Purchase prediction: how to read the Flowpredict score

1. Read the score as a signal of interest.

On a profile, start with four pieces of information: the score, the visit history, purchases and contact recognition. Below, the visitor has 71/100, two visits, seven page views and an add-to-cart, but no recorded purchase. That's a concrete reason to check the offer and the last stage of the journey. The score alone doesn't explain whether the obstacle was shipping, price or an interrupted visit.

The "Anonymous" label (shown in Polish in the screenshot) matters as much as a high score. It lets you analyze behavior, but it doesn't give you a recipient for a tag in your marketing tool. A contact action requires recognition, a link to the contact in the integration and the right consent. Don't read the score alone as permission to communicate.

2. Check intent before the cart, too.

The 40/100 profile has five visits and 23 page views, but no cart and no purchase. The model can assess interest before a product is added. With the condition "score of at least 70," however, this profile won't go any further. Return visits don't guarantee a higher score; the next recalculation takes the available signals into account, not just a growing page view count.

This is a separate visitor, not an earlier stage of the 71 profile. More visits with a lower score isn't a contradiction in itself: the model takes many behavioral features into account. The radar helps you read their estimated contribution, but it doesn't prove that artificially boosting a given feature will lead to a purchase.

The threshold of 70 is a chosen action rule, not a promise that seven out of ten people will buy. You set it separately from the segment thresholds in the ML report. Write down what group size and cost of errors you accept; don't change the threshold just to get a particular profile into a campaign.

3. Check what model accuracy really means.

Accuracy is the share of the model's assessments that are correct. It also depends on the ratio of buyers to non-buyers. With 1,000 observations and 20 purchases, the rule "nobody will buy" would be 98% accurate, even though it wouldn't identify a single buyer. This is an arithmetic example, not a Flowpredict result.

Distinguish a person's score from the metrics of the model as a whole. On new profiles, next to a score of 40, 71 or 95, there's an accuracy of 94.4% with a description of the test: one buyer to at most four non-buyers, not real traffic. That's not the effectiveness of a future campaign or the share of people with a score of 71 who will buy. For an evaluation, you also need the number of cases, the period and the types of errors.

4. Record the score before acting, and check the outcome later.

Start your own evaluation with a note made before the action: date, score, stage reached, the chosen condition and the number of observed cases. Set the window in which you'll check for a later purchase. Don't reconstruct the assessment from a profile after the purchase — you'd be using information that didn't exist earlier.

Evaluate a sample from a later period than the data used to design the rule. Keep the same definitions and observation window. If you change the threshold, the discount and the message content at the same time, you won't be able to tell what affected the result. With a new promotion or a change in your product range, check again whether the threshold is still useful.

5. Turn intent into conditions and a contact tag.

The recognized profile below has 95/100, six visits, 40 page views and activity in the cart and order form, but still zero purchases. You can check whether it meets the high-intent rule. The "Recognized" label doesn't yet confirm consent or a connection to the chosen platform — verify these before acting.

For a score-based scenario, set an ML trigger of at least 70, add a check for a purchase in the last 24 hours and send the "NO" branch to the wysoka_intencja (high intent) tag in GetResponse. The optional +25 points change the contact's score in GetResponse, not the Flowpredict score. The tag lets you pick an audience; you configure the message, exclusions and messaging limits in the target platform.

In the pre-cart variant, start for example with a visit to an address containing /kolekcja/. Check the Flowpredict score. To check that there's no recorded cart, use the product interest condition: any product, add to cart, at least one, selected period — "NO" branch. Check for the absence of a purchase separately. To react to a later rise in the score, choose a periodically evaluated ML trigger; a single listing view doesn't wait for the threshold to be crossed in the future.

6. Separate choosing the right group from message effectiveness.

First check whether the chosen score sorts interest: how many cases entered the group, how many later bought and how many buyers stayed outside it. Use a fixed window and the same purchase definition. Don't compare only percentages from groups of very different sizes.

Evaluate the impact of communication separately. People close to buying might have come back without a message, so a purchase after tagging isn't proof of recovered sales. If you're evaluating impact, plan a comparable control group and the cost of the action. When a high score stops helping, check new events, traffic sources and sample journeys before you change the threshold.

7. After a purchase, clean up earlier labels.

Another, independent profile shows three visits, 22 page views and one PLN 669 order. The panel shows "Purchase completed" and no current intent score. That's information about the stage reached, not proof that earlier tags and messages were switched off automatically.

Set up a separate scenario: recorded purchase → remove the tag used before the purchase. In GetResponse, select the existing tag that the earlier automation actually added. Check that the action ran and the segment rules on the platform side. A drop in score alone doesn't remove the tag, and removing the tag doesn't guarantee that a message already queued for sending is withdrawn.