Tecof • September 15, 2026
AI-Driven E-commerce Email Campaigns: A Personalization Guide
In Brief
E-commerce email marketing is the practice of generating sales by sending behaviour-driven automated and manual messages to a list a brand has collected with explicit permission; artificial intelligence's role here is not writing copy, as most people assume, but scoring who belongs in which segment, when they should hear from you and how often. As of 2026 the question is no longer whether email still works, but how accurately you split your list and how clean you keep your consent records.
Wednesday evening, 20.10. The marketing lead at a home textiles brand sent a single discount campaign to all 48,000 people on the list. The results were clear by morning: 41,900 delivered, 22 percent opens, 1.4 percent clicks, 190 orders, 214,000 TL in revenue. That same send produced 96 unsubscribes and 31 spam complaints. Two weeks later the same list was split into four: new-season pieces to people who had bought in the last 90 days, a reminder to those who had bought within 12 months and then gone quiet, a first-order voucher to registrants who had never bought, and nothing at all to anyone who had already received two messages in the last 30 days. Total volume fell from 48,000 to 31,400, revenue rose from 214,000 TL to 268,000 TL, and complaints dropped from 31 to 7.
The problem was never the copy. The two campaigns had nearly identical subject lines and the same visuals. The difference was that in the second one, 16,600 people never saw the message at all. In email, most of the gain comes not from writing better but from writing to fewer and more relevant people; and that is precisely what artificial intelligence makes possible at scale.
Where Email Stands in 2026
An owned list versus rented reach
On advertising platforms you rent reach. The day you cut the budget the traffic stops, costs swing with competition, and targeting rules change without notice. An email list is yours: it can be exported, carried across platforms, and costs fractions of a lira per send. The range we see in the field is that email and SMS together carry between 12 and 25 percent of total revenue at an established Turkish e-commerce brand. Most of that share comes not from calendar-driven campaigns but from automated flows running quietly in the background. We compared the cost, consent and readership differences between email and SMS in a separate piece; in practice they are not rivals but two ends of the same permission pool.
What AI actually changes
Writing subject lines has been cheap since 2023. Today a language model will give you forty variants in thirty seconds and most of them will be acceptable. The real difference lies elsewhere. Predicting when a customer will place their next order, which category they are drifting toward, and which week they will go quiet and never return is a statistical job that cannot be done by hand. Producing three numbers for every person on a 40,000-name list (probability of purchase, expected order value, churn risk) is an hour's work for a model and impossible manually. Those three numbers are what raise your programme's return, not subject line variants.
Manual campaigns versus automated flows
Most brands spend 80 percent of their effort on manual campaigns and 20 percent on the flows that produce 70 percent of the revenue. The ranges below are typical of mid-sized stores in the Turkish market; they shift with category and basket size.
| Measure | Manual campaign | Automated flow |
|---|---|---|
| Trigger | Calendar, promotional window | User behaviour |
| Share of monthly volume | 70-85% of total | 15-30% of total |
| Open range | 18-28% | 38-58% |
| Click range | 0.8-2.2% | 3-11% |
| Orders per 1,000 sends | 2-6 | 12-45 |
| Effort after setup | Repeated every send | Quarterly review |
The conclusion is simple: flows are built once and run for months. Manual campaigns do not replace flows, they sit on top of them.
Segmentation: The Logic of Splitting a List
RFM and the purchase cycle
RFM is shorthand for three simple questions: when did they last buy, how many times have they bought, how much have they spent in total. Five bands on each dimension produce a 125-cell grid; in practice reducing that to eight or ten meaningful groups is enough. Where RFM falls short on its own is the purchase cycle. If you sell coffee the average repeat interval might be 28 days; if you sell furniture it might be 400. The same label of "hasn't bought in 90 days" is an alarm in the first case and perfectly normal in the second. Calculate the median repeat interval per category and tie your definition of a lapsing customer to that number.
