# AI for Ecommerce: practical workbook Free training: https://ecomcoach.xyz/ai-ecommerce/ Real prompts. Real projects. Real results. ## Project brief Store / product: Problem: Baseline and date: Evidence files: Owner: Budget / time ceiling: Success measure: Review date: ## 1. Choose your ecommerce AI toolkit Build a small toolkit around one store problem, with a clear owner and review step. ### Inputs One repetitive task, your current apps, a representative input file and the output you need. ### Steps 1. Write the job before choosing software: for example, turn anonymized support questions into five product-page improvements. Record current completion time and the mistakes you want to avoid. 2. Use Claude for writing, analysis and multi-step work. Claude Cowork and Design capabilities are being brought into the main Claude experience; check what your account exposes. Treat “Claude Work” as a description of work in Claude, not a separate verified product. 3. Use Grok Bot for a bounded research or coordination role. Its marketplace includes SEO & AEO Desk and QA roles; inspect each bot’s instructions and access before adding it. Installing a bot does not validate its output. 4. Use Figma for shared design and interactive prototypes, Shopify Sidekick for supported admin tasks, and Higgsfield for creative production. A general assistant can work from exported files when a native connection is unavailable. 5. Use ChatGPT with GPT-6 Astra for a store-analysis or planning task when it is available in your account. Supply the source files, the question, the required deliverable and the checks it must pass. For example, turn an anonymized product-and-sales export into a weekly store brief, then verify its calculations and proposed actions. 6. Test one task three times. Compare factual accuracy, edit time, permissions and cost. The model handles reasoning; access to store data and connected tools depends on your account and the permissions you grant. ### Prompt Help me choose tools for [task]. Inputs: [files]. Current stack: [apps]. Required result: [deliverable]. Compare doing this manually, with my existing assistant, and with a specialist tool. State required access, review steps, unknown capabilities and a small pilot. Do not invent integrations or prices. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] Every proposed connection exists in your account. - [ ] The pilot uses minimized customer data. - [ ] You can identify where each factual claim came from. Reviewer: Decision: Next action: ## 2. Identify personas from customer evidence Create a buyer brief grounded in real purchasing situations. ### Inputs Anonymized interviews, product reviews, support tickets and customer questions. ### Steps 1. Gather a small, diverse sample: buyers, people who returned products and shoppers who did not buy. Remove names and order details before uploading. 2. Ask AI to group jobs, triggers, objections and desired outcomes. Require a source ID for each observation and keep minority views visible. 3. Write personas around a situation and buying task. Demographics alone rarely explain why someone needs the product. 4. Interview customers about a recent purchase: what happened, what alternatives they considered and what nearly stopped them. Use those answers to challenge the AI’s draft. 5. Create a one-page brief with language to use, proof needed, objections to answer and evidence still missing. ### Prompt Analyze these anonymized records [records]. Group buying situations, triggers, desired outcomes and objections. Cite record IDs. Separate observed evidence, inference and unanswered questions. Do not invent customers, quotes, demographics or market size. Draft two buyer briefs and five interview questions. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] Quotes match the originals. - [ ] A human has checked the clusters. - [ ] The brief includes contradictory evidence. Reviewer: Decision: Next action: ## 3. Validate product ideas before buying inventory Decide what evidence you need before investing in a product. ### Inputs Two candidate products, supplier quotes, competitor pages and customer conversations. ### Steps 1. Write the problem, proposed customer and reason to choose your offer. Record what would disprove your idea. 2. Ask AI to compare alternatives using a source-backed evidence table. Label missing demand evidence rather than filling it with invented search volumes. 3. Model landed cost, fees, fulfillment, discounts, returns and acquisition assumptions in the order calculator. A popular product can still lose money. 4. Test a transparent concept page, customer interview or small sample batch. Do not imply that stock is available or collect orders you cannot fulfill. 