Click-through rate vs conversion rate: compare the same records

A click-through rate reports clicks relative to a stated opportunity to click. A conversion rate reports a defined outcome relative to a stated base. The base might be ad impressions, delivered messages, visits, clickers or another eligible group, depending on the report. The useful comparison starts by naming both the action and its denominator.

For example, “registrations among people who clicked” asks a different question from “registrations among everyone sent a message.” A report can use either definition, but the label needs to tell the reader which one it means. A click is also different from a unique person who clicked; repeated events can change the total.

In this study, the common base is a recipient–newsletter record in the released dataset. We count whether that record contains a click flag and whether it contains an investment flag. We do not have a delivered-message denominator or a timestamped sequence, so we call these recorded rates and overlap shares.

That distinction is useful when a business decides whether to change its message, its offer or its measurement. Two separate totals cannot identify the point where an individual stopped. Linking eligible records is a step toward answering that question; establishing the order and timing of events is another.

Researchers Lars Hornuf and Christoph Siemroth worked with bettervest, a German debt crowdfunding platform. Their experiment compared different ways of describing ten investment projects in newsletters. The released dataset contains 55,357 recipient–newsletter records and 7,864 recipient codes. A recipient can appear in as many as ten newsletters.^[1][2]

The authors recorded whether the recipient clicked a unique project link in a newsletter and whether that recipient invested a positive amount in the project. Each outcome is stored as a yes-or-no flag. An investment flag is not a count of transactions, a purchase value or a measure of completed ecommerce sales.

The original paper examines the effects of newsletter framing. Our reanalysis asks a different question: how much do the two recorded outcomes overlap for the same recipient and newsletter? We pool all three original treatment groups and retain the complete released sample. We do not re-estimate the framing effects.

These are historical records from a specific crowdfunding platform, released in 2023. The reviewed materials do not establish a common observation window or the order of clicks and investments. The study therefore supplies a measurement example, not a current email marketing conversion-rate benchmark.

A recorded click and investment appear together in 327 of the 55,357 records. Another 1,499 records have a click without an investment flag, 284 have an investment without a click flag, and 53,247 have neither. The table keeps all four possibilities visible.

Across the complete file, 1,826 records have a click flag, or 3.30%. A total of 611 have an investment flag, or 1.10%. Those are the two marginal rates. They say how common each flag is, but not how often the flags belong to the same records.

Among the 611 investment-flagged records, 327 also have a click flag. That is 53.52%; the remaining 46.48% have no recorded click. Turn the question around and the denominator changes: among the 1,826 click-flagged records, 327 also have an investment flag, or 17.91%.

Both percentages are correct because they answer different questions. The first starts with investment records and asks about clicks. The second starts with click records and asks about investments. Neither should be presented as the probability that a click led to a later investment.

An investment-only record does not explain its own history. It could not tell us that a person bypassed the newsletter, that a tracker missed an event or that another channel caused the investment. Those explanations would require evidence about collection, timing and other activity that this release does not provide.

For a campaign owner, the immediate implication is narrower and practical: keep the matched count alongside the separate totals. When the business question concerns people who took both actions, that intersection is part of the answer. A ratio made from two unmatched totals is not a substitute.

The overlap varies across the ten source newsletters. Among investment-flagged records, the share that also has a click flag ranges from 29.55% to 78.43%. We show every newsletter so that either endpoint cannot stand in for the complete study.

The newsletter numbers below are source codes, not a time sequence or a ranking. The smallest investment base is 31 records and the largest is 118. A large percentage on a small base should remain attached to that base when the result is quoted.

The full counts below make the comparison auditable. For any newsletter, add “both” and “investment only” to obtain its investment base. Add “both” and “click only” to obtain its click base. All four cells add to that newsletter’s total.

This variation does not identify the best newsletter creative. Projects differ, the same recipients can recur, and the new analysis pools the original treatment groups. There are no per-newsletter significance tests or claims that a particular project caused stronger overlap.

For a business with several campaigns, retain campaign detail alongside the combined report. If you are deciding how to change one campaign, start with its own row and counts. Keep the combined table when reporting across campaigns. Do not select a favorable campaign after seeing its result.

Our main comparison gives every recipient–newsletter record equal weight. A recipient represented in ten newsletters therefore contributes ten records, while someone represented once contributes one. That is appropriate for describing the released records, but it is not the only possible summary.

