To generate alt text with AI, give the tool each image plus its context: the page's purpose, the product name and the details a shopper needs. Then check every description for accuracy, relevance and length, which Shopify recommends keeping to 125 characters or fewer. Decorative images get empty alt text instead of a description.
The part AI gets wrong is judgement. It describes what is visible, not why the image is on the page, and it fills gaps with confident guesses: "leather" for faux leather, "navy" for a colour you sell as "ink", a text-heavy sale banner summed up as "promotional graphic". Each of those is a small misdescription that a screen reader user hears as fact.
The worked example here is an illustrative four-person online clothing shop with 60 new product photos to describe before an autumn launch. Along the way there are examples from a jewellery seller, an events company, a sign maker, a print shop and a software reseller, because each kind of image needs a slightly different kind of description.
The standard most accessibility checks refer to is WCAG 2.2 success criterion 1.1.1, a Level A requirement: non-text content needs a text alternative that serves the equivalent purpose, and purely decorative images should be hidden from assistive technology. "Serves the equivalent purpose" is the test to keep in mind. Alt text is not a caption and not a keyword field. It stands in for the image for someone who can't see it.
Four places AI alt text comes from, and how far to trust each
You may already have an AI alt text generator switched on without realising. These are the common sources for a small business, as of September 2026:
| Source | How it works | Good for | Watch out for |
|---|---|---|---|
| Shopify's media editor | Shopify says the media editor might suggest alt text for images that don't have any; accepting a suggestion saves it to the file | A quick first draft on new product images | It sees the photo, not your fabric composition or colour names |
| Word and PowerPoint | Automatic on Copilot+ PCs when you insert an image; elsewhere you request it from the Alt Text pane or the Accessibility Assistant | Reports, proposals and slide decks | Microsoft itself says generated descriptions may not be fully accurate and should be reviewed |
| Social platforms | Instagram creates automatic alt text and lets you write your own in a post's advanced settings | Nothing, until you replace it | Automatic descriptions are generic ("may be an image of a person") |
| A chat assistant (ChatGPT, Claude, Gemini) | You upload images and paste the product details, and it drafts descriptions in bulk | Batches where you can supply context | It will guess details you don't give it unless told not to |
There are also bulk alt text apps in the Shopify App Store that write descriptions from your product data. They are third-party tools with their own pricing and quality; test one on ten products before letting it touch the whole catalogue, and read what it does to images that already have alt text.
Sort every image by its job before generating a word
The W3C's alt decision tree asks what each image is for, and the answer decides what the alt text should say. AI tools skip this step entirely, which is why they describe a "Get a quote" button as "blue rectangle with white text". Here is the tree applied to real small-business images, all illustrative.
- Informative image: describe what matters on this page. A sign maker's portfolio photo. AI draft: "A shop front with a sign above the door." Better: "Illuminated fascia sign in brushed aluminium with halo-lit individual letters, fitted above a café window." The portfolio page exists to show workmanship, so the materials and lighting are the point.
- Decorative image: empty alt text. The watercolour texture behind an events company's page header adds mood and nothing else. It should have an empty alt attribute so screen readers skip it. Most website builders let you leave the field blank or mark an image as decorative. An AI tool will happily write "abstract pastel background", which wastes a listener's time.
- Functional image: describe the action. A print shop uses an image as its "Get a quote" button. The alt text should be "Get a quote", because that is what happens when you select it.
- Image of text: include the text. The same print shop's homepage banner reads "Business cards from $29 until 31 October. Free delivery over $50." The AI draft said "Promotional banner for business cards". A screen reader user would miss the price, the deadline and the delivery offer, which is the whole message.
- Complex image: short alt text plus the detail on the page. A software reseller's report includes a chart of support tickets. AI draft: "A bar chart showing data over time." Better alt text: "Bar chart: monthly support tickets fell from 84 in January to 31 in June." Then put the month-by-month figures in a table below the chart, which the decision tree recommends for complex images.
