AI alt text review: five image decisions before handoff

An AI alt text review compares a proposed text alternative with the image's job on a page. After this exercise, you can save five decisions covering a descriptive alternative, an empty alternative and a request for more information. Your output is an editorial handoff, not an accessibility certification or a completed website change.

Alt text is an image's text alternative. A screen reader is software that presents page content through speech or another accessible output. Writing a longer picture description is not always the right task. An illustration can explain a service, repeat nearby wording or provide the only name for a link. We will keep those uses separate.

Content last updated: 15 October 2026. The repair workshop, images and proposed edits below are fictional teaching examples. No customer material, observed AI output or participant test is presented. The fifteen-minute plan assumes an existing working text assistant and a plain-text editor. Registration, website access and implementation are outside that allowance.

Why does page context change the answer?

Imagine a small bicycle repair workshop updating its service page. Someone asks an assistant to describe every image. The result sounds detailed, but a decorative stripe receives a paragraph while a booking icon is called a green square. The missing input is not a stronger adjective. It is what each image does in that location.

The W3C decision tree distinguishes images by context, including information, decoration and interaction. It also says its branches do not cover every case. Treat it as a starting point for a decision, not a stamp approving an entire site. Our exercise turns those distinctions into a short revision ledger with explicit unresolved items.

The separate image SEO guide covers upload preparation and discovery. This review does not resize files, choose compression or promise search improvements. Its deliverable is a clear answer to a narrower question: what should this particular image communicate through its alternative in this particular use?

What belongs in the five-image context card?

Copy the table into your working document. Preserve the surrounding text and interaction details; do not submit only the picture descriptions. You will send written context, not upload photographs. This deliberately separates a text-alternative decision from the different problem of whether a vision model recognizes an image correctly.

The B1 card states that the photograph visibly shows a frayed cable end. It does not identify the cause, the bicycle owner or whether riding is safe. B3 leads to the booking page, not directly to a confirmed appointment. B5 intentionally omits the numbers. Those gaps are boundaries, not invitations to generate plausible details.

Add columns for role, suspected issue, proposed alternative, human decision, open question and owner. Leave every decision pending at first. Keep the B IDs even if a proposed alternative is removed. A reviewer needs to trace a decision to its original object rather than infer the relationship from a polished final paragraph.

How should you ask the assistant for proposals?

Use an existing text assistant such as Gemini. Open gemini.google.com in a browser, sign in with an account you are allowed to use and start a new conversation. If text chat is unavailable, complete the ledger manually instead. This exercise does not require an image attachment, an integration or a new paid plan.

A prompt is the instruction you send the assistant. Paste the following into its message box, replacing [CONTEXT CARD] with all five table rows and the extra B1, B3 and B5 notes. Read the message before sending it. The fictional card is sufficient; customer photographs and private appointment details add nothing to this task.

This is a fictional alt text editing exercise. CONTEXT CARD: [CONTEXT CARD]. For every B ID, identify the image's role on the page, the issue with its current alternative and a proposed change. Separate appearance from function. Ask a question when context is missing. Do not invent visual details, safety conclusions or chart data. Explain when an empty alt attribute may be appropriate; do not confuse it with a missing attribute. Do not squeeze an entire complex chart into a short alternative. Leave the human-decision column empty. Do not certify accessibility or change website code.

Count the IDs in the response before judging the wording. If B2 becomes informative simply because its colour is described, compare that with its stated decorative role. If B3 says an appointment is confirmed, correct it: the link only opens booking. If you need broader task framing, the SEO prompt examples cover other editorial jobs, not approval of these decisions.

Which proposed changes should pass your review?

For B1, a candidate is “Close view of a frayed cable end.” That reflects the supplied visual detail without adding a diagnosis. Remove the sales slogan. Do not turn the alternative into advice about whether the bicycle is safe to ride. During a real implementation, someone must compare the wording with the actual photograph; this exercise checks the written card only.

For B2, the expected editorial decision is an empty alternative: alt="". Do not type the word empty as its value. Do not simply remove the attribute and call that equivalent. Record the reason: the stripe adds no information or action. Have the technical owner check how the real site renders it rather than applying a blanket instruction to every green image.

For B3, “Book a repair appointment” communicates the link purpose more usefully than the icon's geometry. It names the booking action without claiming that a slot has been reserved. The card explicitly states that there is no other name. If visible wording or another accessible name is added later, review the image decision in that new context.

For B4, the link already contains Workshop home as visible text. In this exercise, an empty image alternative avoids repeating that same destination. This is not a rule that every logo should be silent. A standalone linked logo is a different case, and a logo communicating additional information would need its own assessment.

For B5, keep the item open. Request the underlying repair counts and a suitable explanation or data presentation elsewhere on the page. Do not accept an invented trend as a convenient replacement. A short title does not deliver all the information in a complex chart. The content owner must supply what the current card does not contain.

Finish this pass with five accounted-for objects, not five nonempty strings. Accepting a blank alternative with a reason and retaining a chart as unresolved are legitimate ledger outcomes. A response that fills every field with confident prose may be less useful than one that exposes the remaining work accurately.

What can you save in fifteen minutes?

1. Minutes one to three: copy the context card and its version. Expected output: five identifiable source items.
2. Minutes four to seven: send the request and inspect ID coverage. Expected output: proposals attached to the correct objects.
3. Minutes eight to twelve: compare each proposal with the checks above. Expected output: accept, revise or unresolved beside every ID.
4. Minutes thirteen to fifteen: save and reopen the document. Expected output: a readable handoff retaining the chart question.

This is a suggested work allocation, not a measured completion guarantee. If login trouble or a difficult decision interrupts it, record that blocker. You can still finish a useful source card manually. Do not replace a missing fact with a guess just to satisfy the time target.

Open Notepad or another plain-text editor. Add the title Alt review 01, draft only. Paste the context and checked decisions. Press Ctrl+S, choose Documents and save alt-review-01.txt. Close the editor, reopen the file from that folder and check that B1 through B5 remain present, with B5 visibly unresolved.

In my earlier writing about effectiveness and perfection, I described how pursuing every detail delayed my projects. The application here is to finish an inspectable five-item handoff rather than claim that a whole website is perfect. That is an application of a working principle, not a report of an accessibility project or a measured customer outcome.

What remains before a real website change?

Give the technical owner the page address, image ID, current alternative, approved proposal and open questions. The website audit guide provides a separate framework for assigning and checking site work. Do not transfer this fictional exercise straight into a production editing instruction without matching it to real objects.

Where visual details are uncertain, the image analysis guide covers that different first step. Verify the observation before using it in an alternative. The artificial intelligence guide gives the broader background. Neither an assistant response nor this text review replaces implementation checks and appropriate user testing.