This year “artificial intelligence” has become the favorite phrase of every sales pitch. Every email tool, every ad platform, every online store plugin is now “powered by AI”. You see it in brochures, on conference banners and in the offers landing in your inbox. Sometimes it’s true. Other times it’s an IF-ELSE statement with a big marketing budget.
I’ve spent the past few years working on Retargeting, a personalization and marketing automation platform for online stores, so I see both sides: what machine learning can genuinely do, and how often it’s used as a label. I wanted to write down, as concretely as possible, what actually works today, in the autumn of 2017, and what is still science fiction.
From AlphaGo to your online store
Last year AlphaGo, the program built by DeepMind (owned by Google), beat Lee Sedol at Go, a game many people said would resist computers for another decade. The headlines were apocalyptic. The reality is more mundane and, for a merchant, more useful: the same families of techniques, neural networks and statistical models that learn from data, have meanwhile made their way into tools a store with ten employees can use.
The key distinction is this: what we call AI in marketing today is almost always machine learning. A system is fed lots of past examples and learns to predict one specific thing: who will buy, what they will buy, when they will open an email, how much a click is worth. It doesn’t “think”. It makes narrow predictions, but it makes them very well and very fast.
What already works, concretely
Here is my list of applications that are no longer experiments and actually make money for those who use them properly:
- Product recommendations. Amazon has done this for years, and so has Netflix. Based on what similar customers viewed and bought, the system shows you the products you are most likely to care about. I wrote at length about personalization that goes beyond a first name in the subject line.
- Send-time optimization. Instead of sending the newsletter to everyone at 10 a.m., the model learns when each subscriber tends to open and delivers the message then. It sounds like a detail, but it matters, especially since email marketing still drives the most sales for many stores.
- Predictive segments. “Likely to buy in the next 7 days”, “likely to churn”. Instead of hand-built segments by age and city, you get lists built on behavior and probability.
- Bid optimization in Google AdWords and Facebook Ads. Automated bidding strategies (target CPA, target ROAS) and Facebook’s conversion optimization adjust the bid at every auction, using signals no human could weigh manually. The old debate of Facebook Ads or Google AdWords for an online store has partly shifted from “how do I bid” to “what data do I give the algorithm”.
- Image recognition. Services like Google Cloud Vision or Amazon Rekognition can tag products automatically from a photo: color, type, style. For a catalog with thousands of items, that means hours of work saved and better filters.
- Fraud detection. Payment processors have used models for years that flag suspicious transactions within milliseconds.
- Basic chatbots. For repetitive questions (where is my order, how much is shipping) they work reasonably well. I looked at them recently, wondering whether chatbots are a passing fad.
What it can’t do yet
This is the part sales decks skip. An algorithm can’t write campaign copy that makes you smile. The automatic text generators I’ve seen produce sentences that are grammatically correct and completely soulless. Creativity, humor, the right tone of voice for a brand remain, for now, a human job.
Second: context. A system sees you bought a stroller and recommends three more. A human would understand you need a car seat, not a second stroller. Or that it was a gift and you don’t need anything else. Models learn correlations, not reasons.
And most importantly, AI doesn’t replace strategy. It doesn’t decide which market you enter, at what price, with what positioning, which customers you want and which you leave to competitors. It optimizes the road you chose very well. It doesn’t choose the road for you.
Step one: clean data
Every application above has one thing in common: data. Lots of it, accurate and connected. If a customer appears three times in your database with three email addresses, if tracking on your site is half set up, if store sales don’t connect to your campaigns, the smartest algorithm in the world will learn nonsense.
Say you run an online sporting goods store. Before shopping for “AI solutions”, check three things: that every order is tied to a single customer, that you know where every visitor came from, and that key on-site events (product view, add to cart, purchase) are recorded correctly. It sounds boring. It is. But it’s the foundation.
Step two: tools with machine learning built in
A small or mid-sized store doesn’t need a team of data scientists. It needs to pick tools that already have the models built: an email platform with send-time optimization, a recommendation engine, automated bidding in AdWords, conversion optimization in Facebook. Someone has already invested millions in the algorithms. You just feed them your data.
A practical tip: let the algorithms learn. Automated bidding strategies need a minimum number of conversions and a few weeks to stabilize. If you change the settings every day because “it isn’t working”, you never give them a chance to work.
Step three: measure the lift, not the promise
Every vendor will show you a beautiful chart. The only chart that matters is yours. How do you get it? With A/B tests: part of your traffic or subscribers gets the “smart” experience, the other part gets the old version. You run the test long enough to get a statistically significant result, and you compare sales, not clicks.
If personalized recommendations raise your average order value, great, keep them. If you see no difference, either your data isn’t good or the tool isn’t right for you. There’s no shame in dropping it. The shame is paying for years for something you never measured.
What comes next
I think machine learning will become invisible over the next few years. We’ll stop saying “AI-powered store”, just as we no longer say “internet-powered store”. It will simply be how tools work. The advantage won’t go to those who buy the most “AI”, but to those with the best data and the clearest strategy.
Until then, my advice is simple: ignore the noise, pick two or three concrete applications from the list above, clean up your data and measure. The rest is marketing about marketing.
What are you already using in your stores? Has a “smart” tool ever brought you growth you could actually measure, or just one more invoice?