TapCub is live — analytics, insights and live chat, free on one platform
Glossary

Novelty effect: definition, how to spot it, and a worked example

Launch-week numbers run high because users are trying the new thing, not because the feature won. Split new from existing users, read usage by days since first exposure, and wait for the curve to flatten.

Su He Reviewed by Lin Zhou Also called novelty bias Last updated Nov 16, 2026 Glossary

Quick answer

The novelty effect is a temporary lift (or dip) in metrics caused by users reacting to something being new — curiosity-driven trials inflate early usage, then decay toward a steady state. Judge a launch by days since each user's first exposure, not by its opening days.

Formula: Novelty premium(t) = existing-user usage rate(t) − new-user usage rate(t); steady state ≈ the flat new-user rate. Premium decaying toward 0 = novelty wearing off.

What is the novelty effect?

The novelty effect is the temporary change in user behavior caused by something being new: people notice a new feature, try it out of curiosity, and usage runs above its long-term level until the curiosity fades. Data from a feature's opening days often reads high — or low, when users stumble over an unfamiliar interface — because people are "trying the new thing"; judge the effect only after the novelty wears off, measured by days since each user's first exposure. The pattern is well documented in online controlled experiments: Microsoft's analysis of common metric pitfalls describes a coach-mark feature whose measured lift existed only during the visit where users first saw it, and vanished on every later visit.

The term comes from psychology, where novel stimuli attract attention and are remembered better than familiar ones. In product analytics it describes something narrower and more practical: the gap between what a metric shows right after launch and what it shows once the launch stops being news. The mirror image is the primacy effect: existing users resist or fumble a changed interface, so the metric first dips below its true value and recovers as users relearn it. Both effects make opening-day numbers a biased forecast of the steady state, in opposite directions.

How to tell novelty from real adoption

Three checks separate a curiosity spike from durable adoption, and none of them requires reading a single chart's first points:

  • Re-express time as days since first exposure, not calendar days. A calendar-time chart mixes users who saw the feature yesterday with users who saw it three weeks ago. When you re-index each user's behavior to their own day 0 (first use or first exposure), novelty shows up as a decay curve that flattens; real adoption shows up as a flat line from the start, or a curve that settles above the pre-launch baseline. Microsoft's pitfall paper recommends exactly this: segment the effect by days of the experiment or by user visits instead of trusting the aggregate.
  • Split new users from existing users. Users who joined after the launch never knew the old product, so they have no novelty to wear off and no habits to unlearn. If the launch-week lift appears only among existing users and new users behave the same before and after, you are looking at novelty, not a better product. If both groups settle at the same higher level, the change is real.
  • Watch repeat use, not first touches. Curiosity produces first clicks; value produces second, third and tenth uses. Track the share of first-time users who come back in week 2 and later — the same logic as reading a retention curve for the feature rather than for the whole product.

One caution before you diagnose anything: short windows are noisy. Microsoft's team notes that many apparent novelty or primacy trends in brief experiments are artifacts of cumulative plots, whose early segments simply have too little data. Confirm a decay on per-day rates with adequate denominators, not on a wobbly cumulative line.

A recomputable example: 14 days of a new feature

Sample data: a SaaS product ships a "Saved views" feature on Monday, October 5, 2026. On an average day, 1,000 users are active: 800 existing users (active before launch) and 200 new users (first seen that day). You record, for each day after launch, how many users in each group used the feature at least once.

Days after launchExisting users who used it (of 800)RateNew users who used it (of 200)RateAll active usersRate
1–7 (week 1)240240 ÷ 800 = 30%2424 ÷ 200 = 12%264264 ÷ 1,000 = 26.4%
8–14 (week 2)120120 ÷ 800 = 15%2424 ÷ 200 = 12%144144 ÷ 1,000 = 14.4%

Read the three columns against each other:

  • The aggregate lies twice. Week 1's 26.4% looks like a hit; week 2's 14.4% looks like collapse. Neither is the feature's steady-state value, because the aggregate blends a decaying existing-user curve with a flat new-user curve.
  • New users are the control. Their rate is 12% in both weeks — for them the feature was never new. That 12% is your best observational estimate of steady-state usage.
  • The novelty premium is the excess. Existing users ran 30% − 12% = 18 percentage points above steady state in week 1, then 15% − 12% = 3 points in week 2. The decay from 18 to 3 points is the novelty wearing off, on a schedule you can now see.

If you had judged the launch on week 1 alone (26.4% vs the 12% steady state), you would have overstated adoption by more than double: 26.4 ÷ 12 = 2.2×. Two weeks in, the honest headline is "12% steady-state usage among new users; existing users still converging" — a modest feature, not a breakout.

How long should the observation window be?

