What is impression discounting?

Impression discounting is a technique used primarily in recommendation systems to penalize or reduce the weight of items that have been shown (impressed) to a user multiple times without eliciting interaction. Over time, if a user repeatedly ignores a recommendation, its “score” is discounted so those non-engaged items are less likely to appear again.

In other words: if you keep showing someone the same content or product and they never click, impression discounting says, “Let’s dial this down.” That way, fresh or more relevant items can surface instead.

How It Works (Mechanics & Formula)

Baseline scoring: The recommendation engine assigns each candidate item a base score using whatever model or algorithm your system uses.
Discount factor application: For items that have been shown before (impressions), a discount function (linear, exponential, quadratic, etc.) scales down the base score.
Re-ranking: The system reorders the recommendations based on the discounted scores. Items with many non-clicked impressions drop lower.
Feature inputs: Discounting is often based on features like: how many times the item has been impressed, how long ago last impression was, user’s interaction history with that item, and position in the recommendation list.

Formally, a new score might look like:

new_score = orig_score × f(g(X))

where g(X) is a discounting function applied to features X (e.g. impression count), and f aggregates them.

Use Cases & Benefits

Recommendation systems (e.g. “People You May Know,” product recommendations) use impression discounting to avoid redundancy and “ad fatigue” of showing the same suggestions repeatedly.
Content personalization / news feeds: Prevent suggesting the same article over and over if a user doesn’t engage.
Increased acceptance rate: By reducing the likelihood of showing previously ignored items, the system surfaces fresher alternatives, which can improve click-through or interaction rates.

Considerations & Trade-offs

Discount intensity tuning: Apply too harsh a discount and you might suppress content that could still be relevant later; too light, and you don’t reduce fatigue.
Cold start or new items bias: Newly introduced items may get prioritized, possibly displacing stable performers.
User behavior changes: A user’s preferences might evolve—discounting based on past non-engagement may wrongly penalize resurfacing content that becomes relevant later.
Computational complexity: Tracking impression history, computing discount factors, and re-ranking in real time requires infrastructure support.