# Methodology

## 1. What this dataset is

A de-identified, book-level record of the royalties earned by 540 titles in one Amazon KDP catalog, plus monthly and publication-cohort aggregates. Data cut: **2026-04-19**. The dataset is published so other people can check the claims in WriteAIBook's 350-book case study, reuse the numbers with attribution, and test their own assumptions about catalog economics.

It is **not** a random sample of Amazon KDP, a survey of KDP authors, or a controlled experiment. It is one publisher's account, primarily Amazon.de.

## 2. Source and collection window

- Source: Amazon KDP royalty reports (eBook royalties and Kindle Unlimited KENP reads), consolidated in a private workbook (`KDP_Analysis.xlsx`).
- Unit of analysis: one published title (one ASIN). A title can have sales in more than one marketplace; its royalties are aggregated to EUR.
- Lifetime window: all royalties recorded through 2026-04-19, including legacy titles from 2019–2022.
- Monthly window: 2025-01 through 2026-04.
- The catalog is primarily Amazon.de. Other marketplaces contribute a small share; the exact currency mix is not published per row.
- KDP can revise royalties after a report is downloaded. The last 90 days are a snapshot, not a final settlement.

## 3. Inclusion and the 350-book milestone

All 540 titles in the analyzed catalog are included:

- **524 AI-era titles**, published from March 2025 onward. "AI-era" is a **temporal cohort label**, not a verified per-title provenance flag. The catalog was built with an AI-assisted workflow, but the dataset does not contain per-book evidence of how much AI was used.
- **16 legacy titles**, published 2019–2022.

The 350-book figure used in earlier WriteAIBook posts was a mid-experiment milestone. The catalog reached 350 titles in late 2025 and kept growing. This release uses the full 540-title catalog because a partial snapshot would overstate per-book performance if the omitted titles were weaker.

## 4. Revenue definitions

- **Lifetime royalties (`lifetime_eur`)**: all recorded royalties for the title through the data cut, in EUR.
- **Last-90-day royalties (`last_90d_eur`)**: royalties recorded in the 90 days before the data cut.
- **KENP estimate**: Kindle Unlimited page reads are valued at **€0.0027 per KENP page**, calibrated to the reported April total. This is an estimate, not a per-transaction payout. eBook royalties are reported directly by KDP.
- **Gross, not net**: the figures do not subtract generation costs, editing, covers, advertising, software subscriptions, refunds beyond KDP's reported adjustments, or chargebacks. A title earning €30 did not necessarily profit €30.
- **Currency**: EUR. Amazon.de dominates; non-EUR marketplaces are converted in the source workbook. Small conversion differences are possible.

## 5. Status definitions

Each title is labelled by its last-90-day royalties at the data cut:

| Status | Rule | Meaning |
|---|---|---|
| `winner` | > €30 | Still selling; the catalog's current earners |
| `active` | €10–30 | Meaningful but modest revenue |
| `watch` | €2–10 | Fading or never found an audience |
| `dead` | < €2 | Effectively inactive in the last 90 days |
| `too_new` | published after January 2026 | Not enough time to judge |

These are descriptive thresholds, not a recommendation to delete books. "Dead" does not mean zero lifetime revenue; one title in the dataset had zero lifetime revenue at the cut.

## 6. Cohorts

A cohort is the calendar month of publication. `cohorts.csv` reports, for each month:

- books published;
- total and median lifetime revenue;
- median first-calendar-month revenue (available for the AI-era cohorts from the source workbook; blank for legacy cohorts);
- share of titles above €50 and €20 lifetime.

The July 2025 cohort is the outlier: 54 books, median first-month €13.80, median lifetime €82.42, 66.7% above €50 lifetime. The November 2025–March 2026 cohorts had median first-month revenue of €4.10–€5.70; the November 2025–April 2026 group’s lifetime hit rate is 6.7% as of the cut. Later cohorts are younger, so first-month medians are the fair comparison; lifetime hit rates are reported with that caveat.

## 7. De-identification

The public files remove:

- book titles;
- ASINs;
- pen names and author names.

Book IDs (`BK0001`…) are assigned after sorting by cohort, pen name and ASIN. Pen names are normalized with Unicode NFC (so visually identical accented names merge), then replaced with `PN01`…`PN08`. The private mapping is not published. The raw workbook is not redistributed.

This is de-identification, not anonymization against every possible re-identification attack. A determined person with access to Amazon sales data might still infer some books from cohort and revenue patterns. The dataset does not include titles or product links, which makes direct matching impractical.

## 8. Reconciliation and corrections

The builder recomputes every headline from the 540 book rows and refuses to publish if these checks fail:

- 540 rows, unique book IDs, non-negative revenue.
- Lifetime total = **€30,702.00**.
- Last-90-day total = **€6,251.85**.
- Monthly rows (16) sum to **€25,395.00**.
- The remaining **€5,307.00** is pre-2025 legacy revenue not represented in the monthly table. €25,395 + €5,307 = €30,702.
- Cohort rows sum to 540; status rows sum to 540.
- Peak month = **2025-09, €3,671**.

Corrections to earlier public claims:

1. **350 books / €18,000**: a mid-experiment snapshot. This release covers 540 titles through 2026-04-19. The AI-era lifetime total is €24,405.49; the full catalog is €30,702.00.
2. **September 2025 peak €3,825**: the workbook's summary sheet lists this figure, but the monthly table lists €3,671. The monthly table reconciles to the source rows; this release uses €3,671 and marks the €3,825 line as a summary-sheet error.
3. **Average is misleading**: the mean lifetime is €56.86, but the median is €33.62 and 80 titles (14.8%) earned less than €10 lifetime. The mean is pulled up by a small number of winners.
4. **Cohort lifetime hit rates are age-biased**: the November 2025–April 2026 books are younger than the July 2025 cohort. Use the first-month medians for the fair cohort comparison; lifetime hit rates are reported with that caveat.

## 9. Limitations

- **Single account, not a market sample.** Do not generalize to all KDP authors.
- **No causal identification.** The decline in revenue per catalog book coincides with catalog growth, but also with genre mix, cover quality, ad spend, seasonality, competition and KDP algorithm changes. The dataset cannot separate these.
- **No genre coding.** The source report's genre field was empty for all rows, so genre is not included. Claims about dark romance vs. sci-fi in earlier posts are not validated by this dataset.
- **No profit.** Revenue is gross royalties. Generation, editing, cover and ad costs are not included.
- **KENP is partly estimated.** The €0.0027/page figure is calibrated, not a per-read payout.
- **Survivorship and selection.** The catalog is the set of books that were published and remained in the account. Titles removed from sale may be missing.
- **Date cut.** Royalties can be revised after 2026-04-19. The last 90 days are especially provisional. This is a historical snapshot; it does not describe the catalog's state after the cut.
- **Age-biased lifetime cohorts.** Later cohorts have had less time to accumulate lifetime revenue. The first-month column controls for this; lifetime hit-rate comparisons do not.
- **No per-marketplace or per-format split** in the public row-level file.
- **De-identification limits.** See section 7.

## 10. Reproducibility

The builder is `research/kdp_book_study/build_dataset.py`. It reads the private workbook, de-identifies the rows, recomputes the aggregates, validates the totals, writes the CSVs/JSON/ZIP, and renders the public page. The source workbook is not in the repository.

```bash
python research/kdp_book_study/build_dataset.py \
  --source "$HOME/Downloads/KDP_Analysis.xlsx" \
  --data-dir src/static/data/kdp-book-study \
  --page src/static/research/kdp-book-study.html
```

Version 1.0.0. Data cut 2026-04-19. License CC BY 4.0.
