Predictive marketing uses customers' past behaviour to estimate what each one will do next: buy again, when, how much, or drift away, so you can time and target your messages. A small business can use it. Built-in predictions switch on at data thresholds, such as Shopify's 100 sales or Klaviyo's 500 ordering customers, and below those a spreadsheet version works well.
Two misunderstandings put owners off or lead them astray. The first is that it needs a data scientist; for most small businesses it needs a clean export and a few formulas, or a platform that does it for you. The second is trusting predictions without testing them. A prediction is a probability, and a simple rule such as "remind people five weeks after their last haircut" sometimes beats a sophisticated model. The useful question isn't "can I predict?" but "does predicting beat what I already do?"
The four predictions worth having in a small business
Predictive tools can estimate dozens of things. For a small business, four matter:
- When a customer is likely to buy again. A barber's regular every four weeks, a pet owner's food every five, a contact lens wearer's supply every three months. Knowing the rhythm tells you when a reminder helps rather than annoys.
- Who is likely to lapse. A customer who usually visits monthly and hasn't come for ten weeks is at risk. Catching them at week six is easier than winning them back at week twenty.
- Who is likely to be worth more. Predicted lifetime value ranks customers by what they're likely to spend over the next year, so you can decide who gets early access, a phone call, or a better offer.
- What they're likely to want next. The customer who bought a puppy crate will need a bigger collar in a few months; the client who booked a first gel manicure may want infills in three weeks.
Each prediction leads to an action: a timed reminder, a check-in, a VIP invitation, a relevant suggestion. If a prediction doesn't change what you'd do, it isn't worth producing.
The fourth kind is the easiest to start with by hand, because you already know the patterns. An illustrative garden centre listed what customers tend to buy a few weeks or months after a first purchase: seed trays in February lead to compost and small pots in March; a new lawnmower in April leads to lawn feed in May; a first fruit tree leads to stakes, ties and a pruning question the following winter. Ten of these pairs, written down by the staff who serve customers every day, gave the centre a set of "next purchase" emails triggered by the first purchase. No model needed. Where a platform offers product recommendations, compare its suggestions with your staff's list; the useful ones usually overlap, and the odd ones show you where the data knows something you didn't. Building those follow-ups into a sequence is covered in upsell and cross-sell email sequences with AI.
How a prediction is made, without the maths
Most customer predictions start from three facts about each person, known as RFM: recency (how long since they last bought), frequency (how often they buy) and monetary value (how much they spend). Better models add the gaps between purchases, what they bought, whether they open emails, and whether their pattern is speeding up or slowing down.
The model then compares each customer with thousands of past examples. Suppose customers who bought every four weeks and then skipped two cycles have, historically, stopped coming 60% of the time. A customer who matches that pattern today gets a high lapse risk. The output is a score or band (high, medium, low) rather than a yes or no.
Three things follow from how it works:
- It needs history. A customer with one purchase has no rhythm yet, which is why most tools only score repeat customers.
- It assumes the future looks like the past. A price rise, a new competitor or a change of opening hours breaks the pattern until the model catches up.
- It's a ranking, not a verdict. "70% likely to lapse" means that among customers like this, about seven in ten stop. Three in ten carry on regardless of what you do.
Do you have enough data? The thresholds
Built-in predictions only appear once your data passes a threshold. These are the published requirements for the most common tools:
| Tool | What it predicts | What it needs |
|---|---|---|
| Shopify customer segments | Predicted spend tier: high, medium or low | A store with over 100 sales; only customers with a purchase are scored |
| Mailchimp | Predicted customer lifetime value and purchase likelihood, as high, moderate or low | Standard plan or higher, a connected online store, at least one campaign sent, and enough consistent data for Mailchimp to be confident |
| Klaviyo | Predicted lifetime value, churn risk, expected date of next order, average time between orders | At least 500 customers who've ordered, 180+ days of order history with orders in the last 30 days, and some customers with 3+ orders |
| Google Analytics 4 | Purchase probability, churn probability, predicted revenue for website visitors | In a seven-day period within the last 28 days, at least 1,000 returning users who triggered the condition and 1,000 who didn't, plus purchase events with value and currency |
Read that table honestly against your own numbers. A nail salon with 600 regular clients and no online shop meets none of these, because its data lives in a booking system. A pet shop with an online store and 2,000 orders a year probably qualifies for Shopify's and Mailchimp's predictions. Few small businesses have the website traffic for GA4's. If you don't qualify, the next section is for you.
The spreadsheet version: RFM scoring for a list of any size
You can get most of the practical value with a spreadsheet and an afternoon. The steps, using an illustrative nail salon's booking export:
- Export one row per client with: client ID, date of last visit, number of visits in the last 12 months, total spent in the last 12 months, and usual service. Leave names and contact details out of the working copy.
