Can AI tools like Shopify MCP or ChatGPT tell me if my ads are working?

Not reliably. They pattern-match your transactional data against general knowledge rather than analyzing your specific business over time, so they can't separate growth you caused from growth that was already happening seasonally.

A time-series forecasting method built into Google BigQuery's ML platform. It decomposes your historical data into trend, seasonality, and noise, is trained on your own data, and can incorporate external variables like ad spend to test correlation.

How can cutting ad spend not hurt growth?

If a statistical model shows your growth is driven by trend and seasonality rather than ad spend, reducing spend removes cost without removing the thing actually driving growth — so profitability rises while growth continues.

Do I need a data team to do this?

No — but you do need to own your data in a warehouse and use a real model rather than a chatbot. The barrier is approach and tooling literacy, not headcount.

Who is Deacon Bradley?

Deacon Bradley is an ecommerce operator and investor who buys brands and builds the AI systems that run them. He acquires and operates lean DTC health, beauty, and supplement brands.

Why AI Tools Can't Tell You What's Actually Driving Your Ecommerce Growth

A wave of AI tools — Shopify MCP, Meta MCP, email-platform connectors — promise to answer your hardest business questions: Is my marketing working? What’s driving growth? Should I cut ad spend? They answer confidently every time, and that’s the problem. An LLM reading your store data is pattern-matching against general knowledge, not analyzing your specific business over time — so it can’t separate growth you caused from growth that was already coming. The reliable answer comes from a statistical model trained on your own history. In one of my portfolio brands, that distinction let us cut ad spend dramatically with no loss of growth and higher profit.

Why can’t ChatGPT or an MCP tool tell me what’s driving my growth?

Because it has never seen your business as a system that moves over time. An LLM connected to your Shopify store reads transactional data and pattern-matches it against general knowledge. It doesn’t know your seasonal cycles, it doesn’t know the trend underneath your surface numbers, and it can’t tell the difference between growth your marketing caused and growth that was already on its way.

It still answers. Confidently. Every time. That confidence is exactly what makes it risky: you’re making real spending decisions on an answer that was never an analysis.

What’s the difference between an LLM’s answer and real analysis?

An LLM gives you a plausible-sounding narrative. Real analysis gives you a decomposition of what’s actually happening.

Any business’s numbers are three things layered on top of each other: the underlying trend (where the business is really heading), the seasonality (the patterns that repeat year over year), and the noise (everything random in between). An LLM can’t pull those apart for your specific business. A time-series model can — and once they’re separated, the real story usually looks nothing like the surface numbers.

How do you actually measure what’s driving growth?

I model it with ARIMA_PLUS, a time-series method built into Google BigQuery’s ML platform. ARIMA has been the standard for time-series forecasting for decades; the BigQuery implementation makes it practical to run on your own store data.

Three things make it work:

It’s trained on your data. It learns your business — your trend, your seasons — not a generic pattern.
It decomposes the signal. It separates trend, seasonality, and noise so you can see what’s real.
It accepts external variables. You can feed in something like ad spend alongside your history, and the model shows how much of your movement actually correlates with what you’re spending — and how much doesn’t.

That last point is the one that changes decisions. It turns “we spent more and grew, so the ads must be working” into a measured answer about whether the spend and the growth are actually related.

A real example: cutting ad spend with no downside

One of our portfolio brands runs a membership model. As with any subscription business, the question that matters is growth: are we adding members faster than we’re losing them, and is our marketing spend actually doing it?

I modeled it with ARIMA_PLUS and brought ad spend in as an external variable. Two things became clear fast:

The ads weren’t driving membership growth. We’d sometimes raise spend and see growth, but the model showed that was noise, not signal. So we cut ad spend dramatically — and the business was essentially unaffected. Profitability went up. Growth continued.
The hardest season was coming, and no amount of marketing would change that. The model showed it plainly in the history. That reframed the goal: modest improvement through a tough season is a win worth building on, not a failure to spend our way out of.

Neither insight came from asking an AI a question. They came from a statistical model trained on our own data.

What else can this approach measure?

Anything you need to understand over time, once you own your data: sales velocity, inventory demand, customer growth and churn, and seasonal patterns across any part of the business. I’ve been rolling these models out across the portfolio. The membership example is just the clearest one to tell.

What do you need to do this yourself?

Three things, in order: own your data (a warehouse like BigQuery where your history actually lives), the method (a real time-series model, not a chatbot prompt), and the judgment to read what it tells you. The tooling is more accessible than most operators realize. What’s rare is treating the question as an analysis problem instead of a prompt.

The bottom line

AI is genuinely useful in ecommerce — but most of its value is operational, not conversational. The operators who win with it won’t be the ones asking an LLM whether their ads are working. They’ll be the ones who own their data, model it properly, and make decisions on what’s actually true.