Back to selected work
BAIDU · 2022 DIAGNOSIS & RECOMMENDATIONS

Did the product get worse, or did acquisition change?

New-user retention fell as acquisition expanded. I investigated who was arriving before attributing the decline to the product.

My role
End-to-end diagnosis
Original tools
SQL · Excel · Tableau
Deliverable
Diagnosis & recommendations

The problem

Next-day retention was declining. Was this a product problem, a change in acquisition quality, or a wider disruption?

What I did

  • Separated platform mix from changes within Android and iOS.
  • Compared channels, acquisition volume, release timing, and the stages of the decline.
01 / DON’T START WITH THE AVERAGE

Both the user mix and retention within platforms changed.

Overall next-day retention fell from 22.20% to 19.45%. Android’s share of new users rose from 70% to 77%, while retention also declined within both platforms.

EXHIBIT 01

Account for the full 2.75-point decline

An exact accounting bridge from 22.20% to 19.45% next-day retention: platform mix contributes minus 1.015 percentage points, Android retention minus 1.47 points, and iOS retention minus 0.265 points. The vertical axis starts at 18.5%.
Monthly registration cohorts: November → December 2022. A focused scale starts at 18.5%. Components reconcile before display rounding.
View data & definitions

Next-day returners divided by all new users in the registration cohort, with complete follow-up. Symmetric decomposition: mix changes use mean platform retention rates; within-platform changes use mean platform shares. This is an accounting allocation, not causal attribution.

Account for the full 2.75-point decline
ComponentChange (percentage points)
Platform mix-1.015
Android retention-1.470
iOS retention-0.265

What this changed: The aggregate decline could not be read as a single product effect. I needed to investigate the acquisition mix and what changed inside Android.

02 / FOLLOW THE ACQUISITION

More registrations came with weaker next-day return rates.

The historical investigation found weaker retention across the main paid app-store channels, while the cloud-drive referral channel stayed comparatively stable. Expansion timing and channel-level patterns made acquisition quality a stronger lead.

EXHIBIT 02

Acquisition volume and retention in one Android channel

Scatterplot of daily registration cohorts in Android App store 2. December cohorts generally contain more new users and have lower next-day retention than November cohorts. November uses filled circles and December open squares.
Each point is one daily registration cohort in App store 2. Association supports an investigation; it does not establish an acquisition effect.
View data & definitions

61 daily cohorts, November 1–December 31, with complete next-day follow-up. Next-day returners / registrations. Generic channel aliases are used. No regression effect or statistical significance is inferred from the reconstructed points.

Acquisition volume and retention in one Android channel
Registration dateNew usersNext-day returnersNext-day retention
2022-11-0137616.22%
2022-11-0243716.28%
2022-11-0338615.79%
2022-11-0436616.67%
2022-11-0545715.56%
2022-11-0648816.67%
2022-11-0741717.07%
2022-11-0843716.28%
2022-11-0953916.98%
2022-11-1052815.38%
2022-11-1144715.91%
2022-11-1251815.69%
2022-11-13601016.67%
2022-11-1454814.81%
2022-11-1549816.33%
2022-11-1660915.00%
2022-11-17661015.15%
2022-11-18561017.86%
2022-11-1956916.07%
2022-11-20691115.94%
2022-11-21691115.94%
2022-11-2258915.52%
2022-11-23651015.38%
2022-11-24771316.88%
2022-11-25701115.71%
2022-11-2662914.52%
2022-11-27751216.00%
2022-11-28831315.66%
2022-11-29711115.49%
2022-11-30691217.39%
2022-12-011552012.90%
2022-12-021742212.64%
2022-12-031482114.19%
2022-12-041331813.53%
2022-12-051602213.75%
2022-12-061652313.94%
2022-12-071351813.33%
2022-12-081351813.33%
2022-12-091632112.88%
2022-12-101531912.42%
2022-12-111261814.29%
2022-12-121401913.57%
2022-12-131602012.50%
2022-12-141391913.67%
2022-12-151221613.11%
2022-12-161451913.10%
2022-12-171531912.42%
2022-12-181261814.29%
2022-12-191221613.11%
2022-12-201471912.93%
2022-12-211421812.68%
2022-12-221161512.93%
2022-12-231261511.90%
2022-12-241461812.33%
2022-12-251291713.18%
2022-12-261111412.61%
2022-12-271301612.31%
2022-12-281401712.14%
2022-12-291171512.82%
2022-12-301101412.73%
2022-12-311321511.36%
Compare all Android acquisition channels
Monthly registration cohorts · Android
ChannelNovember new usersDecember new usersNovember retentionDecember retention
Cloud-drive referral6,0007,40017.00%17.00%
App store 12,0004,80020.00%17.00%
App store 21,7004,30016.00%13.00%
App store 31,5003,40019.00%16.00%
App store 41,2002,30019.00%16.00%
Other channels1,60090019.69%16.78%

Recommendation: Evaluate acquisition on retained users and downstream value, alongside registration volume.

03 / TEST THE OTHER EXPLANATIONS

One explanation did not fit every stage of the decline.

