How to Analyze Amazon Reviews for Product Research (2026)

The Thirdwatch Amazon Reviews Scraper converts Amazon product URLs or ASINs into structured written-review data for product research. It returns ratings, titles, text, authors, dates, verified-purchase and Vine labels, helpful votes, media, and source links. Product managers, marketplace sellers, researchers, and data teams can use the output to identify recurring complaints, compare competing products, track new feedback, and build sentiment datasets without an Amazon account.

▶ Skip the setup: Run the product-research Task on Apify → — it is preloaded with the small, tested configuration from this guide. No code required.

Disclosure: Thirdwatch's Apify links include its referral parameter. This does not change your price.

Why scrape Amazon reviews for product research

Amazon review analysis turns unstructured buyer feedback into evidence for product, positioning, and merchandising decisions. A product manager can group complaints by feature, a marketplace seller can compare verified-purchase feedback across competing ASINs, and a research team can quantify how sentiment changes after a new version launches.

The commercial signal is meaningful. Northwestern University's Spiegel Research Center found that purchase likelihood for a product displaying five reviews was 270% greater than for a product with no reviews. That does not mean every review is equally useful. Star-only ratings help summarize satisfaction, while written reviews explain why customers were satisfied or disappointed. Verified-purchase and Vine labels add context, helpful votes identify feedback other shoppers found useful, and attached images can expose quality problems that rating averages hide.

The practical workflow is therefore to collect written reviews, preserve their product identity, normalize the fields, and analyze themes alongside ratings. The Amazon Reviews Scraper actor page provides that repeatable data layer.

How does this compare to the alternatives?

Amazon review research can be done manually, with a custom script, or through a maintained pay-per-result Actor.

Manual review reading remains valuable for final judgment, but it is a poor collection method. A DIY parser offers control, yet the engineering work is rarely the differentiator in a product-research project. The Thirdwatch Actor is best when the goal is to obtain consistent rows quickly and spend analyst time on categorization, sentiment, and decisions.

How to scrape Amazon reviews for product research in 4 steps

How do I choose the Amazon products to compare?

Start with a deliberate competitive set rather than a random list. Include your product, the category leader, a value alternative, and one or two products aimed at the same customer segment. Record the ASINs because they remain more stable than marketing titles.

products = {
    "our_product": "B0D1XD1ZV3",
    "category_leader": "B09V3KXJPB",
    "value_alternative": "B08N5WRWNW",
}

Use Amazon US for the strongest written-review coverage. Other supported markets can be selected with country, but anonymous review availability differs by storefront. A focused set of three to ten products is usually enough for an initial taxonomy of complaints and strengths.

How do I run the Amazon Reviews Scraper with Python?

The official Apify client starts the Actor and returns the dataset items. Store the API token in an environment variable rather than placing it in source control. Apify documents token management in its API authentication guide.

import os
from apify_client import ApifyClient

client = ApifyClient(os.environ["APIFY_TOKEN"])
run = client.actor("thirdwatch/amazon-reviews-scraper").call(run_input={
    "productUrlsOrAsins": list(products.values()),
    "country": "us",
    "maxReviewsPerProduct": 100,
    "sortBy": "recent",
})

reviews = list(client.dataset(run["defaultDatasetId"]).iterate_items())
print(f"Collected {len(reviews)} written reviews")

maxReviewsPerProduct applies separately to each ASIN. Start with 25–100 reviews to validate the analysis, then raise the limit for products with high review volume. Select recent for monitoring or helpful when building a stable research sample.

How do I group review complaints and strengths?

Begin with transparent keyword groups before adding a language model. This creates an auditable baseline and reveals whether your category vocabulary is sensible.

from collections import Counter, defaultdict

themes = {
    "setup": ["setup", "install", "instructions", "assembly"],
    "durability": ["broke", "broken", "durable", "sturdy", "lasted"],
    "value": ["price", "expensive", "worth", "value"],
    "delivery": ["shipping", "delivery", "package", "damaged"],
}

theme_ratings = defaultdict(list)
for review in reviews:
    text = f"{review.get('title', '')} {review.get('text', '')}".lower()
    for theme, words in themes.items():
        if any(word in text for word in words):
            theme_ratings[theme].append(review.get("rating"))

for theme, ratings in theme_ratings.items():
    clean = [r for r in ratings if isinstance(r, (int, float))]
    average = sum(clean) / len(clean) if clean else None
    print(theme, len(ratings), round(average, 2) if average else "n/a")

Review a sample from every theme before treating the counts as findings. Words such as “light” can be positive for portability and negative for perceived durability. Human validation is the difference between a useful taxonomy and a misleading dashboard.

How do I compare verified and unverified feedback?

Segment the dataset instead of deleting one class. Verified-purchase status is useful context, not a universal truth score.

import pandas as pd

df = pd.DataFrame(reviews)
summary = (
    df.groupby(["asin", "verified_purchase"], dropna=False)
      .agg(
          reviews=("review_id", "count"),
          average_rating=("rating", "mean"),
          total_helpful_votes=("helpful_votes", "sum"),
      )
      .reset_index()
)
print(summary.to_string(index=False))

negative = df[df["rating"].fillna(5) <= 2]
negative[["asin", "title", "text", "verified_purchase", "url"]].to_csv(
    "amazon-negative-review-sample.csv", index=False
)

This produces both a quantitative comparison and a readable queue for product teams. Keep the source URL so analysts can check context when a finding appears important.

Sample Amazon review output

Each dataset row represents one written review, not one product or star-only rating.

The record below is an illustrative synthetic example, not a current Amazon observation.

{
  "review_id": "R2X9P4M7K3L8Q1",
  "asin": "B0D1XD1ZV3",
  "country": "us",
  "rating": 5,
  "title": "Comfortable and easy to use",
  "text": "Setup took only a few minutes and it has worked reliably every day.",
  "author": "Morgan",
  "date": "Reviewed in the United States on June 18, 2026",
  "verified_purchase": true,
  "vine_review": false,
  "helpful_votes": 7,
  "image_urls": [],
  "media_urls": [],
  "url": "https://www.amazon.com/dp/B0D1XD1ZV3",
  "source": "woot-amazon-reviews"
}

Use asin as the product join key, rating for quantitative cuts, and title plus text for qualitative analysis. Treat verified_purchase, vine_review, and helpful_votes as context variables. Optional media arrays can be empty because most reviewers do not attach files.

Common pitfalls in Amazon review analysis

The first pitfall is confusing ratings with written reviews. Amazon's headline total includes feedback with no review text, so the dataset count will often be smaller. The second is sampling only the most helpful reviews, which favors older feedback; use recent when measuring current product quality. The third is comparing products with very different launch dates without controlling for time. The fourth is using average rating alone, which hides specific failure modes and polarized feedback.

Privacy also matters. Reviewer display names are public but can still be personal data. Collect only what the research requires, limit retention, and avoid attempting to identify individuals. Finally, storefront availability changes. Amazon US provides the strongest coverage, while other markets are best effort. Thirdwatch normalizes supported records, deduplicates overlaps, and exposes the limitations so the analysis can remain honest.

Related Amazon product research use cases

• Build an Amazon product research tool
• Monitor Amazon competitor pricing and ratings
• Scrape Shopify reviews for product research
• Build a product sentiment database
• Read the complete guide to scraping reviews
• Browse all Thirdwatch data guides