|
ECOM HEADS • August 18 • Read online
|
|
AI & ANALYTICS • 6 MIN READ
The model found the pattern we missed
|
|
Once your ad, analytics, email, commerce, POS, and content data are available in one analysis session, a model can look for relationships your channel reports never test.
|
|
Quick community update: the Ecom Heads Slack group went live yesterday 🥳. We've already had some great knowledge-sharing, a job posting for a great gig, and we're about to have a discussion about this issue! I anticipate that not everyone who I invited from last weeks applications will accept, so there may be some spots available. Apply below for a chance to join as a founding member and get free lifetime access.
|
|
|
|
Your dashboards see channels. Your business does not.
|
|
Most ecommerce reporting is organized by platform.
Google Ads tells you what happened in Google Ads. Meta explains Meta. Google Analytics reports on site activity. Klaviyo shows campaign and flow performance. Shopify, BigCommerce, or Adobe Commerce records your orders.
Each view is useful. The best questions usually cross those boundaries.
Which content changed store revenue? Which email campaigns brought in customers who later bought in person? Which paid channel creates customers with the fewest returns? Which product is profitable after discounts, shipping, support, and repeat purchases?
You could answer questions like these years ago. The work started with exports, spreadsheet cleanup, date alignment, attribution arguments, and a dashboard that needed to be rebuilt or refreshed next week.
Let's be honest. It was a huge chore, so it got pushed off.
The cost of running that analysis has changed.
|
|
Connect the sources, then ask a better question
|
|
An AI application can now work with tools from multiple connected systems in the same analysis session. MCP is one clean way to make those tools available to the model.
Google Analytics, Google Ads, Meta Ads, Microsoft Advertising, TikTok Ads, and Klaviyo all have official MCP options. Other sources may require an API, a warehouse, or a trusted custom MCP server. What matters is giving the model read-only access to the data needed for the question.
|
| CONNECT THE FULL OPERATING PICTURE |
|
Paid media: Google, Meta, TikTok, Microsoft, marketplaces
Site behavior: Google Analytics
Lifecycle: Klaviyo or another email and SMS provider
Commerce: Shopify, BigCommerce, or Adobe Commerce, including POS when available
Operations: returns, inventory, support, promotions, content activity
|
|
|
Once the connections, time zones, and metric definitions are set, the recurring review can begin with one strong prompt and a good model.
The unlock is the model's ability to inspect the whole business while it answers, instead of making you carry numbers from one tab to another.
|
|
A 15% to 25% pattern hiding in a bike retailer
|
|
One of my clients sells bikes on Shopify. They have five physical locations across the Northeast, and every store uses Shopify POS. Their online and in-store transaction data lives in the same system.
The locations are impressive bicycle showrooms. Shoppers can take test rides inside, over obstacles, and across different terrain with help from very knowledgeable salespeople.
The owner is also incredible at content.
Every week he posts videos comparing bikes, brands, tires, pedals, and other products. He has become a familiar face in the online biking community. Sometimes he tells viewers that he will be at a specific location on a specific date.
A few weeks ago, I ran a deep analysis across the available performance data. The model surfaced a relationship we had never thought to test.
|
|
THE MODEL'S FINDING
15% to 25% higher
Average revenue at the specific store on days the owner had announced he would be there, compared with a typical day.
|
|
|
We never would have put that together without the model finding the pattern.
The relationship was hiding between content, a person, a location, a date, and POS revenue. No individual channel report had a reason to surface it.
The result is a correlation, so I would not call it proof that the owner's appearance caused the lift. It was repeatable enough to build a strategy around and test more deliberately.
We are now more intentional about announcing when he will be at each store. Future appearances can give us a cleaner set of data to measure.
|
|
A prompt I would run
|
|
Act as a senior ecommerce analyst. Use every connected read-only source to analyze performance across paid media, analytics, email and SMS, ecommerce, POS, and content activity. Normalize dates, time zones, currencies, and metric definitions before comparing results.
Organic content may not be available through an MCP. Use your browser to visit the brand's Instagram, YouTube, TikTok, and any other supplied content channels. Review every relevant post from the analysis window. Watch the videos when your tools support playback. Otherwise, read the closed captions or transcripts along with the post copy, title, publication date, tagged location, and visible metadata. Build a dated content timeline. If you cannot access a video, captions, or transcript, record the gap and do not guess what the video says.
Cross-reference that content timeline with marketing and revenue data by store, product, customer type, and day. Look for repeated relationships that would be invisible inside one platform's reporting.
For every finding, show the source fields, comparison baseline, sample size, exceptions, and confidence level. Treat correlations as hypotheses. Recommend the next test that would confirm or reject each one. Keep every source read-only. Do not publish, post, comment, like, message, or change a campaign or account setting.
|
|
|
Set it up safely
|
| 1 |
Start read-only. Give the model the minimum access required. Analysis should not require permission to edit campaigns, customer records, or storefront settings. |
| 2 |
Define the numbers. Write down what revenue, new customer, conversion, profit, and return mean. Align time zones and currencies. |
| 3 |
Make it show its work. Require source fields, date ranges, baselines, sample sizes, and exceptions for every useful finding. |
| 4 |
Verify one finding manually. Spot-check the records before changing a budget or operating plan. Models can join data incorrectly and turn correlation into a causal story. |
| 5 |
Run the review every week. A recurring prompt makes deep analysis part of the operating rhythm. |
|
|
The new starting point
|
|
The old version of this job started with exports. Now it can start with a question.
There is still setup work. Permissions, metric definitions, and validation matter. Once those pieces are in place, the analysis that used to lose an afternoon can run while the context is still fresh.
Your best insight may already exist. It may simply be split across five tools that have never been asked the same question at the same time.
|
|
|
|
A few founding spots are still open
The Ecom Heads Slack group puts ecommerce operators and specialists in one room to ask for help, compare notes, share jobs, and talk honestly about what is working. If you want to help shape it and keep your membership free for life, apply for the chance to grab a founding spot.
|
|
TALK SOON,
John Sciacchitano
Ecom Heads: Scale or Die Trying
|
|