Attribution models and metrics
The Frosmo Platform measures the performance of recommendations in two ways:
-
Modification statistics measure every modification, including recommendation modifications. They use the modification-based conversion attribution described in Conversion attribution.
-
Recommendation attribution statistics measure recommendations only. They offer six attribution models, and they track what visitors do with recommended items before they buy.
Both count displays, true displays, and clicks the same way. They differ in how they give credit for conversions. As a result, the same recommendation can show different conversion figures in each.
This page describes both, and explains why their figures differ.
The recommendation statistics in More > All Reports > Recommendations use a third, separate rule: a 30-day period from a click on a recommended item to the purchase of that item. For more information, see Conversion attribution for recommendations and Recommendation statistics.
Where you can view the statistics
| View in the Frosmo Control Panel | Modification statistics | Recommendation attribution statistics |
|---|---|---|
Yes | No | |
Recommendations overview > Analytics, for several selected recommendation modifications | Yes, as the Modification revisions analysis | Yes, as the Recommendation attribution analysis (default) |
Recommendations overview > Analytics link of a single recommendation modification | Yes, as the Modification chart | Yes, as the Recommendation attribution chart |
API placements > Analytics for a single API placement | No | Yes |
API Recos Analytics | No | Yes |
An API placement is a named location on the site for which the site requests recommended items directly from the Frosmo Platform. It has one or more recommendation strategies that serve the recommended items. An API placement is not the same as a modification placement.
Only Frosmo users can currently access Recommendations overview and API Recos Analytics. For more information, contact Frosmo support.
Basic concepts
Before reading the statistics, note the following:
-
Displays are not views. A recommended item gets a display when it is rendered on the page, even if it is below the fold. It gets a true display only after it has been visible in the browser viewport for 3 seconds without interruption. Recommendation attribution statistics track true displays separately for each recommended item. For example, if a slider shows two of its six items, only those two items get a true display.
-
Attribution requires a tracked visitor. A conversion or add to cart that the platform cannot link to a visitor cannot be linked to the visitor's earlier displays and clicks either, so it is never attributed.
-
Attributed does not mean caused. Attribution gives a recommendation credit for a conversion because the visitor saw or clicked the recommendation before converting. It does not prove that the recommendation caused the conversion: a visitor who would have bought anyway produces exactly the same statistics. To measure how much a recommendation actually adds, compare its visitors against the comparison group or run an A/B test. Use attributed figures to compare variations and to follow trends.
Attribution period
Both kinds of statistics use the site's attribution period, which is the same for clicks and displays:
-
If conversion tracking on the site is session-based, only displays and clicks from the visitor's current session count. This is the default.
-
Otherwise, displays and clicks count for 1 day, unless the site has a different period configured.
For more information, see Conversion attribution time limit for modifications.
Modification statistics
How modification statistics attribute conversions
For each visitor, the Frosmo Platform keeps only the following across the whole site:
-
The last modification the visitor clicked
-
The last modification that got a true display for the visitor
-
The last modification that got a display for the visitor
When the visitor completes a conversion, the platform attributes the conversion as follows:
-
If the visitor clicked a modification during the attribution period, that modification gets the conversion after click, the conversion after true display, and the conversion after display.
-
If the visitor did not click a modification, but a modification got a true display, that modification gets the conversion after true display and the conversion after display.
-
If no modification got a true display, the modification that last got a display gets the conversion after display.
The platform attributes the conversion to the variation the visitor saw, with the full value of the conversion.
This has the following consequences for recommendation modifications:
-
All modifications compete for the same conversion. If the visitor clicks a recommendation and then clicks a banner, the banner gets the conversion.
-
A click takes all the credit. A modification that the visitor saw after the click, but did not click, gets no credit at all.
Modification metrics
| Metric | Description |
|---|---|
Displays, True displays, Clicks | Number of displays, true displays, and clicks the variation got. |
CTR | Clicks divided by displays, or by true displays if you calculate with true displays. |
Conversions (click) | Number of conversions attributed to the variation because the visitor last clicked it. |
Conversions (display), Conversions (true display) | Number of conversions attributed to the variation because it last got a display or a true display. These include the conversions counted in Conversions (click), since a click also gives the variation the display credit. |
CR (click) | Conversions (click) divided by clicks. |
CR (display), CR (true display) | Conversions (display) divided by displays, and conversions (true display) divided by true displays. |
Total revenue (click), (display), (true display) | Total value of the conversions counted in the corresponding conversions metric. |
A conversion counts once, however many items the visitor purchased.