Category affinity and churn risk
Category affinity is the distribution of categories across what a person viewed and bought in the last 180 days. Weighting purchases twice as heavily as views and then normalising works well enough in most stores. Churn risk is harder: you combine how far a person has overshot their expected repeat interval, their engagement across the last three sends, and the time elapsed since their first order into a score between 0 and 1. Trigger the reactivation flow for people approaching your threshold, and remove those who have passed it from the sending pool entirely. The second half sounds counterintuitive but is one of the most valuable decisions for list health.
Model scores versus hand-built segments
A hand-built segment is a rule: "bought in the last 60 days and spent over 1,500 TL". A model score is a probability: "0.42 likelihood of purchase in the next 30 days". The rule is transparent and explainable, and when it is wrong you can see why. The score is more accurate but opaque, and it degrades silently when the underlying data drifts. What works in the field is combining the two: draw the coarse boundary with a rule, then rank within it using the score. Define the pool with "at least one order in the last 12 months", then send first to the riskiest 15 percent by churn score.
The limits of segmentation on a small list
Below 5,000 people, segmentation loses meaning fast. Split the list six ways and each segment holds 800 people; at a 2 percent click rate that is 16 clicks, and 16 clicks teach you nothing. On small lists the right move is to settle for three segments (new registrants, active buyers, going quiet) and put the energy into growing the list instead. Calculate the sample size an A/B test needs to produce a meaningful result; if you cannot reach it, believing you are testing is more dangerous than not testing at all.
| Segment | Message angle | Frequency | Expected clicks |
|---|---|---|---|
| New registrant, never bought | Brand introduction, first-order incentive | Weekly | 4-8% |
| Bought in last 90 days | Complementary products, new season | 1-2 per week | 3-6% |
| High spend, regular | Early access, stock priority | Twice weekly | 5-9% |
| Overdue on their cycle | Reminder, limited incentive | Every two weeks | 2-4% |
| No engagement in 180 days | One-off win-back | Monthly, capped at 3 attempts | 0.5-1.5% |
Automated Flows and Realistic Expectations
Welcome series and cart abandonment
The welcome series is the highest-engagement message on your list because the person has just handed you their address. Three messages are enough: a greeting with any incentive immediately after signup, brand story and bestsellers 48 hours later, and a category preference prompt on day five. The cart abandonment flow is the highest revenue producer. A three-step structure works well: a neutral reminder after 30 minutes, product detail and stock status after 24 hours, a final nudge after 72 hours. Do not put the discount in the first message; if you do, customers learn to abandon carts deliberately and you erode your own margin.
Browse abandonment, post-purchase and back-in-stock
Browse abandonment targets people who lingered on a product page without adding to cart. It converts at roughly a third of the cart flow's rate, but its volume is far higher, so its total contribution is significant. The post-purchase flow does two jobs: it manages delivery expectations and it prepares the repeat order. Send a complementary product suggestion 7 to 10 days after delivery, and a reorder reminder at 80 percent of the category's median repeat interval. The back-in-stock notification is the highest click-rate message on any list, because the person has already declared they want that item. Approaches to raising average order value feed directly into how you build the content of these flows.
Reactivation and the exit door
The reactivation flow must not be endless. Send three messages: a reminder, your best offer, and a direct question ("do you still want to hear from us?"). If none of the three gets a response, remove the person from the sending pool. Skipping this step is the most common way deliverability erodes over time. Every send to a dormant address lowers your domain reputation and reduces the chance of reaching the inbox of the customers who do buy.
| Flow | Trigger | Timing | Conversion per send | Note |
|---|---|---|---|---|
| Welcome | List signup | 0 hr / 48 hr / day 5 | 1.5-4% | Incentive in one message only |
| Cart abandonment | Add to cart, no purchase | 30 min / 24 hr / 72 hr | 3-9% | Discount in the final step |
| Browse abandonment | Product page, no cart | 4 hr / 36 hr | 0.8-2.5% | High volume |
| Post-purchase | Delivery confirmation | 7-10 days | 1-3% | Complementary products |
| Reactivation | 120-180 days silent | 0 / 7 / 14 days | 0.3-1% | Capped at three attempts |
| Back in stock | Stock alert signup | 1 hr after restock | 8-20% | Highest click rate |
Generating Subject Lines and Copy with AI
Producing variants and holding the tone
Do not ask the model for a single subject line; ask for eight or ten variants and choose among them. For those variants to be usable, though, you have to give the model the brand's voice file: how many words, what form of address, which words are forbidden, which concrete examples are allowed. Including five previously approved subject lines as examples is more effective than pages of stylistic description. You can automate the tone check too: use a second prompt to score the generated text against your brand rules and filter out anything below threshold. We covered the structure of prompt engineering for e-commerce in detail; the same framework applies to email.
Spam triggers and forbidden phrasing
Language models lean toward inflated marketing patterns because of what they were trained on. For a serious brand most of those patterns are both off-tone and a filter risk. Write the following out as explicit prohibitions inside the prompt:
- Urgency inflation: stacking "last chance", "don't miss out" and "today only"; if there is no real deadline, none of them should appear at all.
- Capitals and punctuation: all-caps words, consecutive exclamation marks, currency symbols piled into the subject line.
- Unverifiable claims: superlatives you cannot document, such as "the best in Turkey" or "scientifically proven".
- Fake personalisation: an empty name field when the data is missing, or filler like "dear valued customer".
- Misleading subject lines: a headline that does not match the offer in the body; this is the fastest way to raise your complaint rate.
Output review and acceptance criteria
Every generated piece should pass a checklist before it goes live. Write the acceptance criteria down: the subject line stays under 42 characters, the preview text does not repeat the subject line, the body has a single primary call to action, price and stock come from the product database, nothing from the forbidden list appears, and the footer carries the unsubscribe link and the legal trading name. Running this list manually is possible but does not scale; at Tecof the content generation step runs within defined authority limits, variants that fail the acceptance criteria are filtered out automatically, and only what passes reaches human approval.
Send Time, Frequency and List Health
Send-time optimisation
General advice like "Tuesday at 10.00 is the best slot" no longer works, because everyone follows the same advice and the inbox is congested at that hour. Per-person timing does better: derive an individual window from the distribution of each subscriber's opens and clicks over the last six months and spread sends across those windows. The lift we see in the field is an extra 5 to 12 percent in opens; not a transformation, but free. Fall back to the segment average for subscribers without enough history, and switch to the individual window as data accumulates.
Frequency, fatigue and the cap rule
Fatigue is when subscribers stop opening but do not unsubscribe. That is the dangerous state, because it is invisible in your headline metrics while it quietly erodes deliverability. The remedy is a frequency cap: no more than three marketing messages per person per week, with automated flows excluded from that count. Define a quiet window as well; if someone has received a message in the last 48 hours, exclude them from the next campaign. In our experience those two rules alone cut the complaint rate by about a third.
Technical setup and thresholds
Deliverability is a matter of technical setup before it is a matter of content. Publish all three of SPF, DKIM and DMARC for your domain; start DMARC at p=none, watch the reports for two or three weeks, then move to p=quarantine. The thresholds to monitor: hard bounces below 0.5 percent, soft bounces below 2 percent, spam complaints below 0.1 percent, unsubscribes below 0.3 percent. Once the complaint rate passes 0.3 percent, the large mailbox providers begin throttling you, and the only correct response at that point is to cut volume and go back to list hygiene.
Commercial Electronic Messages in Turkey: Consent and Privacy
Consent and the consent record
In Turkey, sending commercial electronic messages requires prior consent, and those consents must be registered with the Message Management System (İYS). In practice the sequence is: you collect consent through your own channel (signup form, checkout step, in-store form), you upload it to İYS, and before every send you reconcile your list against the İYS records and remove anyone without consent or who has opted out. The proof of consent matters as much as the consent itself: store the date, time, IP address, the channel through which it was given, and the version of the consent text as it stood on that date. A pre-ticked box, consent bundled into a condition of service, or a checkout form with the box checked by default does not count as valid consent.
The right to refuse and how fast you must act
Every commercial electronic message must carry a means of refusal, and that refusal must be free and easy to reach. Do not bury the unsubscribe link in the footer; make it work in one click, with no second login screen. Once a refusal arrives, stopping sends should not be delayed; in practice aim for the same day and no later than 24 hours. The refusal must be reported to İYS as well. Avoid "did you unsubscribe by mistake?" recovery screens; they create compliance risk and raise the complaint rate.
Privacy notice and profiling disclosure
İYS consent does not substitute for your obligations under Turkey's personal data protection law (KVKK); they are separate regimes. The duty to inform arises the moment you start processing personal data: state plainly which data you process (address, order history, on-site behaviour), for what purpose, how long you retain it, who you transfer it to, and what rights the individual holds. The critical point for this article is profiling: if you are building segments and scores from behavioural data, your privacy notice must say that you profile and that the analysis is carried out by automated systems. An individual has the right to object where a decision made solely by automated analysis produces an adverse result for them; in e-commerce this typically arises around price differentiation or restricted access. If you use a third-party email platform, review your data processor agreement and the conditions for any transfer abroad.
Measurement: From Open Rate to Revenue
Why the open rate no longer stands alone
Apple's mail privacy protection pre-loads the tracking pixel on the user's behalf. The result, on lists where those users make up a meaningful share, is an open rate that reads higher than reality and bears little relation to actual behaviour. Do not discard opens entirely; they still work as a trend line and a deliverability warning. But make decisions on clicks, orders and revenue. Tie your subject line tests to revenue per send rather than opens; otherwise you will crown the variant that gets opened more and sells less.
Holdout groups and incremental revenue
The only reliable way to know whether a flow actually makes money is to send nothing at all to a small slice of its audience. Hold back 5 to 10 percent of the cart abandonment audience as a control group and measure their natural conversion. The gap is the flow's incremental contribution, and it is almost always lower than the raw figure; some of the people who completed the purchase would have come back without your message. Keep the holdout permanent and read it once a quarter.
Channel attribution and double counting
Email, advertising and organic search can each claim the same order in their own dashboard. Add them up and you find twice your actual revenue. Pick a single source of truth, ideally your order database or ERP, and compare channel dashboards only against themselves over time. The same discipline applies to work on raising conversion rates: without fixing where you read the number, you cannot claim an improvement.
| Metric | What it measures | Reliability | Decision it supports |
|---|---|---|---|
| Open rate | Pixel loads | Low | Trend and delivery warning |
| Click rate | Real engagement | High | Content and offer decisions |
| Revenue per send | Monetary yield of a message | High | Frequency and segment decisions |
| Incremental revenue | Net contribution of a flow | Very high | Whether to keep the flow |
| Complaint rate | Level of annoyance | High | Cap and hygiene decisions |
Building the Email Programme in 30 Days
Days 1-7: consent, hygiene and technical groundwork
Reconcile the list against İYS records and remove addresses without consent or with a recorded refusal. Move anyone with no engagement in the last 12 months into a separate bucket and pull them out of the main pool. Publish SPF, DKIM and DMARC records, starting DMARC at p=none. Review your signup forms: the consent box unticked by default, the consent text explicit, the record stored with date and channel. Update the privacy notice so it includes the profiling disclosure. Send no campaigns at all this week.
Days 8-14: two core flows
Build only the welcome series and the cart abandonment flow. Three messages each, six templates in total. Pull product blocks dynamically from the catalogue; never type price or stock by hand. Hold back 10 percent of the cart flow as a control group. Run each of the six templates through the acceptance checklist and read them on a phone; more than 60 percent of lists are opened on mobile. The first numbers arrive by the end of the week, but do not interpret them yet.
Days 15-21: segmentation and scoring
Build the RFM grid and reduce it to eight or ten meaningful groups. Calculate the median repeat interval per category and tie your lapsing-customer threshold to it. Produce the churn score and test it against historical data: do the scores from three months ago actually catch today's quiet customers? Write down the message angle and frequency cap for each segment. Add the browse abandonment and post-purchase flows this week. Enter the frequency cap and the 48-hour quiet window into the system as hard rules.
Days 22-30: measurement and scaling
Build the revenue-per-send report with a segment breakdown. Take the first reading from the holdout group and calculate the cart flow's incremental contribution. Turn on per-person send-time optimisation for subscribers with enough history. Launch the reactivation flow with three messages and a firm exit rule. At month end, look at a single table: which segment, how many sends, how much revenue, how many complaints. Build next month's plan from that table, not from intuition.
Here is the job for tomorrow morning: open your email platform, filter for subscribers who have produced not a single open or click in the last 12 months, write the number down, and move those people out of the main sending pool into a separate list; then in the afternoon download your İYS records and clear out every address without valid consent, because every improvement you make from here rests on those two steps.
Frequently Asked Questions
I have 2,000 people on my list. Should I start segmenting?
Do not split into more than three segments. New registrants, buyers from the last 12 months, and those going quiet is enough. At that size, put most of your energy into growing the list and building the two core flows; adding segments will not produce statistically meaningful results, only more work.
Can I let AI write the email copy entirely?
You can let it write the draft; you cannot send that draft live. Verifiable details such as price, stock, delivery times and campaign terms must come from system data. The copy itself must pass a written acceptance criterion; otherwise the first wrong price or misleading subject line costs more than the tool ever saved you.
Can I use an old list without registered consent?
No. Sending commercial electronic messages to addresses whose consent is not registered with İYS carries a risk of administrative penalty, and it will raise your complaint rate and damage deliverability as well. Rather than sending those addresses a single "renew your permission" message, fix your consent channels and rebuild the list properly from scratch.
How many emails should I send per week?
Do not exceed three per person per week, and keep automated flows outside that count. The right number varies by list; if complaints start passing 0.1 percent or unsubscribes 0.3 percent, your frequency is too high. Find the number by stepping it up every two weeks and watching the metrics, not by changing it all at once.
My open rate is 45 percent. Is that good?
On its own it tells you nothing. Apple's privacy protection can inflate opens artificially. If over the same period your click rate is below 1 percent and revenue per send is falling, a high open rate is not good news but measurement noise. Make the decision on clicks and revenue.
Should I put a discount in the cart abandonment message?
Not in the first two messages. In the third, offer a limited incentive only to people above a certain basket value who are abandoning for the first time. If you attach a discount to every abandonment, your regular customers learn to leave the cart sitting deliberately and you erode your own margin.
Should I build email or SMS first?
Email. It costs less, gives you more room for content, and suits automated flows better. Keep SMS to short, time-sensitive messages such as the final cart abandonment step and back-in-stock alerts. The consent pools are separate; holding one does not grant you the right to use the other.
Doesn't keeping a holdout group cost me revenue?
In the short term, yes, a very small amount. A 5 to 10 percent control group amounts to a fraction of a percent of total revenue. In return you learn the flow's true contribution; without that, the risk of running an ineffective flow for months is many times the cost of the holdout.
Which data do I actually need for personalisation?
Three things are enough: last order date, the category distribution of past orders, and on-site behaviour over the last 90 days. Fields like first name, birthday and city sound appealing but their effect on revenue is not measurable. Every field you collect creates a data protection responsibility, so do not collect what you will not use.
Do I need special infrastructure for AI segmentation?
Not necessarily, but order, product and on-site behaviour data have to come together in one place. If the data sits in three separate systems, solve the integration before you build any model. Agent setups operating within defined authority limits can automate the scoring and segment refresh steps, but no model works until the underlying data is unified.