5. Decide in advance what would justify a next test. Keep expressions of interest separate from purchases and repeat purchases. ### Prompt Compare [idea A] and [idea B] using only [evidence]. Produce a table of customer problem, alternatives, demand evidence, costs, supplier risk and unknowns. Recommend the cheapest test of the biggest uncertainty. Do not predict revenue or declare a winner without evidence. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] Supplier costs have dates. - [ ] Interest and paid demand are reported separately. - [ ] You have a stop rule and a test budget. Reviewer: Decision: Next action: ## 4. Use Shopify AI and dropshipping tools carefully Create accurate product drafts and a supplier verification workflow. ### Inputs Verified product specifications, supplier terms, sample photos and current shipping estimates. ### Steps 1. In Shopify admin, open Sidekick and ask it to explain a store question or prepare supported content. Review the actual proposed output before applying changes. 2. Use a supplied specification sheet to draft a product title, benefit-led description and FAQ. Mark missing dimensions, materials and compatibility for human confirmation. 3. Evaluate dropshipping tools such as AutoDS by supported suppliers, draft imports, synchronization behavior, fees and error handling. Its product-finding tools can surface candidates; they do not establish profitability. 4. Order a sample and test quality, tracking, packaging and the return process. Recalculate margin using the real delivery experience. 5. Keep imported products as drafts until descriptions, images, inventory and shipping promises are verified. Create an exception queue for stock or price changes. ### Prompt Using only [verified specification sheet], draft a Shopify product page. Output title, benefits tied to facts, dimensions, care, shipping questions and missing information. Do not invent certifications, stock, delivery dates, customer reviews or performance claims. Return a supplier-checklist before publication. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] Sample quality has been checked. - [ ] Delivery promises match current supplier evidence. - [ ] Automated price and stock changes have limits and alerts. Reviewer: Decision: Next action: ## 5. Turn keyword research into a store content plan Map actual search language to useful category, product and advice pages. ### Inputs Search Console queries, site-search terms, customer questions and a dated keyword-tool export. ### Steps 1. Export queries with impressions, clicks, dates and landing pages. Keep country and device filters visible. 2. Ask AI to group terms by buying intent, product attributes and questions. Treat estimated volume from a keyword tool as an estimate; AI-generated numbers are not keyword research. 3. Map each cluster to one primary page. Compare current pages before creating new ones so several URLs do not compete for the same intent. 4. Choose long-tail topics where you can provide specific evidence: sizes, materials, use cases, comparisons and care instructions. 5. Write a brief with the customer question, answer, product facts, internal links and success measure. Review impressions, qualified visits and outcomes after publishing. ### Prompt Cluster this dated keyword export [data] by intent. Preserve original metrics and flag missing values. Map each cluster to [existing URLs]. Suggest one primary page per cluster, explain overlap and produce three briefs grounded in our actual products. Do not invent demand or ranking difficulty. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] Metrics have a named source and date. - [ ] Intent matches the destination. - [ ] New pages add distinct value. Reviewer: Decision: Next action: ## 6. Make your store understandable in search and AI answers Improve product facts, helpful answers and crawlable content without promising rankings. ### Inputs Your top product and category pages, Search Console and current catalog data. ### Steps 1. SEO helps search engines discover and understand pages. AEO focuses on answering questions clearly; GEO is a common term for visibility in generative answers. These overlap and none guarantees a citation. 2. Check indexing, internal links, canonical URLs and mobile page usability. Make the essential product information readable in the page itself. 3. Add direct answers about fit, materials, compatibility, shipping and returns. Use original photos and explain the basis for comparisons. 4. Keep product structured data and merchant feed values consistent with visible price, availability and policies. Validate markup; do not mark up invented reviews. 5. Monitor search queries, qualified referral visits and conversions. Use repeatable, dated AI-answer checks as observations, not comprehensive market-share measurements. ### Prompt Audit [page content and verified product data] for unanswered buying questions, vague claims and inconsistent product facts. Suggest clearer copy and internal links. Separate crawl/index checks from content changes. Do not promise AI citations, rankings or special AI schema. