Do not turn tourism pageviews and coupon redemptions into merchant sales
Testing a new app requires separating observable platform actions from transactions the platform cannot verify.
Decide what a new app can observe
A tourism platform can test a limited improvement to learn whether a feature helps visitors. Testing more clicks and testing more merchant purchases require different evidence. Truthound can check an aggregate table that preserves metric type and denominator. It does not track people across services, infer revenue contribution or establish that a new application caused a visitor to make a purchase.
The official reply identifies an evidence boundary
In a parliamentary reply on 21 October 2026, Hong Kong described pageviews, outbound clicks and some campaign-specific coupon measures. It said merchant transaction information was unavailable and accurate new-versus-returning visitor ratios were difficult to establish. Turning those unknowns into 0 or a conversion rate for every visitor would exceed what the published evidence actually supports.
Do not put different units into one funnel
A proposed metrics table records metric name, denominator scope, channel, period and availability state. Coupon downloads and redemptions are different stages even within one campaign. An outbound click is not a merchant purchase. Begin the small experiment by preventing the platform’s total pageviews and a particular coupon programme’s redemption rate from sharing a denominator that neither source defines.
Choose a comparison the evidence can answer
If one information screen is improved, identify the behaviour observable on that screen first. A move to an external information page can be reported as a navigation measure. Testing purchases requires separate transaction evidence and appropriate authority to use it. Without that evidence, hold the revenue-effect question and narrow the exercise to a question the available records can actually answer.
Approve the metric pair before checking it
Truthound Core Python SDK 3.3.4 (open-source SDK, Polars backend) documents ColumnPairInSetValidator for operator-defined metric and denominator combinations. NotNullValidator could separately check required periods or availability states. The owner defines permitted pairs and missingness policy. A passing format check does not establish cookie identity accuracy, purchase occurrence or a causal campaign effect; those are different evidence questions that the aggregate contract cannot answer.
A maintained analytics contract can remain
Pandera 0.35.0 (MIT open-source Python DataFrame validation library; pandas, Polars and other supported backends) custom checks can handle the same metric-denominator relationship. Retain an existing contract that correctly manages its scope. Compare Truthound through the owner’s ability to handle exceptions and unknown measures, rather than an assumed ability to discover more conversions. Neither library automatically supplies missing merchant transactions or turns an unobserved outcome into a confirmed value merely by producing a clean report.
Define when to abandon the hypothesis
For the first exercise, confirm that compared periods and screen scopes match. If tracking or campaign population changed, do not interpret a simple increase as improvement. These are proposed review steps, not an executed experiment. If necessary transaction evidence cannot be obtained, the available choices include abandoning the sales hypothesis or preparing a separate, appropriately scoped measurement plan instead of estimating an unsupported conversion figure.
Leave room for an honest unknown
A platform experiment does not need to identify every visitor to be useful. Separating observable actions from unavailable transactions makes the next question clearer without exaggeration. A good metrics table explains whose action was counted, over which period and what remained outside observation. That is more useful for choosing the next experiment than a single impressive conversion rate whose denominator cannot be defended.
Conditions
Define metric denominator, period, channel and availability state in the aggregate table.
Alternatives
Pandera 0.35.0 custom checks can maintain the same metric-denominator combinations.
Limitations
Clicks and coupon redemption are not merchant sales, and unavailable conversion cannot be asserted as known.
Evidence
- Event checked
- Subject checked
- Modified
