AI Remarketing: 6 Steps to Win Back Past Customers on Autopilot

Across more than 14,000 projects, MONA often sees business owners who have phone numbers, emails and order history, yet still look after past customers from each salesperson's memory. AI remarketing gathers these signals to recognize who just visited, who has bought, and who hasn't returned in a long time, then sends the right message at the right moment by email, Zalo or Messenger.

You can set up a basic flow yourself in 6 steps with a data sheet and tools you already have, with free software. This article goes straight into how to do it, the places it easily goes wrong, and when a shop needs to turn order data into an automated customer care system. Start with one flow.

AI remarketing starts with the group of customers your business has already paid to acquire

A past customer already knows the product, has been through checkout and left contact details, so inviting them back usually takes less effort than explaining everything to a stranger. We always review this group before recommending a bigger budget for pulling in new customers. The original ad spend already brought the customer to the website or store. If the profile sits idle after the order is complete, the shop wastes the purchase history, the product usage cycle and the contact channel the customer chose.

AI remarketing uses that same data to create the next touchpoint. A customer who just abandoned a cart needs a short reminder; a customer who bought 30 days ago needs a how-to guide or a reorder suggestion; a loyal customer needs content that is very different from the first two groups. Start with 3 numbers: the 30-day repurchase rate, the average number of days between 2 orders, and revenue from past customers. After 4 weeks of running, those 3 numbers tell you whether the system is truly retaining customers or just creating more messages.

If you have never measured your repurchase rate, don't rush to build 10 scenarios. Pick one group with clear behavior, run it for 1 month and record how many customers come back before expanding. Measure first, then expand.

Manual remarketing costs staff time and is always slower than customer behavior

In a small shop, the seller still remembers which woman often buys black, which man usually orders at the end of the month and who asked about prices last night. As orders grow, that memory ends up scattered across notebooks, Excel files, the Fanpage inbox and personal phones, so no one has the full profile to message accurately. Manual remarketing is also slow at exporting and matching data. The person in charge downloads yesterday's orders, filters out those who haven't bought, writes the content, and enters it on each channel; by the time it's sent, some customers have already ordered or changed their needs.

At a fashion shop selling through its website, Fanpage and store, we once saw a customer who had already paid still get 2 abandoned-cart reminders within 6 hours. We connected order status to a single profile, set a stop rule at the first purchase, then re-ran the test with 12 internal orders. All 12 orders left the cart-reminder group as soon as the payment status changed. Fix the wrong sends first.

Three manual tasks that easily go wrong

First is recognizing the same person. The email on the web, the phone number in the order and the messaging account may belong to one customer, yet 3 systems treat them as 3 different profiles. Merge the right person first.

Second is stopping a scenario. The person in charge usually remembers to turn a campaign on but forgets to set conditions to stop it the moment a customer buys, requests a cancel of delivery, or moves to another care group. We always test at least 3 states: purchase completed, delivery canceled and moved to support. Set the stop first.

Third is keeping frequency consistent across channels. Each team looks at its own calendar, so the same customer gets an email in the morning, Zalo at noon and Messenger in the evening with 3 mismatched offers. Use one shared send calendar.

You can do this yourself while the list is small, but you must assign 1 person to hold the rules and the shared send calendar. Without a final owner, adding more tools only increases the number of places to check. Pick one person to hold the calendar.

AI picks up customers who visited or bought, then chooses the next message

AI remarketing needs one input signal, one set of rules and one output action. The signal can be viewing a product 2 times, abandoning a cart for over 2 hours, completing an order or 45 days without returning; the action is sending content through a channel the customer agreed to receive. AI reads the history to suggest groups, writes several versions of content and schedules the send time based on past reactions. The automation system handles updating the profile, triggering the scenario and stopping the moment the order status changes.

Email, Zalo and Messenger do three different jobs

Email suits content that needs explanation, post-purchase guides, collections or nurture sequences that run for several days. Zalo suits short notifications tied to a phone number, while Messenger is convenient when the first conversation already took place on the Fanpage. We don't force every customer through all 3 channels. Someone who opens email regularly stays on email; someone who only replies on Messenger should stay on Messenger, unless they choose another channel themselves.

If you want to understand the identification and re-engagement layer before building scenarios, the article what is remarketing will help you tell remarketing apart from a mass promotional blast. Reading it first helps the sales team use the same terms. Pick just one channel.

How to set up basic AI remarketing yourself

A first flow only needs to serve 1 group, 1 goal and 1 channel; keeping it narrow lets you see data errors before they spread to thousands of profiles. We often choose the group that has already bought, because order status and purchase date are usually clearer than page-view-only data. Follow the 6 steps below, and test each step with internal profiles before touching real customers. Don't switch everything on at once.

Step 1: Collect enough columns to identify the customer

Create a table with customer ID, full name, email, phone number, the channels they agreed to receive messages on, products purchased, most recent purchase date and order status. If the same email appears twice, keep 1 main profile and merge the purchase history into it. We also add a “last messaged date” column so one person is not touched constantly. That one column alone helps the operator spot conflicts before pushing data to each channel.

Step 2: Choose one group and one destination

An easy example is customers who received their order 30 days ago and have no new order, with the goal of bringing them to the category page they bought from. Write the condition as a readable sentence, and don't use vague labels like “customers about to leave”.

Pick 10 random profiles in the group and check each row by hand. If someone is filing a complaint, has canceled delivery or just reordered, fix the rule before writing the content. Fix even a single wrong row.