In the clothing shop's product galleries every photo is informative, because each one shows something the others don't. Decorative images live elsewhere: the homepage background, a divider graphic, a repeated brand pattern in the footer.
A product alt text prompt that asks for doubts, not guesses
A chat assistant writes much better alt text when you give it the facts it can't see and forbid it from inventing the rest. The clothing shop uses this prompt, uploading all five photos of one product at a time with the details pasted from the listing:
Write alt text for each of these product photos for an online clothing shop.
Product facts (use these, don't contradict them):
Name: [product name] | Colour name: [our colour name] | Fabric: [composition]
Fit/length: [details] | Features: [buttons, pockets, lining, etc.]
Rules:
- One description per photo, 125 characters or fewer.
- Say what THIS photo shows that the others don't (back, detail, fabric, fit).
- Start with the item, not "Image of" or "Photo of".
- Use our colour name and fabric exactly as given. Never guess a material.
- Don't describe the model's age, ethnicity, body type or gender.
Mention the model only to show fit ("on model", "tucked in").
- No marketing words (stylish, perfect, must-have).
- If you can't tell something from the photo or the facts, write [CHECK: ...].
Return a numbered list matching the photo order.This is an illustrative first run for a coat, with the facts pasted in as "Colour: camel. Fabric: 60% wool, 40% polyester. Knee length, single-breasted, three buttons, centre back vent, half belt":
- "Camel wool-blend coat on model, knee length, single-breasted with three buttons, worn open over a white shirt."
- "Back view of the camel coat showing a centre vent and a half belt at the waist."
- "Close-up of the coat's twill weave, with a soft brushed surface. [CHECK: is the weave twill?]"
- "Detail of a dark horn buttonhole and button on the camel coat. [CHECK: button material]"
- "Camel coat laid flat, showing two flap pockets and the full length."
What the owner fixed: photo 3 was indeed twill, so the placeholder came out; photo 4's buttons are plastic in a horn effect, so "dark horn buttonhole" became "dark horn-effect button", which matters because "horn" is a material claim. Photo 1 mentions a white shirt, which is styling rather than product, but it helps a shopper picture the outfit, so it stayed. Each description is under 125 characters and each says something different. Without the "[CHECK]" rule, the assistant had written "genuine horn buttons" on an earlier test, which is exactly the kind of confident guess that turns into a complaint.
Sixty autumn product photos: the clothing shop's run-through
The shop launches 12 products for autumn with five photos each: front on model, back on model, flat lay, fabric close-up and one detail. Before the project, 41 of the 60 images had no alt text and the rest had the product title pasted in five times. Here is how the afternoon went, with illustrative figures.
Step 1: export to a spreadsheet (10 minutes)
Shopify's product CSV export includes an "Image alt text" column, with a 512-character maximum and 125 or fewer recommended. Exporting the 12 products gave one row per image, which makes it easy to see gaps, duplicates and lengths at a glance. A spreadsheet formula such as =LEN(C2) next to the alt text column flags anything over 125 characters.
Step 2: generate per product, not per image (15 minutes)
Uploading the five photos of one product together, with its facts, meant the assistant could see which photo showed what. Twelve prompts, about a minute each including pasting.
Step 3: check every description (25 minutes)
This is where the value is. The owner checked each description against the photo and the listing, at roughly 25 seconds per image. What she found in the 60 drafts:
| Problem found | Count | Example |
|---|---|---|
| Wrong or guessed material or colour | 7 | "Navy knit jumper" for a colour sold as "Ink" |
| Placeholder left for a fact to confirm | 9 | "[CHECK: lining fabric]" |
| Over 125 characters | 4 | A flat-lay description listing every seam |
| Detail wrong (collar, pocket, fastening) | 3 | "Button-down collar" on a spread collar |
| Fine as drafted | 37 | — |
Step 4: import and spot-check on the live site (10 minutes)
After re-importing the CSV, she opened three product pages with her phone's screen reader switched on (VoiceOver on iPhone, TalkBack on Android) and swiped through the galleries. Hearing the descriptions in order is the fastest way to notice two photos that say the same thing.
