What can my Steam reviews tell me?

Your reviews tell you why players stop, what they would pay for, and which problems are worth fixing first. The review score only tells you how many players were happy. This guide is about reading the text and the fields behind each review. How reviews affect sales, and how to get more of them within Valve's rules, is a separate guide.

Expect most reviews to carry little for the team. Lin, Bezemer, Zou and Hassan hand-coded a random sample of 472 English Steam reviews and found that 42% held feedback a developer can use: a pro, a con, a suggestion or a bug. They also coded 71% of the sample as holding something not helpful to a developer, such as an emotion with no reason, like "Good game!".

The same study found that players complain more about design than about bugs. Only 8% of the sampled reviews mentioned bugs, while 34% named a con of the game design. In negative reviews, 17% mentioned bugs and 57% named design cons. So a codebook that only tracks bugs misses most of what negative reviews say.

Sources: Lin, Bezemer, Zou and Hassan, Empirical Software Engineering (2019).

How do I export or collect my Steam reviews?

Call Valve's GetAppReviews Web API and page through the results. It is a GET request to api.steampowered.com, it needs no key, and it returns up to 100 reviews per page. The older store.steampowered.com/appreviews endpoint is deprecated, and Valve says GetAppReviews accepts the same filters and returns the same fields, with a few changes that it lists, such as a new refunded field.

Parameters go in one input_json value. To collect every review, sort by Recent (filter 1) or Updated (filter 2) and follow the cursor until a page comes back empty. Valve warns that the Helpful sort widens its day range until it finds results, so it is not a way to page through everything. Start with cursor "*", then send the cursor from each response.

Set purchase_type to 1 (All) for analysis. The default, 0, returns only reviews from players who bought the game on Steam, which matches the score but hides reviews from key activations. Set languages to ["all"]: an empty list means English only. Language values are Steam's API language codes, such as english, german, schinese and brazilian.

Mind the limits. Anonymous responses can be cached for up to 10 minutes, calls without a key associated with the app share a lower rate limit, and a call that is limited gets HTTP 429. Valve says to pass a key associated with the app for the higher rate limit. Our practice is a slow full export once, then a daily call with a date range for new reviews.

We ran the request below on 19 November 2026 against a live game. It returned 100 reviews, a cursor for the next page, a total_matching count and, on that first page only, the score totals. App 413150 is the public game we tested with: put your own app ID in its place.

GetAppReviews inputs and fields that matter for analysis, as Valve documents them, read 19 November 2026.
NameInput or fieldWhat it does
filterInput0 Helpful (default), 1 Recent, 2 Updated, 3 Funny. Use 1 or 2 to page through every review.
languagesInputA list of API language codes. Empty means English; "all" means every language.
review_typeInput0 All (default), 1 Positive, 2 Negative.
purchase_typeInput0 Steam purchases only (default), 1 All, 2 Non-Steam purchases.
date_range_start, date_range_endInputUnix times. Both must be set, or the range is ignored.
playtime_min_hours, playtime_max_hoursInputAuthor playtime in the app when the review was written.
cursor, num_per_pageInputPaging: start with "*"; up to 100 reviews per page (default 20).
filter_offtopic_activityInputOff-topic review bomb reviews are left out by default. Pass false to include them.
voted_up, language, reviewFieldPositive or negative, the language code of the review, and its text.
timestamp_created, timestamp_updatedFieldWhen the review was written and last edited.
author.playtime_at_reviewFieldPlaytime in minutes when the review was written.
steam_purchase, received_for_free, refundedFieldHow the author got the game, and whether they were refunded.
written_during_early_access, developer_responseFieldEarly Access flag, and the text of your official response.

