Ecommerce MCP prompts: how to talk to your store’s data
Ecommerce teams have plenty of data. The challenge is getting from that data to a useful answer without sending a request to the data team and waiting days for an analysis. MCP servers help solve that by connecting LLMs to analytics solutions like Mixpanel, so ecommerce teams can query their data in natural language.
With MCP, the right prompt turns a data request into a conversation. You can map your data, measure a behavior, investigate a change, and connect the answer to context from the rest of your stack in real time.
To help you get started, here is a practical list of ecommerce prompts from our Mixpanel MCP Prompt Library to use with AI and your MCP server.
(Note: The prompts below assume MCP is already set up. They’re written with Mixpanel MCP server in mind, but will work with any connected analytics solution. For setup help, check out our docs.)
1. What’s in my store’s data?
Before running ecommerce analytics, map your project’s schema. You don't want to build a query around an event called “Purchase Completed” if your project actually calls it “Order Completed.” Wrong event names, bad filters, and missing properties can all break your analysis.
To better understand your store’s data, start with these ecommerce MCP prompts:
- "What Mixpanel projects do I have access to? List them with their IDs."
This question is a good starting point if you work across multiple projects.
What comes back: A project list so the AI can target the right dataset.
- “I want to understand completed purchases in this project. Suggest the top 3–5 candidate events that represent an order, and how often each fires.”
This question maps schema to a business question.
What comes back: Candidate events ranked by volume, helping you identify the real purchase event.
- “What properties are available on [Order Completed]? Show actual values for product category, payment method, and discount code.”
Asking this allows you to explore properties and their actual values.
What comes back: The actual property names and values available on the event, so you don't build a breakdown around a property that doesn't exist or use a value that doesn't match.
You can also use MCP to ask AI to suggest an activation signal:
- “What’s the best activation event for a first-time shopper? Share the early action that predicts they’ll come back and buy again.”
What comes back: A suggested activation signal to use as the starting point for retention analysis.
Why does an ecommerce team need to orient the AI first?
Before asking analytical questions, have MCP map the events and properties in your Mixpanel project. It can't guess that your purchase event is called "Order Completed" or that category values are lowercase. A quick schema check helps prevent prompts from returning results for the wrong event.
2. Where am I losing revenue?
Once you know what your events and properties mean and you’re confident that your AI does too, measure the behavior that matters. You can use these prompts to analyze funnels, retention, feature adoption, and trends.
A good ecommerce LLM prompt specifies four things: behavior (which events), population (who), timeframe (when), and answer shape (rate, trend, or breakdown). Leaving one of these out forces AI to fill in the gap.
For example: product-view-to-purchase conversion, mobile shoppers, the last 30 days, and a conversion rate.
- “What’s the conversion rate from product view → add to cart → purchase for [mobile shoppers] over the last 30 days? Show me where the biggest drop-off is.”
What comes back: A step-by-step funnel with conversion at each stage. You can go deeper by asking for a breakdown of the information gathered: “Break that down by [plan type/device/channel].”
- “Show week-over-week checkout funnel trends for eight weeks and flag any step where conversion drops more than 10%.”
What comes back: A trended funnel with the weakening step surfaced automatically. The flagging is AI reasoning over the query results, rather than a separate tool call.
- “What’s the 1-, 7-, 14-, and 30-day repeat-purchase rate for customers whose first order used a discount code vs. full price?”
What comes back: Two retention curves compared, giving you a direct view of repeat purchasing among discount and full-price customers so you can see how their behaviors differ.
- “Compare this month to last month across revenue, add-to-cart rate, and 30-day repeat-purchase rate. Summarize what improved, what got worse, and what stayed flat.”
What comes back: A compact comparison of the three metrics, showing how each changed over the 30 days.
What makes a good analytics prompt for an online store?
A reliable ecommerce MCP prompt names the behavior, population, timeframe, and answer shape. That gives AI enough information to measure the intended behavior without making assumptions about which shoppers, dates, or output you mean.
3. Who’s dropping off, and why?
The prompts in the previous section help you surface friction points and areas to investigate. Once you have that, use these prompts to zoom in on specific users, accounts, and sessions, moving from aggregate data to individual behavior. You can also pair these prompts with a solution like Session Replay to witness the user experience firsthand.
