-apply: generate and write straight to the attachment. It is fast, and it is how sites end up with hundreds of descriptions that are technically true and practically useless.
The second is review-first, which is how AltMint is built. The plugin scans for images with no alt text, stores each generated sentence as a suggestion in a separate table, and shows it next to the image. Nothing touches the live alt attribute until you approve it. You can edit a suggestion, skip it, or approve the batch.
Review-first costs a few minutes per hundred images and catches the failures that matter: the invented product detail, the chart described as "a chart", the decorative swoosh that should have been skipped. It also protects descriptions you already wrote, because described images never enter the scan.
If you want the manual version of the same discipline, the step-by-step guide to adding alt text in WordPress covers where the field lives and what to write.
Bring your own key, keep your images
Most hosted alt text services upload your images to their servers — a real decision for a store with unreleased products. AltMint takes the bring-your-own-key route instead: you connect an OpenAI, OpenRouter, or Google Gemini key, it is stored in your WordPress database, and the only outbound request is the generation call to the provider you chose. There is no telemetry and no AltMint account in the loop.
Read your provider's data policy: the image does leave your server. If an image is confidential, do not generate alt text for it with a cloud model.
A quality checklist for AI suggestions
- Does it describe what matters here, not everything in the frame?
- Does it avoid invented specifics — model names, materials, locations?
- Would it still be true in six months, after a price change or a redesign?
- Does it avoid "image of" and other filler?
- If the image is decorative, should the alt attribute be empty instead?
- Does it read naturally out loud? That is the real test.
Frequently asked questions
Is AI-generated alt text bad for SEO?
Not inherently. Search engines evaluate whether the alt text describes the image and fits the page. Generic or repetitive AI output is weak for the same reason generic human output is weak. Reviewed, specific alt text helps regardless of who drafted it.
Can I let it apply automatically?
You can, but the failure is quiet: a wrong description is invisible to sighted visitors and obvious to screen reader users, so nobody reports it. Review before save, at least for product pages and charts.
Which images should I skip?
Decorative images, tracking pixels, and images whose meaning is already in the surrounding text. Those should get an empty alt attribute, not a generated sentence.
For stores, the context problem gets sharper: every product photo needs the product name, and the same shirt appears in a featured image and a gallery. That is covered in the WooCommerce alt text guide.