Shopify Image Descriptions: $1,500 Catalog Sprint

Disclosure: this article contains affiliate links. If you open an account through one of them, Cashflow Abroad may earn a referral commission at no extra cost to you.

A 2,000-image Shopify catalog can hide hundreds of blank, duplicated, or useless alt attributes, yet a disciplined audit can be sold for $1,500 and delivered in about 10 operator hours. The offer is not “AI accessibility compliance.” It is a controlled inventory, draft, review, and import service that gives a merchant measurable coverage and a human-approved change log while you earn in dollars from abroad.

This is a narrow service for Shopify agencies and established merchants, not a generic AI side hustle. It fits the practical workflows in AI Income & Cash Flow: sell a defined outcome to a strong-currency buyer, keep infrastructure portable, and avoid storing more client data than delivery requires.

What exactly should you sell?

Sell an evidence-backed catalog remediation sprint

Sell a fixed-scope “product image alt-text audit and remediation sprint.” The buyer receives a baseline export, a rule set, AI-assisted drafts, a human-reviewed approval sheet, a controlled Shopify update, and before-and-after metrics. Do not sell a legal certification, an accessibility overlay, or a promise that alt text alone makes a store conform to WCAG.

The best initial buyer is a US Shopify agency managing stores with 500–10,000 product images. The agency already understands catalog access, can resell the work, and may bring several stores. A secondary buyer is a direct-to-consumer merchant with frequent launches, multiple product angles, and no catalog operations specialist.

Put hard boundaries in the proposal

Define one store, one language, up to 2,000 product images, one review round, and one approved import. Exclude theme graphics, blog images, video, PDFs, translation, and legal conformance testing unless separately quoted. A merchant must approve the writing rules and final rows before anything is written back.

Offer three tiers: $750 for an audit-only report up to 2,000 images; $1,500 for audit, drafting, review, and import; and $2,500 for up to 5,000 images plus a 30-day new-image monitoring workflow. Price by governed inventory and review burden—not by pressing a button.

What tools do you need, and what do they cost?

Use client-owned access and a capped API budget

As of August 2026, the lean version has no mandatory monthly software subscription beyond metered model usage. Use a client-approved Shopify custom app or collaborator access, a local CSV or private GitHub repository for versioned rules, and a small Worker to queue batches. Keep API keys in secrets, never in a spreadsheet or source file.

Tool	Current price or limit	Role	Constraint
Shopify Admin GraphQL API	Included with the client’s Shopify store	Read image IDs, current alt text, product context, and write approved changes	App scopes, API versioning, and rate limits require testing
GPT-5.4 mini API	$0.75/M input tokens; $4.50/M output tokens	Classify images and draft structured candidates	Image inputs and generated tokens are metered; human review remains required
Cloudflare Workers	$0 with 100,000 requests/day; paid starts at $5/month	Queue, validate, and log batch jobs	Free plan allows 10 ms CPU per invocation; paid tier is safer for production batches
n8n Cloud Starter, optional	€20/month billed annually; 2,500 executions	No-code scheduling and approval notifications	One workflow run counts as one execution; do not trigger once per image
GitHub Free	$0; unlimited private repositories with limited features	Store code, schemas, and non-sensitive test fixtures	Never commit store tokens or raw customer exports

The official OpenAI API pricing page lists token rates, while Cloudflare’s Workers pricing lists the free allowance and $5 paid minimum. Model cost per image varies with dimensions and prompt/output length, so quote a budget ceiling rather than inventing a universal per-image rate.

Exact startup and monthly cost math

For a first paid pilot, budget $5 for Workers Paid and cap OpenAI usage at $10: exact cash ceiling, $15 for the month. GitHub Free is $0, Shopify is client-owned, and a local approval sheet is $0. If you add n8n Cloud Starter, the fixed stack becomes $25 plus the €20 subscription; because currencies move, keep euros separate rather than disguising the conversion as a fixed dollar amount.

Do a 50-image calibration before processing the catalog. Record input tokens, output tokens, failures, and total model charge; multiply the measured cost per accepted image by the remaining inventory, then add a 25% retry reserve. Stop the job automatically at the $10 pilot ceiling and ask the client before increasing it.

How should the AI decide what alt text to write?

Context beats object recognition

Alt text depends on what the image does on that page. The W3C WAI alt decision tree says informative images need a brief description that conveys their meaning, functional images should communicate the action, and decorative or redundant images can use an empty alt attribute. Therefore “describe every image” is the wrong prompt.

For a product gallery, provide the model with the product title, product type, variant names, image position, existing alt text, and whether nearby visible text already communicates the same fact. Ask it to return a decision—informative, functional, decorative, complex, or manual review—before it drafts any wording.

Use a structured developer prompt

You are drafting candidate alt text for a Shopify product image.
Use only supplied catalog facts and visible evidence.
First classify: informative, decorative, functional, complex, or review.
Return JSON: classification, candidate_alt, confidence, reasons, flags.
Use an empty candidate_alt for decorative or redundant images.
Do not add SEO keywords, “image of,” subjective praise, or unsupported claims.
If confidence is below 0.85 or facts conflict, classification must be review.