Here are the prompts that will help you investigate the user experience more deeply:
- “Show the distinct IDs of shoppers who added to cart but didn’t purchase in the last 14 days, then pull their session replays.”
What comes back: A list of shoppers who abandoned their cart and replay metadata and links for their sessions. Mixpanel opens recordings in the UI, so you can see exactly what happened.
- “Pull the last five sessions for customer [distinct_id]; show which products they viewed, their cart value, and where they exited.”
What comes back: A session-by-session view of an individual shopper's journey.
- “This high-value customer was active last month but hasn’t returned in three weeks. Pull their last 30 days of activity and replays.”
What comes back: An individual early-churn investigation combining recent activity with session replays.
When you don't have a shopper ID, you can also use a two-step process to identify relevant users:
- “Step 1: Show me the IDs of users who reached [Step N] but didn’t complete [Final Event] in the last 7 days. Return up to 10. Step 2: Pull session replays for those users.”
What comes back: A population of users who reached the specified step without completing the final event, along with replay metadata and links for those users.
How do ecommerce teams find out why shoppers abandon carts?
Use MCP to identify shoppers who added to cart but didn't purchase, then pull their session replays. Instead of stopping at an abandonment rate, you can investigate the behavior around the moment shoppers left.
4. How do I see it all at a glance?
Once you've found an analysis worth monitoring, turn it into a dashboard. Dashboards created through MCP are real, saved Mixpanel Boards that are accessible for the whole team, unlike temporary query results from a conversation.
- “Create a dashboard called ‘Store Health’ with: daily revenue, a product-view → purchase funnel, and 30-day repeat-purchase retention. Add a text card at the top summarizing what this board tracks.”
What comes back: A live Mixpanel dashboard built from one prompt.
- “Build a weekly merchandising dashboard tracking add-to-cart rate and low-stock flags by product category for the last 12 weeks.”
What comes back: A category-level view that a merchandising team can use to monitor engagement and inventory.
- “Duplicate my ‘Store Health’ dashboard, change the range to BFCM week last year, and rename it ‘BFCM 2025 Recap.’”
What comes back: A seasonal comparison without having to rebuild the dashboard from scratch.
Can MCP build a Mixpanel Board from a prompt?
Yes. With Project Owner, Admin, or Analyst access, an ecommerce team can describe a dashboard in plain English and have MCP create it as a saved Mixpanel dashboard. Unlike ad hoc query results, dashboards persist for the team.
5. Is my data trustworthy?
The fastest way to get a bad answer from an AI system is to give it bad data. MCP can help ecommerce teams audit Lexicon entries and data-quality issues before they affect an analysis.
- “Find all checkout-related events without Lexicon descriptions, suggest clear descriptions based on the event name and properties, then apply the descriptions.”
What comes back: A more documented event taxonomy. Always review AI-generated descriptions before confirming them, as AI can guess wrong on internal shorthand.
- “Flag any properties across events that might contain PII (email, shipping address, phone number) but aren’t marked sensitive yet.”
What comes back: A preliminary privacy audit. It can catch obvious cases but may miss obfuscated or custom fields, so it shouldn't be treated as a complete PII audit.
- “Before I run the checkout funnel, check for open data-quality issues on [Add to Cart] and [Order Completed]. Summarize what's broken, when each was detected, and which to fix first.”
What comes back: A check on the data feeding your critical funnel before you trust its results.
- “Tag every event related to checkout with ‘Checkout Flow,’ create the tag if it doesn’t exist so the team can find them.”
What comes back: A tidy, searchable Lexicon.
Why check data quality before analyzing ecommerce data with AI?
An AI will answer confidently even when an underlying event is broken or mislabeled. Asking MCP to surface data-quality issues on your events before running the funnel catches problems like duplicate order events or missing category values that would otherwise turn a wrong number into a business decision.
6. How do I connect the rest of my stack?
With MCP, you can combine product analytics data with attached files, external context, or other connected solutions to add the context needed to get the full picture. These cross-system questions are where MCP server analytics becomes particularly useful. Behavioral data can be combined with ad spend, inventory, or customer feedback to answer questions that previously required custom data pipelines.
