Open rate measures whether an inbox registered an open, not whether a specific customer is moving closer to conversion.
That distinction matters because lifecycle teams often use a population-level metric to make an individual-level decision. A campaign reports a 32% open rate, another reports 24%, and the first campaign is labeled the stronger performer. But the average tells us very little about what any one person will do next. It does not reliably distinguish a customer who opened, compared products, and is ready to buy from one whose mail client preloaded an image and never noticed the message.
For individual decisions, the useful question is different: what sequence of observable behaviors usually precedes a conversion for this person, in this context, at this point in their lifecycle?
Why open rate is a weak individual signal
Open rate has always been an imperfect proxy for attention. It is generally inferred from a tracking pixel, so it depends on image loading, mail-client behavior, and reporting conventions. Apple Mail Privacy Protection made the problem more visible by automatically loading remote content in many cases. That can inflate recorded opens and make the timing of an open less meaningful.
There is also a statistical problem. A high open rate can indicate that a subject line attracted broad attention, but it does not tell us whether the recipient read the message, clicked through, or found the offer relevant. A customer who opens every email may be habitually clearing notifications. Another customer may rarely trigger an open but consistently click product links from a preview pane or use a tracked destination directly.
Open rate can still be useful for diagnosing deliverability and comparing broad creative patterns. It becomes unreliable when treated as a sufficient feature for selecting the next message for one person.
The event patterns that carry more signal
Individual-level prediction usually improves when the event stream captures actions closer to the decision itself. No single event is definitive, but several patterns tend to be more informative than an open in isolation.
- Click intent: A click on a product, plan, feature, or category link expresses more specific interest than an open. The destination matters. A click to a sale collection and a click to a returns policy should not be represented as the same event.
- Sequence and progression: A person who clicks an email, views the same product twice, visits a comparison page, and returns within three days has a different near-term likelihood from someone who clicked once six weeks ago. Order, spacing, and repetition are features, not implementation details.
- Recency and frequency: Recent activity usually matters more than historical activity, but excessive contact can reduce response. Useful features include time since the last click, number of sessions in seven days, number of messages received, and the interval between a message and a site visit.
- Commercial depth: Add-to-cart, checkout initiation, trial activation, account completion, and use of a core product action are closer to conversion than a message interaction. These events should be joined to messaging data without assuming that every later purchase was caused by the last click.
- Negative or cooling signals: Repeated delivery without clicks, a sharp drop in site activity, product returns, unsubscribes from a topic, and long periods of inactivity can indicate that the next message should change, pause, or narrow its scope.
The strongest features are often combinations. “Clicked in the last 48 hours” is useful. “Clicked a product link twice, viewed that product on the site, and has not received a promotion in five days” is more descriptive of the person’s current state.
From event logs to usable features
Raw events are not decisions. A lifecycle system needs to turn a stream of timestamps, URLs, message IDs, and customer actions into features that can be evaluated at send time.
We typically start with an identity map that connects a message recipient to web, application, and commerce events. Then we build time-bounded features: clicks in the last day, product views in the last week, days since the last purchase, and the number of similar messages recently delivered. Time windows prevent old behavior from dominating a current decision.
Context is equally important. A click on a product page may mean interest for one customer and support research for another. The same person may respond differently on a weekday morning than on a weekend, or differently after a purchase than before one. Features can include lifecycle state, product category, channel history, device, geography where appropriate, and the content a person has already seen.
There is a practical modeling principle here: features should describe a choice the system can act on. If the decision is which of three message variants to send, then recent category interest, prior response to similar content, and contact pressure may be more useful than a large collection of loosely related demographic fields.
A realistic campaign comparison
Consider a fictional ecommerce team comparing two reactivation campaigns. Campaign A has a 38% open rate and a 3.1% click rate. Campaign B has a 26% open rate and a 4.4% click rate. At the campaign level, the team might conclude that A has stronger creative because more people opened it.
Now examine the event paths. Among people who received A, many opened once and took no further action. Among people who received B, a smaller group clicked a category link, viewed two products, returned to the site within 24 hours, and added one item to a cart. Another group clicked a discount link but did not view a product. These are not equivalent responses, even though both groups contribute to the click rate.
For the first group, a product-specific reminder may be appropriate. For the second, a broader category message or a lower-frequency follow-up may be less intrusive. For customers with no activity after several messages, sending the same format again is not a neutral choice. A pause, preference prompt, or different channel may be more sensible.
A population report cannot make those distinctions. A per-person decision layer can use the observed sequence to select among eligible variants, provided the team has defined the conversion event and the message options clearly. The point is not to predict every purchase with certainty. It is to make a better conditional choice than “send the campaign that opened most often last month.”
Where the data falls short
Per-person engagement data does not solve attribution on its own. A customer may convert because of a price change, a sales conversation, a product need, or an ad that is not present in the lifecycle event stream. Correlation between a click path and a purchase is not proof that the message caused the purchase. Holdouts, careful experiment design, and clear conversion windows are still necessary.
The data itself can also be sparse. New customers have little history. Some people browse without logging in, use multiple identities, or move between devices. Not every CRM captures product views, checkout steps, message deliveries, or preference changes with enough detail to reconstruct a sequence. Missing events should not be treated as evidence that an action did not happen.
There are operational limits too. Event definitions drift, URLs change, consent rules affect collection, and timestamps from different systems may not align. A model can appear precise while consuming inconsistent inputs. For a small team, a short list of trusted events is often more valuable than a sprawling schema that no one maintains.
The practical standard is therefore modest and testable: collect the events closest to the decision, preserve their order and timing, separate reliable actions from noisy proxies such as opens, and evaluate whether person-level choices improve an agreed conversion outcome. Open rates still belong in the dashboard. They just should not be mistaken for a behavioral profile.