You can limit the conversion and revenue metrics to a single conversion type, or select All to combine them.
If the modification has the comparison group enabled, its statistics are shown in a separate CG row. When you view several modifications together, the table shows totals with and without the comparison group.
Modification chart metrics
The modification chart shows the statistics over time, with one line for each variation. You can view the chart by day, week, or month. In addition to the counts and rates above, the chart offers cumulative metrics, where the value is summed from the start of the selected time range:
-
Cumulative click-through rate
-
Cumulative conversions after display and after click
-
Cumulative conversion rate after display and after click
-
Cumulative average order value after display and after click: cumulative revenue divided by cumulative conversions
The platform calculates cumulative rates from the cumulative counts, not by averaging the daily rates. A day with no activity counts as zero, so a cumulative line stays flat on that day.
Recommendation attribution statistics
How recommendation attribution statistics attribute conversions
Recommendation attribution statistics are built for recommendations only:
-
Only recommendations compete for the conversion. Clicks on other modifications do not take credit from a recommendation.
-
Displays and clicks are tracked for each recommended item.
-
Each conversion is attributed under six attribution models at the same time. You select the model whose statistics you want to view, and you can switch between models without recalculating anything.
The platform attributes a conversion to one of the following:
-
For a recommendation modification, the variation the visitor saw.
-
For an API placement, the recommendation strategy that served the recommended items.
The attribution period is measured backwards from the time of the conversion.
Attribution models
| Attribution model | Gets the credit | Share of the conversion |
|---|---|---|
Last click (default) | The recommendation with the most recent click on any recommended item. | The whole conversion and its full value. |
Direct click | For each purchased item, the recommendation with the most recent click on that same item. Purchased items that the visitor did not click in a recommendation give no credit. | One conversion for each matched item, with the value of that item. |
Last true display | The recommendation with the most recent true display, whether or not the visitor clicked anything. | The whole conversion and its full value. |
Last display | The recommendation with the most recent display or true display, whether or not the visitor clicked anything. | The whole conversion and its full value. |
Fractional (linear) | Every recommendation that the visitor clicked. Displays and true displays do not count. | An equal share. For example, with three recommendations, each gets one third of the conversion and its value. |
View-through | The recommendation with the most recent display or true display, but only if the visitor clicked no recommendation. | The whole conversion and its full value. |
Which model to use depends on what you want to know:
-
Last click is the default and the standard choice. It gives credit for any purchase that follows a click on a recommendation, including purchases of items that were not recommended.
-
Direct click is the strictest model. Use it to find out whether visitors buy the items they clicked in a recommendation.
-
Fractional gives credit for the same conversions as Last click. It is useful when visitors click several recommendations before they buy, and you do not want the last one they clicked to take all the credit.
-
Last true display, Last display, and View-through give credit to recommendations that visitors saw but did not click. They give the highest figures and the weakest evidence.
The Last true display and Last display models were added later than the others, and conversions from before they were added have no attribution under them. For a time range that starts before that point, both models show lower figures than the other models.
Until [deploy date], the Fractional model also gave credit to recommendations that got a true display. Conversions from before that date keep that attribution. For a time range that includes that date, Fractional figures mix the two definitions: before the date, recommendations that visitors saw but did not click get a share of the credit, and after it they get none.
Example: Attributing a conversion to recommendations
During the attribution period, a visitor:
-
Sees recommendation A, which gets a true display.
-
Clicks item X in recommendation B.
-
Sees recommendation C, which gets a true display.
-
Buys item X for 60 euros and item Y for 40 euros in a single transaction worth 100 euros.