- Work out each client's usual gap. Divide 365 by their visits per year. A client with 13 visits has a usual gap of about 28 days.
- Compare with now. Days since last visit divided by usual gap gives an "overdue ratio". Under 1 means on rhythm; 1.5 or more means drifting; 3 or more means probably lapsed.
- Score value. Rank spend into thirds: top, middle, bottom.
- Combine into actions. High value and drifting: a personal message from their usual technician. Middle value and drifting: an automated reminder. Anyone on rhythm: leave alone.
Four rows from the illustrative salon's sheet, filled in:
| Client | Visits (12 mo) | Usual gap | Days since last | Overdue ratio | Value third | Action |
|---|---|---|---|---|---|---|
| C-0142 | 13 | 28 days | 24 | 0.9 | Top | None: on rhythm |
| C-0388 | 12 | 30 days | 52 | 1.7 | Top | Personal message from usual technician |
| C-0571 | 6 | 61 days | 95 | 1.6 | Middle | Automated reminder with booking link |
| C-0903 | 3 | 122 days | 400 | 3.3 | Bottom | Include in quarterly win-back only |
This isn't machine learning, but it answers the same questions: who is due, who is drifting, who matters most. For 600 clients it takes an afternoon to set up and a few minutes a week to refresh.
Asking AI to do the RFM work for you
A chat assistant can build the sheet, write the formulas and suggest thresholds from your actual data. Use an anonymised export, and a business plan that doesn't train on your content.
Attached is an export of our salon bookings: one row per visit, with
client_id, visit_date, service, amount. No names.
1. Build one row per client with: last visit date, visits in the last
12 months, spend in the last 12 months, usual gap in days,
days since last visit, overdue ratio (days since / usual gap).
2. Suggest overdue-ratio thresholds for "due", "drifting", "lapsed"
based on how our clients actually behave, and explain why.
3. Tell me which clients have too little history to score.
4. Give me the spreadsheet formulas so I can refresh this myself.An illustrative extract from the reply:
Of 612 clients, 148 have only one visit and can't be given a usual gap; I've marked them "new". Among regulars, clients whose overdue ratio passed 1.5 came back within the following two months only about a third of the time in your data, compared with most of those under 1.2, so 1.5 looks like a sensible "drifting" threshold. Note: December visits are much higher than other months, so gaps measured over winter may look shorter than normal.
What you'd check: that the "came back" figure was calculated from your data and not assumed (ask it to show the counts), and that the December observation is handled, perhaps by measuring gaps over a full year. The assistant's arithmetic on a clean export is usually right; its explanations of why can be confident guesses, so treat those as ideas to check.
An optician's reorders and recalls, worked through
An illustrative independent optician with about 1,400 contact lens patients and 4,500 spectacle patients shows where prediction helps and where a rule is better.
Contact lenses: prediction helps. Patients buy supplies of daily or monthly lenses, but they don't use them at the same rate. Some wear dailies every day; others only at weekends. The practice's shop system records each order. Working out each patient's actual reorder gap (from their last three orders) rather than assuming the box lasts 90 days showed a wide spread: from about 60 days for daily wearers to 150 for occasional ones. Reminders were then timed for 10 days before each patient's predicted run-out.
The numbers over six months: previously, a single reminder at day 80 went to everyone. About 38% of patients reordered through the practice within three weeks of the reminder; the rest either had plenty left (so ignored it) or had already run out and bought elsewhere. With predicted timing, the figure rose to about 52%. On an average lens order of $70 across roughly 700 reminders over the period, that's around 98 extra orders, or about $6,900 in sales that might otherwise have gone to an online retailer.
Eye test recalls: a rule is better. The practice tried to predict who was "likely to come in" for their eye test and found the model mostly rediscovered the recall interval the optometrist had set for each patient. Clinical recall dates, based on age, prescription and eye health, are the right trigger. Prediction added nothing except a risk of skipping patients whose health needed a check. The lesson generalises: where there's a professional or contractual schedule, use it; predict only where behaviour genuinely varies.
Where predictions go wrong
- Seasonality mistaken for lapsing. A garden centre's model flagged hundreds of spring-only customers as "churned" every October. They weren't lost; they come back in March. Compare customers with the same period last year, or build separate expectations for seasonal buyers.
- Life events the data can't see. A pet shop's replenishment emails kept reminding a customer to reorder food after their dog had died. The data showed an overdue customer; the reality was grief. Make "stop these reminders" a one-click option, train staff to update records when a customer mentions a change, and keep the tone of automated reminders gentle enough that a mistake doesn't wound.
