Alt Texts from the Picture
An alt text is the sentence a screen reader reads out, an image search indexes and a browser shows when the picture does not load. The SEO Engine writes it from the picture itself: a vision model looks at every gallery picture, and a sentence is written only for pictures that demonstrably show their item.
Two sentences describe the procedure:
No alt text for a picture that does not fit. If a picture shows something other than the item — a headset on a charging cradle, a label roll with another part number — it is held back, and no sentence is written for it at all.
No sentence reaches a picture without a human. Every proposal waits in the Alt Text Review until somebody accepts it.
Preview
The feature is registered as Preview. It is fully usable; the label says it may still change after the first catalogue runs.
Prerequisites
- A vision model as a deployment of its own in the merchantCENTRAL AI Setup, for example
gpt-6-sol, with Accepts Images and — for reasoning models — Reasoning Model. Check it with Test Connection. - Under Model per Feature, assign the feature Image Alt Texts to that deployment. The listing drafts keep their model.
- The feature Image Alt Texts is active under Copilot & AI Capabilities (it is after installation).
Without the assignment the feature does not fall back to the standard model; it says what is missing.
The procedure
For every item, in five stages:
| Stage | What happens |
|---|---|
| 1. Blind analysis | The model sees only the picture, not the item — so it cannot see what it expects. It reports the kind of picture, the view, the object, visible parts, legible text, colours and background. The picture is scaled down to at most 1024 pixels first. The analysis is stored per file; the same file on many items is paid for once. |
| 2. Fit check | A separate call compares every analysis with the item data and judges: fits, does not fit or unclear. Reasons for "does not fit": another product type, a different brand or type number, a clearly different body colour, a sample or prototype notice. Packaging, scope of delivery, drawings and logos count as fitting. |
| 3. Writing | Only pictures that fit go on. The writing call gives its own verdict first, then the sentence. If it disagrees with the fit check, the picture is held back. |
| 4. Code checks | At most 125 characters, the product name included, no plain background, no "picture of …", no conflict note in the sentence; different files get different sentences, the same file the same one. On a violation there is one repair round. |
| 5. Approval | The proposal waits in the Alt Text Review for Accept. |
A missing verdict, an unreadable answer or a picture left out of the answer counts as unclear — the picture is held back.
What the model gets
- In the analysis: the picture — nothing else.
- In the fit check and writing: brand, manufacturer, manufacturer part number, marketplace category, current title, item description, item category, item attributes and the analyses of the pictures. When writing, also the existing alt texts of the gallery as the house style, and your style instructions from the setup.
No prices, no stock, no customer or vendor data. The picture and the item data never go to the model in the same call.
Status per picture
| Status | Meaning | Alt text in the shop |
|---|---|---|
| New | Not checked yet, or the check failed (picture not reachable, the AI did not answer). The job tries up to three times. | unchanged |
| Review Image | The fit check or the writer says "does not fit" or "unclear". No sentence is written. | unchanged |
| Review Text | The sentence still breaks a rule after the repair. | unchanged |
| Proposal | Every check passed. | unchanged until accepted |
| Accepted | Approved and written to the item picture, source AI. | new text |
| Rejected | Discarded. Proposed again only when the picture file changes or somebody drafts again. | unchanged |
| Kept | The picture fits and already has an alt text, which stays. | unchanged |
The Alt Text Review
Via search: Alt Text Review, or through the counters Pictures to review and Alt texts to accept in the Waiting for you card of the SEO dashboard.
Every checked picture stands on a card, 20 items a page:
- the picture itself — a click shows it large,
- the status with a colour stripe and the reason in one sentence, for Review Image in the AI's own words ("The picture shows a headset …"),
- Fit check (fits / does not fit / unclear and why), The AI sees (kind of picture, view, main object, colours, text it read) and the alt text today,
- the proposal with a character counter against the limit of 125 and the note AI-generated — read it before you accept it,
- how many pictures are the same file — a click shows them all; a wrong photo on 42 items is found at once,
- small at the bottom: model, prompt versions and date of the check.
At the top, the filters choose what is shown: All Open, Review Image, Review Text, Proposals, Could Not Be Checked and Decided — each with its count. Above them: how far the catalogue is, and what this month cost.
A card offers only what fits its status:
| Action | Where | Effect |
|---|---|---|
| Accept | Proposal, Review Text | Writes the sentence onto the picture, source AI. Refused when the picture's file or alt text changed since the check. For Review Text only after a correction. |
| Edit | Proposal, Review Text | Corrects the sentence right in the card. It is checked again on saving: without a finding it becomes a proposal; with one it stays Review Text and can still be accepted. Saving without a change is no correction. |
| Reject | Proposal, Review Text | Discards the proposal. The picture is proposed for again only when its file changes or somebody drafts again. |
| Picture Fits | Review Image | For a false alarm: confirms that the picture shows the item and drafts the sentence right away. The confirmation holds for this file on this item. |
| Leave as It Is | Review Image | Takes the picture off the list without a sentence. |
| Draft Again / Try Again | everywhere | Checks all pictures of the item again and writes new sentences, replacing rejections and corrections too. |
| Open Item | per item | Opens the item — to remove a wrong picture, or to correct wrong item data. |
Page actions:
- Accept Proposals on This Page writes every proposal on the page, as it was shown: a sentence the background job replaced in the meantime is left out. Held-back pictures and sentences that broke a rule are never part of it.
- Checked Pictures shows the same pictures as a list — to filter, sort, export to Excel, or decide many rows at once. Sentences changed there are checked again too.
Nothing gets past the hold-back
A picture in Review Image has no sentence, and no action of the review writes one under it. When a later run checks a picture again and holds it back, a sentence corrected by hand but not yet accepted is dropped as well.
When the picture is right and the data is not
A white scanner with the colour attribute "anthracite" ends up in Review Image although the picture is right. Correct the attribute on the item then; the next run checks again.
On the marketplace item card
Below the image gallery, the section Alt Texts from the Picture shows the cards of every picture of this item, decided ones included — with the same actions. Open Alt Text Review shows them on a page of their own with more room.
When it runs
- For one item: marketplace item card → Draft Alt Texts with Copilot. The review then opens for that item; the cards also stand below the gallery of the item card.
- For the catalogue and new pictures: the alt text job queue entry, created in the setup. It runs every quarter of an hour, takes up to Pictures per Run pictures and continues on the next run. A picture is due when it has never been checked or its file changed. After three failed items in a row the run ends and tries again next time.
Setup
| Field | Meaning |
|---|---|
| Draft Alt Texts from Pictures | Switches the background job on. Off by default: the first catalogue run costs real money. |
| Picture Size for the AI (px) | The longer side pictures are scaled down to. Default 1024. |
| Picture Detail for the AI | High reads small print — which is what catches a wrong type number. Low costs a fraction. |
| Replace Existing Alt Texts | Off by default: pictures with an alt text are only checked for whether they fit and keep their sentence. |
| Accept Flawless Proposals Automatically | Off by default. Not recommended for the first catalogue run. Held-back pictures and sentences that broke a rule always wait for a person. |
| Pictures per Run | How many pictures one run of the job checks at most. |
Cost
Measured in the pilot with gpt-6-sol at 1024 pixels: about 0.3 cents per file for the analysis and around 0.8 cents per item for the fit check and writing. A new picture therefore costs about 1.2 cents, including the new sentences for its item. Every call appears in the AI usage log with model, picture count and cost, and counts against the token and the cost budget.