The quick sum: writing 60 descriptions by hand at about two minutes each is two hours. The AI route took about an hour including the full check. The time saved is modest; the bigger gain is consistency, because the rules in the prompt apply to every product, including the ones added on a busy Friday.
The five checks every AI description must pass
Whether the draft came from Shopify, Word, Instagram or a chat assistant, run it through these five checks. Two to three seconds each once you're used to them.
- Is every fact true? Material, colour name, count ("set of three"), size and any text in the image. If the image contains words, are they transcribed exactly?
- Does it say what matters on this page? The same photo of a jewellery seller's ring needs "hammered silver ring, 4mm band, shown on a finger for scale" on the product page, but "the maker's bench with tools and a half-finished ring" on the About page.
- Is it short enough to listen to? Aim for one sentence, 125 characters or fewer for product images. Long detail belongs in the product description.
- Does it avoid filler? No "image of" or "picture of" (screen readers already announce an image), no marketing adjectives, and no keyword lists. Google's own example of good alt text is "Dalmatian puppy playing fetch": specific, short, no padding.
- Does it avoid guessing about people? Leave out age, ethnicity, gender or body shape unless it matters to the page and you know it rather than inferred it. If fit depends on the model's height, state it in the product text ("Model is 175cm and wears a size 10").
Google explains how it reads images in its image SEO guidance, including the warning that keyword-stuffed alt attributes may cause a site to be seen as spam. Accurate descriptions help search and accessibility at the same time; stuffed ones help neither.
Six alt text mistakes and how customers would notice them
Each of these is illustrative, but all are common with AI-generated descriptions:
- The guessed material. A faux-leather bag described as "genuine leather tote" by a bulk app. The product text was correct, so the page contradicted itself, and a shopper who used a screen reader relied on the alt text. A misdescription like that is a returns problem at best and a trading complaint at worst.
- Five identical descriptions. A bulk tool copied the product title into every image. A screen reader user swiping through the gallery hears "Oversized cord shirt, rust" five times and learns nothing about the back, the fabric or the fit.
- Keyword stuffing. An "SEO alt text" setting produced "rust cord shirt oversized shirt womens shirt autumn shirt cord overshirt". It reads as spam to a listener and, by Google's account, may be treated as spam by search.
- The untranscribed offer. The print shop's sale banner had alt text "Sale banner". Customers using screen readers never heard the price or the closing date.
- Describing bodies. One draft opened with "slim young woman with long blonde hair". It is irrelevant to the jumper, it's a guess, and it's the kind of detail that makes an otherwise helpful page feel careless.
- The useless chart description. Word's generated alt text for a reseller's chart read "A graph of a line". Microsoft's own guidance is to review AI descriptions, and for charts the fix is always the takeaway in the alt text plus the numbers in a table.
AI-generated product photos add a twist: the image itself may contain errors, such as a buttonhole with no button or a pocket on one side only. Check the picture before you describe it; the tutorial on checking AI-generated images for errors before posting covers what to look for, and the one on creating product photos with AI on a small budget covers making them.
Social posts, emails and documents: the images people forget
Product pages get attention because they sell. Alt text elsewhere is often missing entirely.
Social posts
Instagram generates automatic alt text, but you can write your own in a post's advanced settings before publishing and edit it afterwards. For a carousel from the events company's latest wedding, "Long banquet tables under festoon lights in a barn, set for 120 guests" is far more useful than the automatic "may be an image of indoor". If you also label AI-generated visuals, the tutorial on labelling AI-generated images and video on social media explains how the two fit together, and the pre-publish checklist for AI-drafted posts puts alt text on the list.
Newsletters
Some email apps don't load images until the reader allows them, so the alt text is what many subscribers see first. A newsletter hero image with the text "Autumn sale: 20% off coats this weekend" needs that text in its alt attribute, or the offer is invisible until someone clicks "show images".
Proposals, reports and slide decks
Word and PowerPoint will draft alt text for you, automatically on Copilot+ PCs and on request elsewhere. Review each one, and for charts write the takeaway yourself. A proposal a client forwards to a colleague who uses a screen reader should make the same case to both of them.