- Self-fulfilling loops. If you only ever email the "high likelihood" group, they buy, and the model concludes it was right. The low group never gets the chance. Keep a random slice in every campaign.
- Tiny samples. A barber with 90 regulars can calculate rhythms but shouldn't trust percentages. With small numbers, a few people's habits swing the averages.
- Predictions you shouldn't use. Some platforms infer attributes such as gender from first names. That's unreliable and easy to get offensively wrong; target on behaviour, not guessed identity.
The lapse-risk version of these problems, with a gym as the example, is covered in AI churn prediction for a small gym.
Using customer data for predictions responsibly
Analysing purchase history to decide what marketing people receive is a form of profiling. Data-protection laws in many places regulate it, and customers are entitled to know it happens. Practical steps for a small business: make sure your privacy notice says you analyse purchase or visit history to personalise offers and reminders; only use data you collected for a related purpose; keep exports anonymised where you can; and give people an easy way to opt out of marketing. If predictions would ever affect prices, credit or access to a service rather than just which emails someone gets, talk to a data-protection adviser before you start.
Testing whether prediction beats your current approach
Every prediction should earn its place against something simpler. A fair test takes one campaign cycle:
- Pick the decision the prediction is meant to improve, for example who gets a lapse reminder this month.
- Split the eligible customers at random into two groups of similar size.
- Group A is handled by the prediction (reminded when the model says they're drifting). Group B is handled by your old rule (reminded at a fixed interval), or not contacted at all if you want to know the baseline.
- Measure what matters: bookings or orders within a set window, not opens.
- Decide in advance what difference would justify the extra effort.
A quick illustration with a dry cleaner's 900 regular customers: 450 get reminders timed by predicted gap, 450 get the old "every six weeks" email. Over two months, 81 of the predicted group return versus 63 of the rule group. That's 18 extra returns; at an average order of $26, about $470. Worth keeping, since the setup is done. If the difference had been three or four customers, the simpler rule would have been the sensible choice.
Predicted segments slot naturally into email once you have them; segmenting your email list with AI shows how to turn scores into lists, and winning back lapsed customers with AI-personalised emails covers what to send the drifting group. If your customer data sits in several systems that don't talk to each other, that's usually the first thing to sort out, and it's the kind of groundwork my AI implementation consultation covers.
Questions about predicting customer behaviour
Is predictive marketing the same as AI?
It's one use of machine learning, a branch of AI. The models look for patterns in past behaviour and output a probability or estimate for each customer. It's different from generative AI, which writes text or images, though the two are often combined: a model predicts who to contact, and a chat tool helps write the message.
How accurate are predictions for a small customer base?
Less accurate than for a big one, because there are fewer examples to learn from. Treat any prediction as a ranking of who is more or less likely, not a promise. The only reliable measure of accuracy for your business is a test: compare what the predicted group actually does against a similar group chosen at random.
Do I need a data scientist to use predictive marketing?
No. The built-in predictions in email and shop platforms need no technical work once your data qualifies, and the spreadsheet method in this tutorial needs only basic formulas. You might want specialist help if you want to combine data from several systems, or build predictions your platforms don't offer.
Can predictive marketing work for a service business with no online shop?
Yes, if you record visits or bookings with dates. Salons, barbers, opticians, dry cleaners and garages all have repeat intervals that can be predicted from booking history. The data usually lives in the booking or till system; export it with customer IDs and dates and the spreadsheet method works the same way.
Further reads
- What Is AI Lead Scoring and Does a Small Business Need It? — The same prediction idea applied to enquiries rather than customers.
- How to Create Customer Personas With AI From Real Data — Turn real customer data into personas you can market to.
- How Bakeries Can Use AI to Predict Demand and Cut Unsold Stock — Predicting demand for stock rather than customer behaviour.
- What Can AI Do for Your Email Marketing? 10 Practical Uses — Where predicted segments fit among other AI email uses.
- How to Build a Customer Loyalty Programme With AI Personalisation — Use predictions to personalise a loyalty programme.
- How to Measure Whether AI Is Improving Your Marketing Results — Measure whether prediction-led campaigns really pay.
- Generative AI vs Traditional AI: Which Does Each Task Need? — How generative and traditional AI differ in what they need, cost and get wrong, four questions that sort any task, and a print shop's six tasks sorted.
- Is AI Marketing Automation Worth It for a Small Business? — When AI marketing automation pays off for a small business and when it doesn't yet, with a payback sum, a scorecard and three worked cases.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Shopify Help Center on predicted spend tier; Mailchimp Help on customer lifetime value and purchase likelihood; Klaviyo Help Center on predictive analytics requirements; Google Analytics Help on predictive metrics and their prerequisites.