One GetAppReviews request and the next page

curl -G "https://api.steampowered.com/IUserReviewsService/GetAppReviews/v1/" \
  --data-urlencode 'input_json={"appid":413150,"filter":1,"languages":["all"],"review_type":0,"purchase_type":1,"num_per_page":100,"cursor":"*"}'

# The response, shortened:
# { "response": {
#     "query_summary": { "num_reviews": 100, "review_score_desc": "...",
#                        "total_positive": ..., "total_negative": ..., "total_reviews": ... },
#     "cursor": "AoJ4iNCh+KEDeMXGlQc=",
#     "total_matching": ...,
#     "reviews": [ { "recommendationid": "...", "language": "german", "voted_up": false,
#                    "timestamp_created": ..., "refunded": false,
#                    "author": { "playtime_at_review": ..., ... }, "review": "..." } ] } }
#
# Next page: send the same input_json with "cursor" set to the cursor above.
# Stop when a page comes back with no reviews.

Sources: Steamworks IUserReviewsService and User Reviews Get List documentation; Languages Supported on Steam.

What does Steamworks offer developers for reviews?

Steamworks gives you tools to respond to and flag reviews, and points you to the API for the data. Valve's User Reviews page describes no download or export button for reviews. For reviews in a date range or with certain criteria, it links to GetAppReviews.

On each review, the moderator controls let you write an Official Developer Response, which shows under the review, and flag a review as abusive, off-topic or against the community guidelines. You cannot delete a review: a flagged review stays listed until a moderator removes it or clears the flag.

Two flags in the API come from that system. developer_response holds the text of your reply, so you can count which negative reviews already have one. filter_offtopic_activity drops reviews from a period Valve marked as an off-topic review bomb, which is the default; pass false when you want to study the bomb itself.

Sources: Steamworks User Reviews documentation; IUserReviewsService documentation.

How do I tag reviews with a codebook?

Write a short list of tags with a one-line rule for each, then give every review all the tags that fit. A codebook keeps two people, or a person and a model, tagging the same way, and it turns review text into counts you can chart.

Start small. The table below has eight complaint tags and one praise tag with sub-themes, which is our starting point, not a standard. Lin and colleagues built their categories by open coding: read a review, use an existing tag if one fits, and add a new tag only when none does. When you add a tag, go back over the reviews you already tagged.

Keep tags about the game, not about the mood of the reviewer. "Too short for the price" gets both length and price. A review that only says "10/10" gets praise and nothing else. Keep the codebook in a file next to the data, with the date of each change, because a new tag makes old weekly counts look lower.

A starting codebook for Steam reviews. Copy it into a sheet and change the rules to fit your game.
TagUse it when the review saysExample phrase
bugsSomething does not work as designed: a crash, a broken quest, a lost save"crashed when I opened the map"
performanceFrame rate, stutter, loading time or hardware load"drops to 20 fps in town"
priceThe price, the value for money, or the DLC price"not worth 20 dollars"
lengthThe game is too short, too long or padded"finished it in 90 minutes"
difficultyToo hard, too easy, unfair or grindy"the second boss is a wall"
controlsInput, key bindings, controller or Steam Deck support, camera"no controller support"
contentToo little variety or content, or a missing feature or mode"only three levels and no co-op"
localizationTranslation quality or a missing language"the German text is cut off"
praiseWhat the player liked; add a sub-theme: art, music, gameplay, story"the music is beautiful"

Sources: Lin, Bezemer, Zou and Hassan (2019) for the open coding method; the tags are our practice.

How do I count themes by week and by build version?

Group your tagged reviews by the week of timestamp_created, and by the build that was live at that time. Then count reviews per tag, and divide by all reviews in that week or build, so a busy sale week does not look like a new problem.

GetAppReviews has no build field. Keep your own table of build numbers and the time each went live, from your patch notes, and give each review the build that was live when it was written. Our practice is to treat the first day after a patch as mixed, because players who have not restarted Steam may still be on the old build.

Edited reviews need a rule. timestamp_updated changes when a player edits a review, and Valve says Steam may prompt a negative reviewer who keeps playing to update the review. For trends, count each review once at its created date. To see whether a patch changed minds, sort by Updated and read the reviews edited since that patch.

Read shares, not raw counts, and wait for enough reviews. Lin and colleagues found that 96% of games receive a median of fewer than 10 reviews per day. With 30 reviews in a week, one review moves a share by more than 3 points (our arithmetic), so compare builds rather than single days.