Android weakened while iOS was initially steadier; later, both platforms declined. Releases overlapped acquisition changes, and regional activity added evidence of a wider disruption. Those overlaps limited a clean product-versus-channel attribution.

EXHIBIT 03

Separate platform-specific changes from shared movement

November and December next-day retention trends for Android, iOS, and all new users on one percentage scale. iOS begins higher and is steadier initially; both platform trends weaken later. Line styles distinguish the series.
Trailing seven registration days, pooling returners and new users. All three lines use the same percentage scale.
View data & definitions

Registration-date cohorts with complete next-day follow-up. The first six points use shorter available history. Overall retention is calculated from all returners divided by all registrations, not an unweighted mean of platform rates. This timeline is illustrative; it is not a release experiment or a causal estimate of external events.

Separate platform-specific changes from shared movement
Registration dateAndroid pooled retentioniOS pooled retentionOverall pooled retention
2022-11-0118.62%32.16%23.18%
2022-11-0218.10%31.96%22.67%
2022-11-0317.97%31.94%22.67%
2022-11-0418.26%31.90%22.89%
2022-11-0518.20%31.96%22.74%
2022-11-0618.19%31.94%22.64%
2022-11-0718.23%31.96%22.73%
2022-11-0818.18%31.92%22.72%
2022-11-0918.27%31.95%22.72%
2022-11-1018.26%31.98%22.59%
2022-11-1118.12%32.00%22.47%
2022-11-1218.15%31.98%22.54%
2022-11-1318.11%32.02%22.51%
2022-11-1418.07%31.99%22.39%
2022-11-1518.05%31.98%22.31%
2022-11-1617.97%32.00%22.22%
2022-11-1717.96%32.02%22.21%
2022-11-1818.06%32.04%22.28%
2022-11-1918.03%32.06%22.26%
2022-11-2017.92%32.03%22.09%
2022-11-2117.80%32.04%21.91%
2022-11-2217.88%32.06%21.97%
2022-11-2317.96%32.06%22.09%
2022-11-2417.92%32.04%22.03%
2022-11-2517.79%32.02%21.81%
2022-11-2617.77%32.00%21.73%
2022-11-2717.92%32.03%21.87%
2022-11-2817.90%32.03%21.85%
2022-11-2917.78%32.04%21.75%
2022-11-3017.83%32.00%21.73%
2022-12-0117.57%32.13%21.44%
2022-12-0217.25%32.31%21.08%
2022-12-0317.14%32.43%20.92%
2022-12-0417.08%32.48%20.83%
2022-12-0516.92%32.59%20.59%
2022-12-0616.65%32.61%20.20%
2022-12-0716.51%32.63%20.05%
2022-12-0816.54%32.49%20.12%
2022-12-0916.52%32.33%20.06%
2022-12-1016.38%32.16%19.84%
2022-12-1116.31%32.01%19.76%
2022-12-1216.37%31.82%19.86%
2022-12-1316.45%31.67%19.93%
2022-12-1416.35%31.55%19.79%
2022-12-1516.31%31.41%19.70%
2022-12-1616.34%31.29%19.72%
2022-12-1716.25%31.15%19.65%
2022-12-1816.16%31.06%19.59%
2022-12-1916.06%30.94%19.53%
2022-12-2015.91%30.83%19.37%
2022-12-2115.78%30.74%19.19%
2022-12-2215.81%30.66%19.23%
2022-12-2315.75%30.52%19.25%
2022-12-2415.73%30.42%19.24%
2022-12-2515.65%30.24%19.07%
2022-12-2615.60%30.13%18.99%
2022-12-2715.65%30.02%19.06%
2022-12-2815.60%29.87%19.03%
2022-12-2915.41%29.69%18.86%
2022-12-3015.41%29.58%18.83%
2022-12-3115.34%29.40%18.72%

Recommendation: Continue release-specific checks while tracking channels and platforms separately; timing alone is not enough to assign blame.

THE DECISION IMPLICATION

Investigate acquisition quality before treating the whole decline as a product failure.

What I delivered: A staged diagnosis that narrowed the acquisition, platform, release, and external explanations. The work supports recommendations, without claiming a measured retention recovery.

Methods, definitions & source notes

Based on my historical new-user retention investigation. Original analysis: SQL, Excel, and Tableau. This English presentation, exact decomposition, and Python/vector charts are portfolio reconstructions.

The cohorts contain 20,000 November registrants and 30,000 December registrants. Android channel rows partition Android users; platform rows partition each month. Rates are computed from integer returner counts.

Registration cohorts and next-day returners
MonthPlatformNew usersNext-day returnersRetention
NovemberAndroid14,0002,52018.00%
NovemberiOS6,0001,92032.00%
DecemberAndroid23,1003,69616.00%
DecemberiOS6,9002,13931.00%

Releases, channel expansion, and external disruption overlapped. The original report’s correlations do not establish a controlled causal effect. The decomposition is newly calculated from the replacement figures, not a historical company attribution estimate.

Original reports, internal channel IDs, SQL, and customer records remain private.

Download chart data