# Impact Analytics

The **Impact** page (Recommender > Impact in the dashboard) shows what the
recommendations are doing for your store: how often they are shown and
clicked, how much revenue and how many orders they are credited for, whether
they really *add* revenue (A/B holdout), how baskets with a recommended
product compare with the rest, and a month-by-month history that never
expires.

All numbers come from the events your storefront sends
(`rec-impression`, `rec-click`, `add-to-cart`, `buy`, `detail-page-view`);
see the widget embed and WooCommerce plugin docs. Shopify stores do not get
this page yet: their events are recorded by the Shopify app without the
shopper identity attribution needs.

## Time windows and months

- The **7 / 30 / 90 days** switcher recomputes everything live from raw
  events, ending right now.
- **Month by month** uses UTC calendar months (`2026-08` = 1 Aug 00:00 UTC to
  31 Aug 23:59 UTC) and is refreshed every night. The current month is
  marked *in progress*; a month is final once it has been refreshed at least
  a week after it ended, so late order syncs are included.
- The **request counter on the Billing page** follows the same convention:
  it restarts from zero on the 1st of every month (UTC). Nothing is lost:
  every earlier month is listed under *Requests by month*.

## Attributed revenue and orders

A purchase is credited to the recommender only when the shopper actively
engaged with a recommendation for **that same product** within 30 days:

- **click-through**, where the shopper clicked the product in a recommendation
  widget and later bought it, or
- **cart-through**, where the shopper added the product to the cart after
  seeing (or clicking) it as a recommendation, and later bought it.

Merely *seeing* a recommendation never credits a sale. Attributed revenue is
the value of the credited order **lines**, not the whole basket. *Share of
sales* is attributed revenue divided by all order revenue the store reported
in the window.

## Engagement, funnel and efficiency

- **Recommendations shown / clicks / CTR** count impressions and clicks of
  recommended products, broken down per widget type and per recommendation
  source (order history, AI complements, visual similarity, fallbacks…).
- **Recommendation funnel**: shown, then clicked, then added to cart, then
  ordered, with the conversion rate of each step.
- **Revenue per 1,000 impressions** and **revenue per click** are attributed
  revenue divided by impressions (×1,000) and clicks. They are the quickest
  way to compare periods, widgets or sources of different size.

## Per widget and per source

The engagement tables break the same numbers down by widget type ("You may
also like", "Frequently bought together", cart, trending…) and by the source
that produced each recommendation (order history, AI complements, visual
similarity, fallbacks). Both carry the **revenue and orders credited** through
that widget or source and the **revenue per 1,000 impressions** (and, per
source, per click), so you can compare slots and algorithms of different size
fairly.

## Click-through by position

CTR per display slot (1st, 2nd, 3rd… recommendation). Clicks concentrate in
the first slots; slots that never earn a click only cost page space, so this
table answers "how many recommendations should I show?". Slots beyond 20 are
folded together; events logged before position tracking appear as *untagged*.

## Catalog coverage

The share of your catalog that was recommended at least once in the window
(distinct products shown ÷ products in the catalog), plus how many distinct
products the recommender has sold. Low coverage with high revenue means the
recommender leans on a few bestsellers; rising coverage means it is moving
the long tail.

## Shoppers who click recommendations vs those who don't

Shoppers who clicked at least one recommendation compared with shoppers who
viewed products but never clicked one: how many bought, their conversion rate
and revenue per shopper, summarised as "clickers convert N× more often". Only
identified shoppers count (customer or session id); server-synced orders
without a shopper id are excluded. Like the basket split this is
**correlational** (engaged shoppers may simply be more motivated buyers), so
use the A/B holdout for the causal answer.

## Basket size with vs without recommendations

Orders are split into those that contained **at least one credited
(recommended) product** and those that did not. For each group you see the
number of orders, the average order value and the average number of items.
The headline ("baskets with a recommended product are X% larger") compares
the two averages.

This is **basket value, not attributed revenue**, and it is correlational:
shoppers who engage with recommendations may also be more motivated buyers.
Use it as a strong indicator; use the A/B holdout below for the causal answer.

## Incremental revenue (A/B holdout)

With the holdout enabled, a fixed percentage of shoppers (default 10%) is
deterministically assigned to a **control** group that sees no
recommendations. Comparing revenue per visitor and conversion between the
control and **exposed** groups gives the revenue the recommender genuinely
*added*. The page flags results as statistically significant (~95%) or warns
while the control group is too small to mean anything.

How the groups are built:

- A shopper's group follows their **widget session**, not their customer id,
  so someone who browses anonymously and logs in at checkout stays in the
  group that decided what they saw. Custom integrations should send
  `session_id` on every event and order (see the
  [API reference](v1-api-reference.md)); the WooCommerce plugin already does.
- **Visitors** are shoppers with at least one product view, add-to-cart or
  purchase. Recommendation impressions and clicks never count a visitor on
  their own. Only exposed shoppers can produce them, so they would inflate
  the exposed denominator.
- Revenue is heavy-tailed, so a handful of large orders in the small control
  group can swing revenue per visitor by several percent in either direction.
  Read negative lift as "no measurable effect yet" until the window is long
  enough for the significance flag.

## Top recommendations

The products that earned the most attributed revenue, with the number of
credited orders, units (when your integration sends quantities), add-to-carts
and the current unit price, so a large amount next to a cheap product reads
as "many units", not a mispriced item.

## Export

*Month by month* has an **Export CSV** button with one row per month (all the
columns above, plus incremental revenue, the clicker conversion multiple,
catalog coverage and the currency).