  1. Export all reviews with filter 1, languages all and purchase_type 1, and save recommendationid as the key so a review is never counted twice.
  2. Add a build column from your release table, using timestamp_created.
  3. Tag each review with the codebook.
  4. Make a weekly table: one row per week or build, one column per tag, with the share of negative reviews that carry it.
  5. Mark the dates of patches, sales and review bombs on the same chart.

Sources: IUserReviewsService documentation; Steamworks User Reviews documentation; Lin, Bezemer, Zou and Hassan (2019).

How do I read negative reviews by playtime?

Sort negative reviews by author.playtime_at_review, which is in minutes, and read them in bands. Each band points to a different part of the game, so the same tag can mean different fixes.

Lin and colleagues found a median of 6.6 hours of play before a negative review and 15.5 hours before a positive one, in 28,159 reviews with exact playtime. So more than half of the negative reviews in their data came within the first 6.6 hours. When they read the negative reviews with the longest playtime, many came from players who liked the game but were let down by other players, a recent update or the price of DLC.

The first two hours matter twice. Valve refunds a game on request within two weeks of purchase if it has been played for less than two hours, and GetAppReviews marks a review whose author was refunded. A negative review under two hours with refunded set to true is a player you lost on both counts.

Playtime bands for reading negative reviews. The bands are our practice; the 2 hour line is Steam's refund rule.
Playtime at reviewWhat it often points toCheck next to it
Under 2 hoursCrashes, performance, controls, a first hour that does not hookCrash rate in the first session, refunds
2 to 10 hoursDifficulty spikes, pacing, a mismatch with the store pageWhere players stop in the funnel
10 to 50 hoursLength, content, grind, endgameRetention after week 1
Over 50 hoursA recent update, other players, DLC priceReviews edited after the last patch

Sources: Lin, Bezemer, Zou and Hassan (2019); Steam Refunds policy; IUserReviewsService documentation.

How do I split reviews by language?

Group reviews by the language field and compute the share of positive reviews for each language. A low score in one language often points to something you can fix, such as a translation, a price or a server region.

Valve made this matter on 18 August 2025. Steam now shows players the review score of their primary language when a game has enough reviews in it. The change applies to games with more than 2,000 publicly visible reviews and at least 200 written reviews in at least one language. Players can switch back to the overall score in their store preferences, and the store page has a language breakdown.

Valve gave its reasons: translation issues, cultural references and poor network connections can give players in different regions very different experiences of the same game. Those are good tags to look for in a weak language. Lin and colleagues found a median of 36% non-English reviews per game in their data, so an English-only export can leave out a large part of your players.

Sources: Steam News, Updates to User Review Scores Based on Language (18 August 2025); Lin, Bezemer, Zou and Hassan (2019).

Can I use an AI model to tag my reviews?

Yes, as a first pass that you check by hand. A language model can apply your codebook to thousands of reviews in many languages in minutes. It also makes mistakes that look confident, so you need a measured error rate before you trust its counts.

The research supports both halves. Gilardi, Alizadeh and Kubli found that ChatGPT beat crowd workers in zero-shot accuracy on four of five annotation tasks, on a sample of 2,382 tweets, at less than $0.003 per annotation. Pangakis, Wolken and Fasching repeated 27 annotation tasks from 11 datasets with GPT-4 and found that performance depended heavily on the dataset and the task, so they argue that every task must be checked against human labels.

Our practice: send only the recommendationid and the review text, never the steamid. Give the model the codebook, ask for a quote for each tag, and run reviews in small batches so a bad batch is easy to find. Then tag a random sample of about 100 reviews yourself without looking at the model's tags, and compare per tag. Lin and colleagues used 96 reviews per group for a 95% confidence level with a 10% confidence interval, a useful size for a check.

If a tag agrees on less than about 8 of 10 reviews in your sample, rewrite its rule and run it again (our practice). Watch for sarcasm, jokes and very short reviews, and have a native speaker check a sample of the languages you sell most in.

Sample tagging prompt with two made-up reviews

You tag Steam reviews of our game. Use only these tags:
bugs, performance, price, length, difficulty, controls, content, praise:art,
praise:music, praise:gameplay, praise:story, other.