As a prespecified secondary comparison, we give each recipient code a total weight of one. Someone appearing in ten newsletters contributes one-tenth per record; someone appearing once contributes one. Each recipient code therefore contributes the same total weight. This does not make the sample representative of a country or market.

Under this weighting, 51.69% of the investment weight also carries a click flag, compared with 53.52% in the main analysis. The reverse share is 19.12%, compared with 17.91%. We retain both summaries; the secondary one is a different descriptive target, not a correction chosen because its result is preferable.

The weighted cell totals include fractions because they add contributions such as one-half or one-tenth. They sum to 7,864, the number of recipient codes. They should not be described as fractional people or as counts of unique investors who took each action.^[3]

Before interpreting a campaign as a funnel, write a short measurement brief. The following ungated template is for planning your own reporting. It does not assume that this historical dataset used the same definitions, or that changing measurement will improve sales.

Decision: “We want to decide [message, offer or reporting decision] for [campaign].” Choose the decision before selecting a convenient rate.

Unit and base: “One row represents [recipient–campaign, visit or another supported unit]. Our eligible base is [definition]. Repeated events and repeated recipients are handled by [rule].” Keep a count of events separate from a count of identifiers.

Actions: “A click means [specific recorded event]. An outcome means [registration, qualified inquiry, completed purchase or another defined action].” Decide how cancellations, refunds, test events and duplicates will be treated.

Link and timing: “We connect records using [permitted identifier]. We observe clicks during [window] and outcomes during [window]. We can establish order using [available timestamps].” If a link or timestamp is unavailable, say so instead of filling the gap with an assumed journey.

Report: “We will show click only, outcome only, both and neither, plus the eligible base. Conditional percentages will name their denominator.” Keep unavailable or unlinked records visible as their own reporting issue, rather than silently calling them “neither.”

Use: “This report describes [observed relationship]. To claim [a campaign caused more outcomes], we would need [a suitable comparison design].” The overlap table itself does not supply that causal comparison.

Consider a fictional creator with 1,000 eligible recipient–campaign records. Suppose the creator’s defined windows and permitted matching process produce 80 records with a click, 30 with a registration, and 18 with both. These are invented teaching values, not another result from the crowdfunding data.

The complete table contains 18 with both, 62 with a click only, 12 with a registration only and 908 with neither. The click rate is 80 divided by 1,000, or 8%. The registration rate is 30 divided by 1,000, or 3%.

Among click-flagged records, 18 out of 80 also have a registration: 22.5%. Among registration-flagged records, 18 out of 30 also have a click: 60%. The creator can report all four numbers without implying that they measure the same thing.

The 12 registration-only records deserve a neutral description until the creator has evidence about their paths. Even when timestamps establish that a click preceded a registration, order alone does not prove that the newsletter caused it. The creator can use the table to specify the next question without treating it as a revenue-lift estimate.

A useful completed brief would say: “We are describing recorded registration overlap for this campaign and these windows. We will keep all four categories, show both conditional bases and separately report records we cannot match.” That is more actionable than a conversion-rate label whose denominator nobody can identify.

We used the complete final Stata dataset from the authors’ version-one public release. Before calculating the joint outcomes, we checked the file identity, the seven required numeric fields, all 55,357 unique recipient–newsletter pairs, 7,864 recipient codes and ten newsletter codes. Both flags had to be present and coded exactly as zero or 100.^[2]

We checked the three treatment-group counts and reproduced the paper’s rounded marginal click and investment percentages. No records were removed, imputed or deduplicated. We then counted all four combinations overall and within every newsletter, and calculated the same quantities using the fixed equal-recipient weighting.

Two independently written implementations produced identical counts and exact fractions. The downloadable package contains original calculation code, the method and all aggregate tables. It excludes raw records and recipient identifiers. Obtain the source file separately from the authors and retain their attribution.

We do not attach sampling confidence intervals to this enumeration of the released file. The counts are exact for those records, but that does not establish precision for a wider population. Shared recipients and projects would matter for generalization, which this study does not attempt.

The practical contribution is the complete recorded overlap and a clear way to ask the next measurement question. It complements, rather than replaces, the authors’ original experiment. For campaign decisions, define the outcome, base, matching rule and time window before deciding what a reported conversion rate means.