The following table shows how each attribution model attributes the conversion. For comparison, the last row shows the result in modification statistics, if A, B, and C are recommendation modifications.
| Attribution model | Result |
|---|---|
Last click | B gets 1 conversion and 100 euros. |
Direct click | B gets 1 conversion and 60 euros. Item Y was not clicked, so it gives no credit. |
Last true display | C gets 1 conversion and 100 euros. |
Last display | C gets 1 conversion and 100 euros. |
Fractional | B gets 1 conversion and 100 euros. A and C only got true displays, so they get no credit. |
View-through | No recommendation gets credit, because the visitor clicked a recommendation. |
Modification statistics | B gets the conversion after click, after true display, and after display, each worth 100 euros. C gets nothing. |
Recommendation attribution metrics
The metrics are listed in the order a visitor moves from seeing a recommendation to buying.
If you select Use true displays to calculate rates, every rate uses true displays as its denominator, and View item and Add to cart count only items that got a true display. The setting also applies to charts.
| Metric | Description | Changes with the attribution model |
|---|---|---|
Displays, True displays, Clicks | Number of displays, true displays, and clicks the recommended items got. | No |
CTR | Clicks divided by displays. | No |
View item | Number of times visitors viewed the product page of an item after clicking that item in the recommendation. | No |
View rate | View item divided by clicks. | No |
Add to cart | Number of times visitors added an item to the cart after clicking that item in the recommendation. | No |
ATC rate | Add to cart divided by clicks. | No |
Attributed conversions | Number of conversions attributed to the recommendation under the selected model. | Yes |
Attributed revenue | Value of the conversions attributed to the recommendation under the selected model. | Yes |
CR (click) | Attributed conversions divided by clicks. | Yes |
CR | Attributed conversions divided by displays. | Yes |
AOV | Attributed revenue divided by attributed conversions. Not shown for the Direct click model, which shows revenue per click instead. | Yes |
Attach rate | Percentage of all transactions on the site that were attributed to at least one recommendation under the selected model. | Yes |
When reading the metrics, note the following:
-
Attributed conversions can be decimal numbers. With the Fractional model, a recommendation can get 0.33 conversions. With the Direct click model, a transaction that includes two clicked items counts as two conversions.
-
View item and Add to cart show a correlation, not a cause. The platform matches a product view or an add to cart to a recommendation afterwards, by checking whether the visitor clicked or saw the same item in a recommendation during the attribution period. The figures are the same for every attribution model.
-
View rate can be over 100%. A visitor can view the same item several times after one click, for example, by reloading the page.
-
Attach rate is calculated for the whole site. It does not change when you select different recommendations.
-
Totals and rates are calculated from the summed counts, not by averaging the rates of individual rows or days.
Time range and data retention
-
Charts do not include the current day. The platform calculates the daily statistics overnight, so charts show statistics up to the previous day, and their default time range is the last 30 days ending yesterday. The Recommendation attribution table in Recommendations overview includes the current day.
-
The statistics of the last three days can still change slightly while the platform processes late data.
-
Conversions are dated by the time of the purchase, not by the time of the click.
-
Detailed displays, clicks, product views, and adds to cart are kept for two months. After that, only daily totals remain. Attributed conversions and revenue are kept permanently. For a long time range, the counts can drop to zero at the start of the range while attributed conversions continue. This means the data is missing, not that the recommendation performed worse.
Why the two kinds of statistics differ
For the same recommendation modification and time range, Conversions (click) in the modification statistics and Attributed conversions under the Last click model usually differ. The following table lists the reasons.
| Modification statistics | Recommendation attribution statistics | |
|---|---|---|
Competing content | All modifications on the site | Recommendations only |
Tracked for | The whole modification | Each recommended item |
Attribution period | Counted in whole calendar days, from the time the conversion is processed | Counted exactly, from the time of the conversion |
Display credit when the visitor also clicked | The clicked modification | The recommendation that last got a display |
Attribution models | One | Six |
Conversion types | All conversion types, or a selected type | Transactions only |
Product views and adds to cart | Not tracked | Tracked |
Comparison group | Shown in a separate CG row | Not shown, since nothing is tracked for visitors in the comparison group |
Use modification statistics to compare a recommendation with other modifications on the same terms. Use recommendation attribution statistics to compare recommendations and their variations with each other, and to follow visitors from a display to a purchase.