This playbook was originally published on G2 AI blueprints.
What problem does this solve?
Shoppers with a question about sizing, compatibility, ingredients or delivery dates often leave instead of asking. Basic chatbots answer from a static FAQ, live agents aren't online at 11pm, and email replies arrive after the shopper has bought elsewhere. Every unanswered question is a lost sale, and the questions themselves rarely make it back to the team that writes the product pages.
How does it work?
- Install Yuma Sales AI on your Shopify storefront (a two-click install). It runs as its own widget on product pages and reads your live catalog, inventory, product specs, reviews, policies and past support history, so there's no FAQ to build. Anything that needs a human escalates to your helpdesk with the full conversation.
- Add sales guidance to your Yuma Handbook: brand voice for pre-purchase conversations, when to recommend add-ons, and discount rules (who can get a code, how much, and how often).
- Connect Claude to Yuma through the Yuma MCP and run the pre-purchase prompts on the last 30 days of chat conversations. Find the top questions by product and the ones that most often end without a purchase.
- Fix your product pages. Add the answers to the most common questions (size guides, compatibility, ingredients, delivery dates) so fewer shoppers need to ask at all.
- Write tests for the tricky cases before you go live: an out-of-stock variant, a compatibility question your catalog can't answer (the agent must not guess), a shopper fishing for a discount, and someone stuck between two sizes.
- Let the built-in split test run. Every Sales AI deployment is tested against a control group without Sales AI by default, so after two to four weeks you can compare conversion rate, revenue per visitor and average order value against shoppers who never saw it.
- Review every week: conversations that ended without a purchase, answers the agent was unsure about, and how discount codes were used.
- Expand once results hold: turn on cart recovery nudges and product recommendations. If your store has no support chat yet, add Yuma Chat AI so the same assistant handles order questions and spots upsell moments mid-conversation, all from one Handbook.
What's the biggest win?
Shoppers get an accurate answer in seconds at the moment of doubt, instead of leaving the site to search or waiting on an email reply. In Yuma's A/B tests, merchants have seen up to an 18% lift in revenue per visitor and up to a 6% lift in conversion rate. The questions shoppers ask also become a prioritized list of product page fixes for the merchandising team.
What's required to run this?
Never let the agent guess on compatibility, allergens, ingredients or sizing beyond what your catalog says; route those to a human with the full conversation. Cap discount codes and make them single-use so the widget doesn't train shoppers to ask for money off. Read results from the control group comparison rather than before and after, so seasonality doesn't distort them, and split results by desktop and mobile. Try it on your own store with a 30-day free Yuma trial.
What are the constraints?
Answers are only as good as your catalog data; missing specs lead to escalations instead of answers. Lift varies with traffic, category, price point and how complete your product pages are, and the figures above are strong results rather than guarantees. Discounts can pull forward sales that would have happened at full price. Brands with regulated claims, such as supplements or skincare, need strict Handbook rules on what the agent may say.
The prompts for this playbook
Use these with Claude connected to Yuma through the Yuma MCP. Replace everything in [brackets].
Prompt 1: what shoppers ask before buying
Search every chat conversation in Yuma from the last [30] days that happened before a purchase (pre-sales questions, not order support).
Return:
1. The top 15 question types by volume (for example sizing, compatibility, ingredients, delivery dates, returns, payment options, stock), with count and share.
2. For each question type, the top 3 products it's asked about.
3. Two conversation IDs and one short shopper quote per question type.
Prompt 2: questions that cost sales
From the same conversations, and using order data where available, find the question types that most often end without a purchase in the same session or within [48] hours.
For each, tell me whether the agent answered fully, answered partly, said it didn't know, or escalated. Flag any answer that looks wrong or unsure, with conversation IDs so I can check.
Prompt 3: product page fixes
Using the results of Prompts 1 and 2, list the product pages that would benefit most from new content. For each page give:
- the product
- the questions shoppers keep asking about it
- the exact information missing from the page (size guide, compatibility list, ingredient list, delivery estimate)
- a suggested short answer we could add, based only on what our catalog and policies already say
Rank by number of questions. Do not invent product facts.
Prompt 4: tests before going live
Write [10] tests for our pre-purchase conversations, using real shopper questions where possible. For each give the shopper message, the expected reply in our brand voice, the expected action (answer, recommend, offer code, escalate) and why the case matters.
Include at least:
- an out-of-stock size or variant
- a compatibility or allergen question the catalog can't answer (expected: no guessing, escalate)
- a shopper asking for a discount who doesn't qualify
- a shopper stuck between two sizes
- a shopper asking about delivery before a holiday deadline
Prompt 5: weekly sales chat report
For the last 7 days of pre-purchase conversations in Yuma, report in under 200 words:
- conversation count and the top 5 question types
- conversations that ended without a purchase, and the most common reason
- answers flagged as unsure or escalated
- discount codes issued and used
- one conversation that went really well and one to learn from
Rules to keep
- The agent never guesses on product facts. Missing data means escalation, then a product page fix.
- Measure lift against a control group, not before and after.
- Save these prompts in a Claude project or skill so anyone on the team can rerun them.