Add client rules after the invariant rules: preferred product naming, color vocabulary, variant order, maximum working length, and banned claims. Treat length as an editorial preference, not a false legal threshold. The reviewer should judge whether the text provides equivalent meaning in context, not whether it hits a character target.

How do you deliver the service safely?

Run an export-review-writeback workflow

Shopify’s current GraphQL Image object exposes altText as the phrase describing an image’s nature or contents. Build around stable media IDs rather than filenames, which can be duplicated or transformed.

Step	Operator action	Acceptance gate	Artifact
1. Authorize	Obtain least-privilege, time-bounded store access and written scope	Client identifies approver and rollback window	Access record
2. Inventory	Export media ID, product ID, URL, position, context, and current alt text	Counts reconcile with Shopify	Read-only baseline CSV
3. Segment	Flag blanks, duplicates, filenames, keyword stuffing, and likely decorative images	Rules reviewed on 50 representative rows	Issue report
4. Draft	Send minimal context and image to the model in rate-limited batches	JSON schema passes; low confidence routes to review	Candidate CSV
5. Review	Human checks meaning, catalog facts, repetition, and risky inferences	Every changed row has approved status and reviewer	Signed approval sheet
6. Write	Update only approved media IDs in small batches	API response logged; errors do not retry indefinitely	Mutation log
7. Verify	Re-query changed IDs and sample storefront rendering	Written values match approvals	Post-change report
8. Roll back	Restore original values if the client rejects the batch	Rollback sample verified before access removal	Final archive

Keep the original and proposed values side by side, hash the export, and assign a run ID. Never bulk-write directly from raw model output. Use idempotency in your own job table so a timeout does not apply the same mutation twice or confuse the audit trail.

Starter path versus operator path

A beginner should export a small catalog, draft in 50-row batches, and have the merchant import or approve changes manually. An advanced operator can use a Worker queue, JSON schema validation, per-store secrets, and automated re-query verification. The service’s defensibility is the governance layer—rules, exceptions, review, evidence, and rollback—not access to a model.

How do you find and close buyers?

Lead with a five-image sample, not fear

Choose stores with large visual catalogs and visibly weak image descriptions. Send the agency or merchant a five-row sample showing current value, issue type, candidate, and why the candidate is more useful. Do not threaten lawsuits or claim that a scanner proves noncompliance.

Use this offer sentence: “We inventory every Shopify product image, draft context-aware alternatives, route uncertain rows to a human, and update only what your team approves.” To test US-side demand, post the fixed-scope offer as a free listing on Brixaz and measure replies before building a full automation stack.

If you operate through a US LLC from abroad, Mercury Bank can be relevant for business banking, but availability depends on the company, owner, and supported-country rules. Banking infrastructure does not change where you owe tax; pair operational setup with the broader guide to running a US business while abroad and get jurisdiction-specific advice.

Make the proposal auditable

State the baseline image count, excluded asset types, languages, approval owner, retention period, access removal date, and maximum API spend. Collect 50% upfront, 50% after the approved import, and price scope changes by added images or added languages. A monthly $350–$750 monitoring retainer can cover new products, a weekly exception report, and one review batch; it should not silently rewrite the live catalog.

What can go wrong?

The expensive failures are operational

• Confident fabrication: the model invents material, shade, size, or compatibility. Require catalog evidence and a conflict flag.
• Redundant noise: every gallery angle repeats the product title. Review the image’s distinct purpose in sequence.
• Decorative over-description: generated prose creates screen-reader clutter. Preserve empty alt attributes when the W3C decision path says the image is decorative or redundant.
• SEO stuffing: the client requests keywords in every row. Explain that the deliverable is useful alternative text, not a keyword field.
• Wrong-object writeback: filenames or URLs shift. Join and mutate by Shopify media ID, then re-query.
• Access leakage: store tokens appear in logs or repositories. Use secret storage, redact logs, rotate credentials, and remove access after acceptance.
• Legal overclaim: an audit is marketed as ADA or WCAG certification. Promise the stated process and artifacts only; refer legal conclusions to qualified counsel.

Minimize the data surface

Product images are often public, but drafts, unreleased products, access tokens, and analytics may be confidential. Put subprocessors and model use in the statement of work, send only necessary fields, define deletion dates, and honor the client’s data-residency requirements. If the client forbids third-party model processing, switch to an approved environment or decline the automation portion.

Conclusion: sell controlled change, not AI text

Your first seven days

Day one: create the five-column sample and decision rules. Days two and three: audit five prospective stores without changing anything and contact agencies with a five-image example. When one buyer agrees, run a paid 50-image calibration, publish measured costs and exceptions to the approver, then process the remaining catalog only after the rule set is accepted.

The portable-income advantage is straightforward: client value is tied to a governed catalog outcome, while the core tooling can cost $15 for the pilot month. One disciplined $1,500 engagement can fund the stack many times over; careless automation can damage a catalog just as quickly. Make the human approval and rollback artifacts part of what the client buys.

Data notes / Sources checked