Keeping descriptions accurate as stock and photos change
Alt text goes stale. A colourway is renamed, a supplier changes the lining, a photo is swapped for a new one and the old description stays attached to the product. Three habits keep it right:
- Write alt text at the moment a product goes live. Make it part of the listing checklist, next to price and stock count, not a separate project.
- Re-export every quarter. Sort the alt text column for blanks, duplicates and anything over 125 characters. A 300-image shop can scan that in 15 minutes.
- Re-check after any bulk change. Renaming colours or importing new photos can leave descriptions describing things that no longer exist.
There is a legal angle too. If you sell to customers in the EU, accessibility requirements for online shops have applied since 28 June 2025, with micro-enterprises that provide services exempt; other markets have their own rules. Whether and how they apply to your business is a question for an adviser, but accurate alt text on every product image is part of meeting them almost everywhere.
Alt text questions owners ask
Should alt text include my keywords for SEO?
Only where the keyword genuinely describes the image, such as the product name. Google says it uses alt text together with computer vision and the page content to understand an image, and warns that stuffing alt attributes with keywords gives a poor experience and may cause a site to be seen as spam. Describe the image accurately and the relevant words appear naturally.
Do all five photos of one product need different alt text?
Yes. A screen reader user moving through the gallery hears each description in turn, so five copies of the product title tell them nothing new. Describe what each photo adds: the back view, the fabric texture, the fit on a model, a close-up of the fastening.
Is it acceptable to leave the automatic alt text that Instagram or Word creates?
Treat it as a first draft. Automatic descriptions are usually generic, such as a person standing outdoors, and Microsoft itself encourages people to review and revise AI-generated alt text. For anything that sells a product or carries information, replace it with a description written for that post or document.
How long does it take to fix alt text across a small shop?
With a spreadsheet export, AI drafts and a proper check, allow about 40 seconds to a minute per image once you have a routine. A shop with 300 product photos is roughly a working day, which is why it is worth doing new products properly as they are added rather than in one big catch-up.
Further reads
- Brand Guidelines for AI Images: Colours, Style and Limits — Set rules for the AI images you describe in the first place.
- How to Audit Your Website for AI Content That Needs Fixing — Find other AI-written parts of your site that need checking.
- AI Product Photos vs a Photographer: Cost and Quality Compared — Compare AI product photos with a photographer on cost and quality.
- Taking a Local Shop Online With AI: Product Listings in a Weekend — Build complete product listings with AI in a weekend.
- How to Spot Bias and Stereotypes in AI-Written Content — Catch assumptions about people in AI descriptions and copy.
- How to Check AI Content for Plain English and Readability — Apply the same checking habits to your written product copy.
- Eight Practical AI Uses for an Independent Optician — Recalls, enquiries, plain-English prescriptions, lens reorders, frame copy, reviews, referral letters and stock: how each works in a small practice.
- How to Write Instagram Captions With AI That Sound Like You — Build a voice file from your own captions, feed the AI raw notes instead of topics, and use a read-aloud edit so every caption sounds like you wrote it.
- How to Create On-Brand Social Media Graphics With AI — Brand rules, a brand kit and six master templates keep AI-made graphics consistent, with a failed poster prompt, a working image prompt and a 40-minute batch.
- How to Keep AI-Made Videos On-Brand: Captions, Voice and Intros — Stop AI video tools restyling your brand: caption settings, an approved voice and pronunciation list, locked intro templates and disclosure rules that apply.
- Email Newsletter QA Checklist Before You Hit Send — A grouped pre-send checklist for newsletters, with why each check matters and how to verify it, plus a worked launch email with the errors it caught.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: W3C Web Accessibility Initiative (alt decision tree; WCAG 2.2 Understanding Success Criterion 1.1.1); Google Search Central image SEO best practices; Shopify help pages on adding alt text to media and product CSV files; Microsoft Support FAQ on AI-generated alt text; Instagram help pages on alternative text.