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] Visible facts agree with structured data. - [ ] Sources and policies are easy to find. - [ ] No promise of guaranteed AI inclusion. Reviewer: Decision: Next action: ## 7. Analyze traffic and ecommerce performance with AI Produce a weekly decision brief with reproducible calculations. ### Inputs Comparable GA4, store and advertising exports with metric definitions, currency and timezone. ### Steps 1. Choose the question first: for example, why did mobile purchase conversion fall? Define sessions, users, orders and revenue before combining exports. 2. Check missing dates, duplicate orders, tracking changes, consent effects, refunds and timezones. Do not sum attributed revenue across ad platforms. 3. Ask AI to show formulas and compare channel, device, landing page and new-versus-returning cohorts where available. 4. List possible explanations and evidence that would distinguish them. A traffic shift is not proof that a design change caused the result. 5. Recalculate a sample manually. End the report with one next investigation, one owner and a date. ### Prompt Analyze [exports] for [question]. First check definitions, missing data, date alignment and duplicate counting. Show calculations and denominators. Separate observed changes from causal hypotheses. Return three findings, confidence limits and one next test. Never invent missing rows. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] Denominators and dates match. - [ ] A sample of calculations has been reproduced. - [ ] The report labels hypotheses. Reviewer: Decision: Next action: ## 8. Build a marketing plan and PPC test with AI Create a focused campaign plan with a spending ceiling and a business outcome. ### Inputs Buyer brief, verified offer, contribution model, historical results and approved budget. ### Steps 1. Choose one product, buyer situation and business outcome. Describe the offer, supporting proof and landing page before selecting channels. 2. Use AI to translate the buyer’s questions into search ad groups or social creative angles. Keep keyword intent and negative-keyword suggestions for human review. 3. Calculate an allowable acquisition cost from contribution and desired retained margin. Use current platform forecasts only as estimates, not promised results. 4. Prepare a small creative test with one clear hypothesis. Check links, mobile checkout, conversion events, feed accuracy and campaign settings before launch. 5. Review spend, purchases, contribution, new-customer mix and returns. Define pause conditions before spending; budget changes require an owner. ### Prompt Build a two-week PPC learning plan for [product, buyer, offer]. Use [margin model] and budget ceiling [amount]. Provide keyword themes, exclusions, two creative hypotheses, landing-page requirements, measurement definitions and stop conditions. Keep all ads as drafts and mark unsupported claims. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] The spending cap is approved. - [ ] Conversion events have been tested. - [ ] ROAS is not treated as final profit. Reviewer: Decision: Next action: ## 9. Plan social content and product creative with Higgsfield Make a two-week content plan and a truthful product-video brief. ### Inputs Rights-cleared product photos, brand guidance, factual specifications and customer objections. ### Steps 1. Choose three content jobs: demonstrate use, answer objections and show how to choose. Match each post to a buying question. 2. Ask AI for hooks, scripts and a calendar with channel, format, source asset, CTA and review owner. Reuse a research insight across formats without publishing identical filler. 3. In Higgsfield, choose an available product-video workflow and supply approved photos or a product URL where supported. Brief framing, motion, duration and the exact facts allowed. 4. Check every frame for distorted labels, changed dimensions, extra accessories and unrealistic product behavior. Keep real demonstrations for claims that need evidence. 5. Test two creative concepts against a defined outcome. Label illustrative or synthetic demonstrations clearly; never present an AI person as a real customer testimonial. ### Prompt Create six social briefs for [product] using [verified facts and approved assets]. Include hook, shot list, on-screen copy, caption, CTA and customer objection addressed. Then write a Higgsfield prompt preserving product shape, colors, label and included items. Flag scenes that need real footage. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] Asset and likeness rights are clear. - [ ] The product remains accurate frame by frame. - [ ] Creative performance is measured beyond views. Reviewer: Decision: Next action: ## 10. Prototype better shopping experiences with Figma and Claude Design Create a testable product-page prototype from a real shopper problem. ### Inputs Screenshots, brand components, product facts and a prioritized usability finding. ### Steps 1. Write the customer task and the obstacle. For example: shoppers cannot tell whether a container fits their shelf. 2. Ask Claude’s available design capabilities or Figma Make for two ways to present the same verified information. Supply brand styles and content constraints. 3. In Figma, refine hierarchy, responsive layout, focus states, readable type and component behavior. Treat generated screens as a starting point. 4. Run a task-based session with representative shoppers. Ask them to find the dimensions and decide fit without coaching. 5. Give development an annotated specification, assets, acceptance checks and analytics needs. A prototype is not a working payment or inventory integration. ### Prompt Design two mobile product-page concepts for [task] using [facts, screenshots, styles]. Preserve claims and pricing. Show how a shopper finds fit, delivery and returns. Include loading, unavailable and error states. List accessibility and implementation checks. Do not simulate a successful purchase as a real transaction. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] Real shoppers attempted the task. - [ ] Keyboard and small-screen behavior are specified. - [ ] Developers can distinguish prototype behavior from integrations. Reviewer: Decision: Next action: ## 11. Plan conversion optimization and A/B testing with AI Write an experiment brief that can produce a credible decision. ### Inputs A funnel problem, baseline conversion, traffic estimate, event QA and business constraints. ### Steps 1. Identify the earliest relevant friction point using analytics and customer evidence. Fix broken functionality without calling the repair an A/B test. 2. Write a hypothesis linking one change to a shopper behavior. Pick a primary metric and guardrails such as order value, errors, margin and returns. 3. Estimate required sample size with a suitable calculator or analyst using baseline rate, minimum detectable effect, significance and power. AI can explain inputs; it cannot create sufficient traffic. 4. Randomly assign eligible users with stable allocation. Verify tracking, exclusions and sample-ratio balance. Run through the planned duration and business cycles; do not stop when a result first looks positive. 5. Report counts, effect size, uncertainty and guardrails. If traffic is too low, use interviews, usability sessions and sequential improvements, clearly labeled as non-causal observations. ### Prompt Draft an A/B test brief for [problem/change]. Baseline [rate], traffic [eligible users], primary outcome [metric]. Ask for missing sample-size assumptions. Specify unit of randomization, allocation, tracking QA, guardrails, duration and stopping rule. Explain what an inconclusive result means. Do not invent a statistically significant winner. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] The test was planned before launch. - [ ] AI-generated personas are not counted as user research. - [ ] An inconclusive result is reported honestly. Reviewer: Decision: Next action: ## 12. Build your marketing, product, design and development agent team Create a shared brief and a reviewable handoff system for one improvement. ### Inputs One customer problem, a task board, source evidence and explicit access boundaries. ### Steps 1. Appoint a human owner and a coordinator role. The coordinator breaks the goal into tasks and maintains the source brief; start with a few roles rather than many bots. 2. Assign research and marketing to buyer evidence and campaign briefs; product to priorities and acceptance criteria; design to prototype and assets; development to implementation and tests. 3. Use a shared task record: input, output, source links, owner, dependencies, status and acceptance check. Each handoff must identify what changed and what remains unknown. 4. Give QA a separate checklist: product facts, links, accessibility, analytics, mobile checkout and rollback. Having another agent agree is not independent proof. 5. In Grok Bot or another supported agent workspace, configure each role with the same brief and only needed tools. Where orchestration is unavailable, use separate conversations and explicit manual handoffs. Keep spend, publishing, customer messages and production changes under named human control. ### Prompt Act as coordinator for [store improvement]. Create tasks for research, marketing, product, design, development and QA. For every task list input files, output artifact, dependencies, acceptance checks and permissions. Do not fabricate progress. Escalate missing evidence. Return a handoff log and proposed next action; do not publish or spend. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] Each task has one accountable owner. - [ ] QA checks the actual artifact. - [ ] Unfinished work is not reported as done. Reviewer: Decision: Next action: ## 13. Use AI for retention, email and customer support Draft a useful customer journey with accurate policy answers and human escalation. ### Inputs Consent-based audience rules, product-care instructions, current policies and anonymized tickets. ### Steps 1. Map the customer’s next need after purchase: setup, care, delivery help or a complementary purchase. Avoid making every message a discount. 2. Ask AI to draft three messages with trigger, audience, exclusions, purpose and a single CTA. Keep marketing consent and service communications distinct. 3. Build a support knowledge set from approved policies. Require the assistant to cite the policy and route missing or exceptional cases to a person. 4. Test with edge cases: delayed shipment, return-window exception, damaged item and wrong product. Do not let a draft assistant promise refunds or delivery dates it cannot verify. 5. Measure repeat purchase, unsubscribe rate, complaints, resolution quality and margin. Review failures weekly before increasing automation. ### Prompt Draft a post-purchase journey for [product] from [approved policies]. For each message provide trigger, exclusions, helpful purpose and CTA. Then draft support replies to [anonymized tickets], citing the policy used. Flag exceptions for a person. Do not send, promise compensation or invent tracking details. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] Audience exclusions prevent contradictory messages. - [ ] Policy versions are current. - [ ] A person handles exceptions. Reviewer: Decision: Next action: ## 14. Build an AI-assisted inventory and operations review Turn stock and fulfillment records into an exception list. ### Inputs SKU-level stock, sales, lead times, returns and outstanding orders with a dated export. ### Steps 1. Normalize SKU identifiers and dates. Separate available, reserved and incoming units; record uncertainty in supplier lead times. 2. Ask AI to identify stockouts, unusual returns and late orders. Require the underlying rows and calculation for each alert. 3. Model reorder scenarios using demand ranges and lead-time variability. Historical averages can mislead when a promotion or stockout changes observed sales. 4. Assign each exception to an owner. Keep purchase orders, refunds and customer notifications as drafts until reviewed. 5. Compare alerts with actual events and update the rules. Track false alarms and missed issues, not just the number of tasks automated. ### Prompt Review [inventory and order exports]. Validate SKU joins and date ranges first. Produce an exception table with evidence rows, calculation, confidence, owner and proposed action. Show low/base/high demand scenarios. Do not place purchase orders or issue refunds. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] Units and dates are consistent. - [ ] Stockouts are not mistaken for low demand. - [ ] A human reviews financial commitments. Reviewer: Decision: Next action: ## 15. Your 30-day AI ecommerce improvement project Finish one documented improvement with evidence, an artifact and a measurement plan. ### Inputs One store problem, a baseline, access to approved data and a realistic time budget. ### Steps 1. Week 1: choose one problem and create the evidence pack. Run the persona or traffic lesson. Write the baseline and the outcome you want to influence. 2. Week 2: use the relevant lesson to create an artifact: a product-page prototype, campaign brief, support workflow or operations report. Keep source facts attached. 3. Week 3: review and test. Run a usability session, a controlled experiment if traffic supports it, or a limited operations pilot. Log failures and edits. 4. Week 4: compare the result with the baseline using compatible definitions. Explain uncertainty, costs and time spent. Decide whether to continue, change direction or stop. 5. Write a handoff with the final artifact, owner, review cadence and rollback plan. A completed draft is an output; a measured business improvement is an outcome. ### Prompt Help me run a 30-day improvement project for [problem]. Create a weekly plan with evidence needed, deliverable, owner, review checklist and outcome measure. Fit it to [hours/budget]. Use my actual data only. At the end distinguish completed work, measured results and unresolved assumptions. ### My deliverable ### Evidence and unresolved questions ### Review - [ ] The final artifact is usable. - [ ] Reported outcomes have evidence. - [ ] The next decision is explicit. Reviewer: Decision: Next action: