Klaviyo AI SMS Personalisation: What to Test Before You Scale
By Trayan Hristov · Thrivelia · Published 13 November 2026

Two people can trigger the same SMS flow for different reasons. One is comparing a product seriously. Another took a quick look. Better wording may help, but I would first check whether either person should receive that message now.
Klaviyo SMS personalisation deserves a test with a clear commercial question. Can adapting the wording improve the result for the same eligible customers, after the extra cost? That is a narrower and more useful question than whether AI can write convincing texts.
What Klaviyo has released
Klaviyo announced Personalized Content on 4 November 2026. It adapts flow text from an existing message and compares personalised output with the original. Its planned grader model was described as forthcoming, not available at launch.
The current Help Center guide lists the Personalization package and six supported flows: Added to Cart, Viewed Product, Started Checkout, Back in Stock, Price Drop and Win Back. It covers SMS, MMS text and plain text RCS. Offers, links and dynamic tags remain fixed; targeting and timing do not change. An active message A/B test prevents enabling it.
Check that guide against the account before implementation. The release article and operational documentation differ on some scope details. I would avoid making a purchasing decision from the launch summary alone.
Write down what must remain true
My proposed starting point is a short decision brief: customer need, current eligibility, product promise, useful destination and the result being tested. Keep the existing message in that brief. It is both the comparison and a message customers can still receive.
For example, consider a shopper returning to a pair of boots. A useful prompt might direct them to sizing information they have not found. A more urgent sentence cannot fix an unavailable size, an inaccurate delivery promise or a landing page that ignores their question.
Review area | Question for the team |
|---|---|
Customer eligibility | Has a purchase, open support issue or other change made this prompt inappropriate? |
Product truth | Could shortening a name remove a material fit, model or compatibility distinction? |
Destination | Does the link help complete the decision promised by the message? |
Offer | Is the incentive valid for this customer, item and market? |
Measurement | Which order and cost definitions decide whether the treatment is useful? |
These are Thrivelia's proposed review questions. They are not a claim about the signals the feature uses. Keep your own eligibility rules responsible for the decision to send.
Review wording against a product decision
Here are two illustrative message directions for a fictional footwear brand. They are brief concepts, not production-ready SMS and not examples generated by Klaviyo.
Light browsing: 'Still comparing boots? Our fit guide explains the widths and sizing before you choose.'
Repeated consideration: 'A question about fit? Check the size guide or ask our team before choosing your pair.'
Neither depends on telling the customer how often they have been watched. The useful distinction is the buying help offered. Before use, the brand would still need to add its identity, approved link and required messaging elements, then check the final rendered text.
Review samples for product claims as carefully as tone. If a technical item is shortened to a generic category, could the customer confuse it with another model? If a product name includes an essential size or compatibility restriction, does the destination keep that clear?
Keep one reviewer responsible for product accuracy and another for the customer journey where the team has that capacity. Record the baseline, review date and approved guardrails so future changes have a reference point.
Choose a test that can answer the question
The Help Center describes an automatic comparison with a baseline group, assignment that stays consistent, and performance results that become confirmed as sufficient evidence accumulates. Personalised messages may fall back to the original. The result describes the treatment group, not proof that one particular sentence won.
I would begin with messages whose events, audience and destinations already work, and enough eligible volume to make the comparison useful. Fix known routing problems first. Otherwise a test can spend weeks measuring copy inside a broken journey.
Record the question before activation. For example: 'Does personalising this checkout message improve retained contribution per assigned eligible profile compared with our current approved message?' Define the observation window, order exclusions and costs at the same time.
Keep the baseline and surrounding programme stable where possible. If the team changes the offer, acquisition mix or promotion during the observation period, record it. A holiday test may still answer a holiday question, but it should not be presented as an evergreen result without further evidence.
Do not compare clickers with non-clickers
People who clicked already behaved differently. Comparing their purchases with everyone else's does not isolate the value of the treatment. Read the assigned groups using the test's actual reporting definitions. If you cannot obtain a metric at the required group level, label that limitation.
A copy comparison answers which messaging treatment performed better. To estimate the value of sending an SMS at all, you need a separately designed no-message comparison. Our attribution and incrementality guide explains the difference.
Calculate the commercial result beyond clicks
Use retained order contribution after the agreed variable order costs, then subtract messaging costs on a consistent basis. Include any additional software fee and review work attributable to the treatment. Existing fees shared by both groups should not be counted twice.
The following is fictional arithmetic, not a Klaviyo benchmark or a client result. Assume equal groups of 10,000 assigned customers, each qualifying purchaser makes one retained order, and contribution is $35 per order before messaging.
Measure | Baseline | Personalised |
|---|---|---|
Assigned customers | 10,000 | 10,000 |
Qualifying purchasers | 250 | 270 |
Order contribution before messaging | $8,750 | $9,450 |
Messaging cost | $300 | $320 |
Contribution after messaging | $8,450 | $9,130 |
The observed difference is $680 for equal group sizes. If incremental software and review costs allocated to this comparison are $750, the observed net is negative $70. More purchasers have not yet paid for the additional treatment. The example makes no claim about statistical reliability.
With unequal groups, compare contribution per assigned profile before scaling the difference to an explicitly stated population. Do not subtract raw totals from unequal groups. Review uncertainty and refunds before treating a small observed gap as a decision.
Klaviyo says keeping the same segment count is a best effort, not a guarantee. Check actual billed usage. Our SMS ROI guide covers the wider cost model, while the personalisation ROI guide includes production and maintenance.
Decide what the next result would change
Keep the treatment when the evidence, economics and customer experience support it. Investigate if clicks improve while retained purchases do not. Rework the underlying journey if the destination or recommendation fails. Leave an inconclusive result labelled inconclusive.
At Thrivelia, I would use this feature inside a programme with clear product direction, eligibility, creative review and reporting ownership. Those responsibilities remain useful as drafting becomes faster. See how we assess agency value alongside AI, or review our services for help designing the test and implementing the surrounding journey.
Does AI SMS personalisation replace flow strategy?
No. Wording is one decision within a flow. Your team still needs to decide who qualifies, which product or offer is appropriate, what happens after a purchase or complaint, and how to assess the result.
When should we judge the test?
After enough outcome data has accumulated for the question being asked. Do not use an early click difference as proof of a purchase or profit improvement. A fixed number of days cannot guarantee a useful answer for every flow.

.png)
Comments