Rules:
- A review can have several tags. Use "other" only when no tag fits.
- Tag what the review says, not what you guess the player meant.
- For each tag, copy the shortest quote from the review that supports it.
- If the review is not in English, tag it the same way and keep the quote
  in the original language.

Return one JSON object per review, one per line:
{"id": "<recommendationid>", "tags": ["..."], "quotes": {"<tag>": "<quote>"}}

Reviews (id, then text):
1001 | Crashes every time I open the map. Refunded.
1002 | Lovely art but it's over in 90 minutes, not worth 20 dollars.

Sources: Gilardi, Alizadeh and Kubli (arXiv, 2023); Pangakis, Wolken and Fasching (arXiv, 2023); Lin, Bezemer, Zou and Hassan (2019).

How do I turn findings into a patch plan and a reply plan?

Rank the tags by their share of negative reviews in the last two to four weeks, fix the top one that blocks play, and reply to the reviews that named it once the fix ships. One plan for the code, one for the words.

For the patch plan, put blocking problems first. Lin and colleagues found that most bugs named in negative reviews stop players from playing or finishing the game: incompatibility, crashes and freezes. Bugs named in positive reviews were more often performance, audio or visual issues. After blockers, take the design tag with the largest share in the under 10 hour bands.

For the reply plan, follow Valve's advice. It says not to respond to every review and not to argue, and warns that "A developer response will frequently draw more attention than the original statement." Our practice is to reply to negative reviews that report a problem you have fixed, with the build number, and to leave opinions alone.

  1. List the top five complaint tags by share of negative reviews, with the trend over the last three builds.
  2. For each, note the playtime band where it appears and one quote that shows it.
  3. Pick at most two fixes for the next patch: a blocker first, then a design change.
  4. In the patch notes, name the problem in the players' words.
  5. After the patch, reply to the reviews that named it, and count the tag again two weeks later.

Sources: Lin, Bezemer, Zou and Hassan (2019); Steamworks User Reviews documentation.

How to measure the result

Put the weekly review themes next to your crash rate, your retention and your refunds for the same builds, so a fix shows up in all of them and not only in the reviews. Manalens is the analytics and growth engine built for games. It counts errors and crashes as events with the build, so you can chart errors per session by build, and its retention reports show whether players from each week come back.

It does not import Steam reviews, so you tag them yourself as above, and its automatic Steam sync of units sold and returned is in early access.

Sources: Manalens documentation on errors and crashes, retention and Steam.

Frequently asked questions

Can I download all my Steam reviews?

Yes, with Valve's GetAppReviews Web API. Sort by Recent, set languages to all and purchase_type to All, and follow the cursor until a page comes back empty. It returns up to 100 reviews per page and needs no key, but a key associated with your app gets a higher rate limit.

Does Steamworks have a review export button?

Valve's User Reviews documentation describes none. It describes official developer responses, comments and flags for moderation, and it points developers to the GetAppReviews API for reviews in a date range or with other criteria. That API is public, so you can also read the reviews of other games.

How many reviews do I need before the themes mean anything?

Enough that one review does not swing a share. With 30 reviews in a week, one review moves a share by more than 3 points. Lin and colleagues used samples of 96 reviews for a 95% confidence level and a 10% interval, so compare builds or months when your game gets few reviews.

Why do negative reviews come early?

Players who dislike a game stop sooner. In the Steam study by Lin, Bezemer, Zou and Hassan, the median playtime was 6.6 hours before a negative review and 15.5 hours before a positive one. Read negative reviews written under two hours next to refunds, because Steam refunds games played less than two hours.

Should I use AI to tag Steam reviews?

Yes, for speed, but check it. Give the model your codebook, ask for a quote per tag, and send only the review ID and text. Then tag about 100 reviews yourself and compare. A 2023 study of 27 annotation tasks found that model accuracy depended heavily on the task and the data.

Do language-specific scores affect a small game?

Not yet. Since 18 August 2025, Steam shows a score per language only for games with more than 2,000 publicly visible reviews and at least 200 written reviews in at least one language. You can still split your own reviews by language to find a translation or price problem early.

Sources and implementation