Reviews and Q&A

Two surfaces carry the social proof: the reviews and the classic question and answer pairs. Both take an ASIN, and both are partly behind a login, which is the first thing to understand about them.

What is public

Amazon requires a sign-in for the review corpus. /product-reviews/ and /portal/customer-reviews/ both redirect to /ap/signin, and amz carries no session and wants none, so there is no version of this tool that pages through twenty thousand reviews.

What is public is on the detail page: the rating, the ratings count, the five bucket star histogram, and a medley of about a dozen reviews Amazon chooses. That is what amz reviews returns, and it says so on stderr on every run rather than printing thirteen rows as though they were the corpus:

amz: 13 of 21095. amazon requires a sign-in for the review corpus, and the detail page carries the histogram and the reviews medley only. the total is the ratings count, which is the largest number the page states
amz:   /product-reviews/ is not readable without a session
amz:   /portal/customer-reviews/ is not readable without a session
amz: run `amz why reviews` for the detail

The same fact is on the record, in reviews.loaded, reviews.total_count and reviews.complete, and in envelope.missed. A product nobody has reviewed and a product whose reviews amz was not allowed to read are different things and they serialize differently.

--deep attempts the standalone corpus anyway. It will fail, and it exists so that the failure is something you can reproduce rather than something you have to take on trust.

Reviews

amz reviews B075F5X8BR
amz reviews B075F5X8BR -n 50 -o jsonl

The review record

Each Review carries review_id, author, reviewer_name, rating, title, text, date, country, verified_purchase, helpful_votes, images, and variant_attrs (the format strip, parsed into key and value pairs such as colour and size). When the page has no stable id, amz derives a stable review_id so the same review hashes the same across runs.

author is a reference and it is almost always unresolved. The profile link goes to /gp/profile/, which robots.txt disallows, so amz has a display name and no identifier to file it under, and the record says that rather than dropping the name or inventing an id.

date is the line Amazon wrote plus the parse of it when one succeeded. A date amz could not read keeps its raw and has no parsed, because a failed parse that became a zero time would be January of year 1 and everything downstream would treat it as a real date.

The histogram

distribution is the five bucket star breakdown, and it carries derived for a reason. Amazon prints the buckets as whole percentages and not as counts, so the counts amz reports are the percentages multiplied by the total and they do not sum to the total exactly. derived: true is the record saying which of those two numbers was read and which was calculated.

Filters

These apply to the medley, locally, after it is read. They narrow what you are looking at, and none of them reaches a review that was not on the page.

Flag	Effect
--stars	only N-star reviews, 1 to 5
--verified	verified purchases only
--with-images	reviews that include photos

amz reviews B075F5X8BR --stars 1 --verified -o csv > one_star.csv
amz reviews B075F5X8BR --with-images -n 20

--sort orders what came back, and it is local too. It takes recent or helpful, and it reorders the medley rather than asking Amazon for a different one, because the surface that takes a sort parameter is the corpus and the corpus needs a session. A review whose date amz could not parse sorts last under recent rather than to the epoch, so an unparsed date does not masquerade as the oldest review on the page.

amz reviews B075F5X8BR --sort recent
amz reviews B075F5X8BR --sort helpful -o jsonl | jq -r '"\(.helpful_votes)\t\(.title)"'

The field is rating, not stars. --stars is the flag that filters on it, and jq 'select(.stars <= 2)' matches every review rather than none, because null compares less than a number in jq.

Just the URL

To open the review pages yourself, render the URL:

amz reviews B075F5X8BR -o url

Questions and answers

amz qa B075F5X8BR
amz qa B075F5X8BR -o jsonl

Each QA record carries qa_id, question, question_by, answer, answer_by, and helpful_votes.

The same login wall applies. The ask region on the detail page states how many questions have been answered and carries the pairs themselves only on some products, and the standalone page redirects to a sign-in whose assoc_handle is amzn_ask_us, which is the Q&A service asking rather than a generic redirect. Most products now have no ask region at all:

amz: this product has no ask region at all, which is now the usual case
no question and answer pairs on the page

That exits 3, the no-data code, rather than printing an empty table, so a script can tell "no questions" from "fetch failed". The answered count, when there is one, is also on the product record under questions.

Compose

A quick sentiment skim of the medley, one-star against five-star:

echo "1-star: $(amz reviews B075F5X8BR --stars 1 -o jsonl | wc -l)"
echo "5-star: $(amz reviews B075F5X8BR --stars 5 -o jsonl | wc -l)"

The histogram is the better instrument for the whole corpus, since it covers all 21,095 ratings rather than the thirteen reviews Amazon chose to show:

amz product B075F5X8BR -o json | jq '.[0].distribution'