Step 3: Have AI write three messages from real data

Give the AI the group name, the product, the time window, the brand voice, the information it may use and the goal of the button. The sample prompt below is enough for a shop to create its first 3 drafts:

Write three messages for customers who bought [product] 30 days ago and have no new order. Friendly tone, don't pretend to know the customer's feelings. Each version under 70 words, mention the exact product, lead to [link], and don't invent any price or offer.

The reviewer must check the product name, price, deadline and link. If the AI adds a gift or discount code on its own, delete it right away and add the line “use only the facts provided” to the prompt for next time. If the facts are wrong, drop it.

Step 4: Set the timing, frequency and stop conditions

Choose a timing based on the real purchase cycle, such as 30 days for a monthly-use product or 90 days for a quarterly maintenance service. Don't copy another industry's timing, because the buying rhythm of coffee, cosmetics and courses is clearly different. Stop conditions must include at least 3 cases: the customer places an order, the customer asks to stop receiving messages, and the person in charge moves the profile to separate handling. Without a stop step, the old scenario keeps running even after the customer's situation has changed.

Step 5: Send a test to a small group

Pick 5% to 10% of the list, watch for 24 hours and only then open up the rest. Check successful sends, replies, clicks, orders and the number of people who asked to stop. We recommend reading the customers' replies too, not just looking at the number table. Lines like “I just bought it” or “why are you sending every day” point straight at the data and frequency errors that charts cannot explain.

Step 6: Compare revenue against a group that wasn't sent anything

Keep a small group that receives no messages over the same period to compare return rates. If the two groups are nearly equal, the campaign is creating more activity but not more business results.

After each round, change only 1 variable such as timing, content or channel. This is slower than changing everything in one night, but you know exactly what brings customers back. Change one variable only.

A coffee chain case shows order data must come before content

The old version used a 6-branch coffee chain as an illustration, and we kept it because this model explains the flow quite easily; this is a sample scenario, not data from a specific client. The shop has purchase history in its loyalty app and point-of-sale machines, but the data is only used to add points. In that model, customers are split into 3 groups: weekly visitors, irregular visitors and those who haven't returned in 60 days. Each group has a different frequency, usual order and purchase hours, so a discount code sent to everyone wastes the valuable part of the data.

How to run the coffee scenario

Weekly visitors get updates on new drinks or points close to the next tier, with no need for constant discounts. Irregular visitors get suggestions based on drinks they have bought, while the group silent for 60 days gets an invitation back during the time window when they usually order. When a customer buys at any branch, the status must update the shared profile and stop any pending re-engagement messages. If the point-of-sale machine in the store doesn't send data back, the AI still thinks the customer hasn't bought and keeps messaging the wrong thing.

The lesson we take from this case is quite blunt: no matter how well the AI writes, it cannot rescue a purchase status that updates slowly. Connect the data from the point of sale first, then debate which wording sounds more appealing. Connect the POS first.

Repetitive content and multichannel are where DIY usually breaks

At a shop selling regular food deliveries, we once collected 4,600 profiles from the website and the sales software and found that 318 people who had already bought were still in the “30 days not returned” group. We fixed the entry condition, set status syncing every 15 minutes and ran a reconciliation for 2 weeks. The new flow stopped correctly after a new order, and only people who had not bought received reminders. Fix the status first.

As the list grows, the same template gets swapped with a new name and sent month after month, and customers notice very quickly. Real personalization changes with the product, the purchase, the reply and the current status, while swapping only the name at the top of the message is just a data merge. Multichannel makes errors worse because each place has its own calendar, format and way of receiving replies. A customer who clicks an email, asks a question on Messenger and then buys in store needs to be recorded as one thread, otherwise 3 teams still think they are 3 different people.

Signs the system is breaking include:

• One customer gets the same message from two channels within 24 hours.
• People who have bought are still in the abandoned-cart or long-time-no-return group.
• The AI repeats the same opening across many campaigns.
• Staff have to export, edit and re-import a file before every send.
• Reports only show views and can't be tied to orders.

At this point, you need a centralized customer profile and channel priority rules. The article re-engaging past customers with AI goes deeper into how post-sale data becomes a proactive care calendar. Our view is quite clear: a small list is fine to do yourself so you understand your customers, but for a large list, stop merging files before you hire more senders. Adding 2 more senders still doesn't solve order status updating late between systems. Don't fix it by adding people.

MONA connects AI customer care with order data

MONA has been working since 2016, has delivered over 14,000 projects and has an 85% returning-client rate, so we understand that the value of the second purchase lies in how the system remembers what actually happened. With AI remarketing, the first thing we review is the path of orders, customer profiles and replies between the website, point of sale, email, Zalo and Messenger. When MONA's AI customer care system is attached to the order data, a customer who has bought leaves the cart-reminder group, moves to the post-sale flow and is prioritized on the channel they responded on. Content is written by AI within an approved framework, while every price, condition and deadline comes from data the shop manages.

The website records the products viewed, the cart and online orders, so a sales website that connects customer care data lets remarketing run from real behavior instead of a weekend export file. We also reconnect in-store data when a shop sells both in person and online. You don't need a dedicated system if you only have 1 channel and a few hundred profiles; run the basic flow above to see exactly where the effort goes. When the list is large, content starts to repeat and customer status drifts between channels, call 1900 636 648 so MONA can review the data and build AI remarketing that follows the orders closely. Review the orders first.