There is no universal number of days; the window has to outlast the decay. Rules you can apply before launch:

  1. Cover at least two full weekly cycles. Weekday and weekend behavior differs, so measure in whole weeks — 14 days minimum for a daily-use product, longer for weekly-use ones.
  2. Extend until the curve flattens. Using the example above: the excess over steady state fell 18 → 3 points between weeks 1 and 2. Declare stability only after two consecutive weeks where the existing-user rate stays within a pre-set band of the new-user rate (say ±2 points), not after the first week that "looks calmer."
  3. Match the window to the decision. A three-day promotion is supposed to produce a three-day bump — that spike is the product working, not a bias. A navigation redesign that will live for years cannot be judged on 24 hours of data.
  4. Do not delete the first week after the fact. Launch-week behavior is real usage and real cost. Report it separately if useful; dropping it post hoc to make the average prettier is how teams fool themselves.
  5. Check that late-window samples are still large enough. Late cohorts contain fewer users; a "flattened" curve computed on 30 people proves little. Size the window against your minimum detectable effect before you start.

Common mistakes

  • Judging on cumulative totals. Cumulative usage can only go up, so it hides decay. Use per-day or per-cohort-week rates.
  • Comparing launch week to the previous week with no control. A holiday, a marketing push or a traffic-mix change moves both weeks. Without a concurrent comparison you cannot attribute the difference to the feature at all.
  • Calling every early spike "novelty." Some spikes are campaign traffic or a logging bug firing duplicate events. Validate the data and the traffic source before reaching for a psychological explanation.
  • Ignoring the primacy direction. A redesign that dips in week 1 may recover as users relearn it; killing it on day 3 destroys a good change. The same days-since-exposure view catches both directions.
  • Treating attention as value. Clicks measure curiosity well and usefulness poorly. Watch a completion or repeat-use metric, and keep guardrail metrics — error rate, support load, unsubscribe — in view while the spike runs.
  • Letting the spike justify itself. Wanting the launch to succeed biases how you read its first chart; that is experimenter bias wearing a novelty costume. Pre-set the window and the stability rule before launch.

Measuring the novelty effect in TapCub

The scenario above — "Saved views" launched October 5, week-1 usage 26.4%, week-2 usage 14.4% — is exactly what you can rebuild in TapCub, with one honest boundary up front: TapCub does not run A/B experiments or split traffic. Everything below is observational. You can separate novelty from steady-state adoption the way the example does, but without a randomized control you cannot fully isolate the feature's causal effect from everything else that changed at launch.

  1. Mark the launch date. Add an annotation for October 5 in Configuration › Annotations so every chart shows where the feature shipped. Without it, next quarter's analyst will read the spike as a traffic event.
  2. Rebuild the daily decay curve. In All-platform analytics › Behavior › Event analysis, chart the feature's usage event by day for the 14 days after the annotation. This is the raw curve behind the example's two rows — check that week 2 is genuinely lower, not just visually calmer.
  3. Split new users from existing users. In All-platform analytics › Users › Segments & tags, build a rule segment on first-seen date (first seen on or after October 5 = "new users"), plus a segment for users active before launch. Segments defined once can be carried as cohorts into other analyses, so the same split works everywhere.
  4. Read retention by first use, not by calendar week. In All-platform analytics › Behavior › Behavior analysis, run retention with the feature's usage event as the start event, aligned to each person's own first use, once per segment filter. The new-user cohort's curve approximates steady state (the 12% line); if the existing-user cohort decays toward it, the gap was novelty. For the product-level view, lifecycle analysis on the same page counts New / Returning / Resurrecting / Dormant users per week, so you can see whether launch week brought back resurrecting users or only re-engaged people who were already active.
  5. Decide with the week-2 numbers. Report steady-state adoption from the new-user cohort and note the existing-user convergence, instead of announcing week 1's 26.4%. If the decision needs causal certainty — "did the feature cause the lift?" — run the experiment in a dedicated experimentation tool; TapCub measures what happened, it does not assign the variants. For what durable adoption looks like after the spike, see feature adoption.

Sources

Was this explanation clear?

We use your answers to fix and expand this page.

Su He

Data Analyst

Explores traffic sources, conversion funnels, and retention, using clear metric definitions and concrete examples to explain what the numbers mean.

FAQ

Questions people ask

Short answers to the follow-up questions that come up most.

Still have a question?

Chat with the team behind TapCub — we usually reply within a few hours on business days.

How long does the novelty effect last?

There is no fixed duration; it depends on how often users encounter the change. For daily-use products it commonly fades within two to six weeks. Extend the observation window until the existing-user rate stays within a pre-set band of the new-user rate for two consecutive weeks.

What is the difference between novelty effect and primacy effect?

They bias early data in opposite directions. Novelty: users explore a new feature out of curiosity, so metrics start above the steady state and decay. Primacy: existing users resist or fumble a changed interface, so metrics start below the steady state and recover as users relearn it.

Should you exclude the first week of data from analysis?

Not automatically. Launch-week behavior is real usage and real cost, so report it separately rather than deleting it after the fact. Base the steady-state judgment on later weeks and on users grouped by days since first exposure, with the window and stability rule set before launch.

Can you detect the novelty effect without A/B testing?

Partly. Observational checks — re-indexing usage to days since first exposure and comparing new-user with existing-user cohorts — reveal the decay and estimate steady state, as the 14-day example shows. But without a randomized control you cannot fully isolate the feature's causal effect from other changes at launch.

See this number on your own site

Add one script tag for web and apps, and the first reports fill in within minutes.

See your data clearly. Find your growth.

Every click, backed by data. Install one line of code and see your first numbers in a minute.

No credit card · Free plans for analytics and chat · Cookieless analytics