Note: Each external source needs its own MCP connection in the same AI client.
When you use MCP to combine data from Mixpanel and an attached file:
- "Checkout conversion dropped between [dates]. I've attached our promo calendar — compare the daily conversion rate against campaign start and end dates and tell me which promotions overlap the drop."
What comes back: A comparison between the metric change and events documented in the calendar.
When you use MCP to combine data from Mixpanel and error monitoring (Sentry, Datadog):
- "Pull shoppers who triggered [Payment Failed] more than twice during checkout in the last 7 days, broken down by payment method. Then check [Sentry] for exceptions on the checkout service in the same window. Same root cause, or separate bugs?"
What comes back: a preliminary error report tying failed checkouts to the exceptions behind them.
When you use MCP to combine data from Mixpanel and team messaging (Slack, Teams):
- Pull my top three drop-off points in the checkout funnel over the last 30 days, then search [#cx-feedback] for shopper complaints mentioning shipping cost, payment, or discount codes. Summarize the quantitative and qualitative signals together."
What comes back: Quantitative drop-off data paired with qualitative feedback that can help explain it.
What cross-system questions can ecommerce teams answer with MCP?
By connecting Mixpanel with authorized sources such as ad platforms, inventory files, or Slack, ecommerce teams can combine behavioral data with the context needed to explain it. That can mean true CAC by channel, restock priorities, or the reasons behind a checkout drop-off.
The ground rules for better ecommerce prompts
You don't need to memorize every prompt in the library. Follow these principles to get more reliable results from your ecommerce data analytics:
| Rule | What it means |
|---|---|
| Rule 1: Specify four things every time | The behavior (which events), the population (who), the timeframe (when), and the shape of the answer (rate, trend, breakdown). Leave any out, and the AI fills in defaults that may not match what you wanted. |
| Rule 2: Start with schema discovery | Before running funnels or retention queries, ask MCP which events represent the behavior you care about. It reads your actual project schema, so you don't have to guess event names. |
| Rule 3: Check property names and values before breakdowns | Add "first confirm what properties are available on this event" to avoid silent failures. Checking actual values avoids filters that match nothing. |
| Rule 4: Use Flows for discovery, Funnels for measurement | Funnels measure the rate between steps you already know. Flows discover the steps you don't. When conversion is low and you're not sure why, Flows shows what users are doing instead. Use Sankey chart type for steps before queries. |
| Rule 5: MCP can't access saved segments or cohorts | Express population filters as event or user properties. "Users where plan_type = enterprise" works. "MyEnterprise cohort" doesn't. |
| Rule 6: Follow up in the same conversation | MCP retains context within a session. "Break that down by plan type" or "compare that to last month" works without restating everything. |
| Rule 7: Check data quality before critical analysis | A broken event produces confident-looking wrong answers. Ask about data-quality issues first on any event you haven't used recently. |
| Rule 8: Results don't save back to Mixpanel (except dashboards) | Query results in a conversation are temporary. Recreate the report in the Mixpanel UI or use Create-Dashboard to persist it. Dashboards created via MCP are real, persistent Mixpanel dashboards. |
| Rule 9: Search-Entities finds more than reports and dashboards | It also discovers experiments, feature flags, metric trees, playlists, heat maps, and cohorts, all useful for auditing project hygiene. Drill into full details with Get-Dashboard or Get-Report. |
| Rule 10: Cross-tool workflows require multiple MCP connections | To chain Mixpanel data with Slack, Jira, Notion, or Sentry, those MCP servers must be connected in the same AI client and authorized. |
| Rule 11: Write operations need Project Owner or Admin role | Editing events, properties, tags, and dashboards requires the right permissions. If a write operation fails silently, check your role first. |
| Rule 12: The rate limit is 600 requests per hour | Heavy analysis sessions can approach this limit. If calls start failing, wait and retry. For 20+ query sessions, break them into multiple sessions. |
The result is a more practical way to work with ecommerce data. Instead of sending a question away and waiting for a custom report, you can ask, refine, investigate, and follow the data wherever it leads you.
Connect the Mixpanel MCP server and try these prompts, or see how other ecommerce teams use Mixpanel.