Half a billion Amazon reviews

The category summaries — five files

The behavioural aggregates — 29 more files

The five files above summarise the review table one category at a time. A second set, merged_results_v2/ — 29 CSVs, 13.7 MB — instead groups every review by its author and by its product, which is what makes reviewer-level and item-level questions answerable at all. It is where every finding in the reviewer analysis comes from. Read merged_results_v2/_meta/manifest_v2.json first — it documents every file’s grain, suppression, and caveats, and records two measures it deliberately did not publish.

Grain is load-bearing, not decoration: C is one row per category, G is corpus-wide with no category column, X is a category × category matrix that must not be joined to per-category files, and C+G carries both, separated by a scope column. The two scopes answer different questions and are not comparable.

import pandas as pd

BASE = "https://ontopic-public-data.t3.storage.dev/amazon-reviews/merged_results/"
V2   = BASE.replace("merged_results", "merged_results_v2")

cats = pd.read_csv(BASE + "category_stats_all.csv")
oad  = pd.read_csv(V2 + "user_one_and_done.csv")

import boto3, pandas as pd

s3 = boto3.client("s3", endpoint_url="https://t3.storage.dev",
                  region_name="auto")
obj = s3.get_object(Bucket="ontopic-public-data",
                    Key="amazon-reviews/merged_results/"
                        "category_stats_all.csv")
cats = pd.read_csv(obj["Body"])

The bucket answers anonymous GETs on virtual-host style URLs (bucket.t3.storage.dev/key); the path-style form t3.storage.dev/bucket/key returns 403.

Six ways to get this wrong

1. Only ts_yearly is a timeline. The monthly, weekday, and hour files pool every year together. Plotting them left to right as a time axis produces a chart that means nothing.

2. Filter on count before trusting a rate. A category-year holding one review reports rating_5_pct = 100.0. Every rate chart here drops cells under 500 reviews.

3. Volumes span four orders of magnitude. 67.4M reviews in Home & Kitchen against 16,216 in Subscription Boxes. Normalise before you compare.

4. Percent columns are 0–100. Not 0–1. Dividing twice, or not at all, is the most common bug against these files.

5. Every mean rating is an upper bound. The Unknown category is excluded — 11% of ratings but 42% of users, and disproportionately one-and-done. Since sparse reviewers rate lower (3.81★ at one review, 4.26★ at ten or more), dropping them pushes every published mean up.

6. The variance shares do not decompose. user 30.2%, item 19.3%, category 0.7% are MARGINAL shares of one factor at a time. They are crossed and unbalanced, sum to 50.2% corpus-wide and 105.6% on Gift Cards, and partition nothing. Do not derive a residual from them.

Derived from McAuley-Lab/Amazon-Reviews-2023 and inherits its terms. Aggregation ran on Google Cloud Run, one job per category, streaming each raw_review_* split; the merged CSVs were migrated to Tigris in August 2026 with every object verified by MD5. Charts on these pages read a 33-category JSON built from those CSVs by scripts/fetch-amazon-aggregates.mjs.