Articles published in the same batch — some are still climbing on day 90, some turn downward on day 45. Looking at data article by article, you’ll never see patterns. Content cohort analysis packages articles published in the same time period into one batch, horizontally comparing each batch’s performance at day 30, 60, 90, 180 after going live — whether topic selection strategy is getting better or worse, one table makes it clear.
Group by publishing batch, use “day N after going live” instead of calendar dates as the horizontal axis, and you can strip out seasonality and market fluctuations, directly comparing topic quality across different periods. Three tables are enough: batch growth table, retention rate table, batch structure table. They answer strategy questions, not single-article questions.
Why the single-article perspective deceives
Single-article reports use calendar dates as the horizontal axis — articles published in June and articles published in November are placed on the same timeline for comparison. When peak season comes, the whole site rises, new article performance gets overestimated; good content launched in off-season gets underestimated. Judgment由此 gets distorted.
The batch perspective changes the coordinate system: all articles’ day 1 aligns to their respective go-live dates. This way, what’s compared is performance difference under the same growth duration — the remaining gaps truly come from topic selection, structure, and conversion ability. To distinguish reader composition across different batches, you can overlay the segmentation method by device and region and look again.
Three steps to build your first cohort table
No complex tools needed — export an article list, add a “go-live date” column, and you can start. Spreadsheet tools are enough to run the first two rounds, once the口径 is stable then consider solidifying into scripts.
- Take the last 12 months of article list, record publish date, category归属, and topic tags, group into batches by month
- Pull each article’s organic click data, calculate cumulative clicks at day 30, 60, 90, 180 after going live
- Calculate median by batch rather than mean, avoid one or two viral articles lifting the whole batch’s numbers
| Publish batch | Article count | D30 median clicks | D90 median clicks | D180 median clicks | D180 retention |
|---|---|---|---|---|---|
| Feb batch | 12 | 85 | 340 | 610 | 92% |
| Mar batch | 15 | 110 | 520 | 880 | 96% |
| Apr batch | 14 | 95 | 300 | 390 | 61% |
| May batch | 16 | 130 | 610 | 1020 | 98% |
The retention rate column is calculated as day 180 month clicks divided by peak month clicks. The April batch’s problem is visible at a glance: not a bad start, but loses momentum after 90 days, only 60% left at half a year. Flip back to the topic list and you can often find commonality — for example, that month concentrated on a batch of time-sensitive topics.
Four types of signals read from the cohort table
- Overall upward shift: new batches outperform old batches at all time points,说明 topic selection standards are tightening, can increase investment in the same type
- Fast start, weak staying power: high D30 but low D180 retention, mostly trending-type topics, need to配 update scheduling to extend life
- Slow start, steady climb: typical long-tail content, don’t pronounce death on day 30, evaluation window at least to day 90
- Whole batch collapse: same batch collectively declines, first rule out algorithm fluctuations and site failures, use traffic drop layered diagnosis method to investigate layer by layer
The second type of signal is most easily misunderstood as poor content quality. The short cycle of trending topics is a topic attribute — to measure it you need to switch metrics: look at peak height and customer acquisition cost, not half-year retention. The premise of judgment is to first tag each batch with topic type labels.
Use cohort results to reverse-engineer topic selection and update scheduling
The most practical output of the cohort table is scheduling basis. Batches with retention below 60% enter the update queue, start refreshing from former high-peak pages — the decay time pattern can be对照 the content decay curve judgment nodes. Batches with retention over 90% then拆解 commonality, reuse the topic template for next quarter.
| Batch performance | Judgment | Scheduling action |
|---|---|---|
| High D90, high D180 retention | Long-term asset | Just补充 data and cases every half year |
| High D90, low D180 retention | Time-sensitive type | Schedule quarterly updates, or convert to evergreen version |
| Low D90, curve still rising | Long-tail climbing | Continue observing, add internal links to accelerate |
| Low D90, curve flattening | Topic deviation | Stop same-type investment, evaluate whether to merge |
Put this judgment table together with cost data, and you can calculate each batch’s actual return — the口径 can directly follow the content ROI calculation method. After doing it for 2-3 quarters, batch curves can also serve as prediction baselines, giving next quarter’s goals an evidence-based range — the method is in organic traffic forecasting modeling思路.
Three easy places to stumble
- Batch division too fine: weekly batches leave only 2-3 articles per batch, median loses meaning — monthly or bi-weekly is appropriate
- Mixing different content types: product pages, tutorials, checklists mixed in one batch, conclusions can’t guide any decision
- Ignoring redesign impact: major site navigation or template changes will整体 raise or lower certain batches, need to add a remarks column in the table
There’s also an operational detail: articles that have been substantially rewritten after going live should be removed from the original batch or separately labeled. Rewriting equals re-launching — mixing in old batches pollutes retention rates, and also makes update action effects impossible to attribute. Building a separate “rewrite batch” for these pages反而 can verify whether refreshing actually works.
Next step: four steps to run the first round
- Export the last 12 months of article list, group by month and tag with topic type and category
- Fill in each article’s click data at D30, D90, D180 three time points, take median by batch
- Use the judgment table to定性 each batch, produce an update queue and a stop-doing list
- Solidify the cohort table into a monthly auto-refreshing view, quarterly reviews only look at batch trends
FAQ
Why look at data by publishing batch?
Single-article with calendar dates mixes in seasonality and market fluctuations — the batch perspective aligns go-live dates, differences under the same growth duration truly come from topic selection and content quality.
How should batches be divided?
By month or bi-weekly — weekly leaves only 2-3 articles per batch, median loses meaning; different content types mixed in one batch, conclusions can’t guide any decision.
Should slow-start articles be abandoned?
Not necessarily. Low D30 but curve still rising is long-tail climbing type, evaluation window at least to day 90; fast start weak staying power is trending type, needs to配 update scheduling.
How is batch retention rate calculated?
Day 180 month clicks divided by peak month clicks. Batches with retention below 60% enter the update queue, start refreshing from former high-peak pages, can recover most traffic.
How to handle articles rewritten mid-way?
Remove from original batch, separately build a “rewrite batch.” Rewriting equals re-launching — mixing in old batches pollutes retention rates, independent labeling才能 verify refresh effects.
