“All reviews on Yelp are fake.”
Someone told me this once awhile back and it is a statement that has stuck with me. Not because I believe that it is true, but because it always makes me wonder about how many reviews on the various websites out there are legit.
But how can you tell? In our case, how can we tell when an App Store user is posting fake reviews? We will get to that in a minute…
Before we do, let’s put this into perspective and discuss why it even matters, if you don’t know already. The shift from desktop to mobile computing has progressed rapidly over the years and will only continue to accelerate as as we move forward.
The dark side of this mobile boom is that some developers are going to try to cheat the system and get their apps on the top charts, assuring themselves almost instant publicity, income and residual downloads. This is nothing new of course. In any market, especially emerging ones, there will always be opportunities to take advantage of inefficiencies.
Apple has taken steps to combat this, but they have to strike a balance between usability and being too restrictive. With over 900,000 apps on the App Store and growing, they also need to be sure not to upset this huge ecosystem.
One thing Apple has done is to place more importance on user engagement when determining app rankings. While this is a step in the right direction, more changes will need to be made in order to ensure that the best apps really are on the Top Charts and showing up in search.
In this post, we will pull the covers back on what a fake reviewer looks like, compare this to a real reviewer, and examine how this information could be used to prevent these shady practices from happening. We will also get into how downloads and reviews are faked, if you are not familiar with the process.
As you look at this data, just keep in mind that these numbers are for the US App Store only.
How To Fake Downloads And Reviews
Scammers setup computers that maintain a ton of bogus user accounts (also known as bots) to download an app multiple times so that the app seems popular and will start appearing on the Top Charts and will have better visibility in App Store search. These computers can use various methods to mask what they are doing.
Since download data is not public information, looking at reviews can be a decent substitute for detecting fraudulent download behavior. After all, if a scammer goes through all the trouble to setup an bogus account, why not leave a 5-star review while they are at it? Reviews also help apps get more downloads by making them look more credible.
Once the dubious accounts are setup, article spinning software can spit out hundreds of variations of one review, making it difficult for spam detection programs to figure out that they were written by the same person. Another way to make fake reviews look more human is to actually pay someone to post reviews with these accounts.
If a shady developer (or PR firm) does not want to go through the trouble of setting this up themselves, the easiest and fastest way to get a lot of 5-star reviews is to simply buy them. There are numerous sites out there that will allow you purchase reviews. Here is a small sample of the options available on Fiverr. Some of these offerings appear to offer honest reviews, but the fact that they are being paid to do it, makes them very questionable.
Some developers even go so far as to post job listings to get people to review their apps. Here is an example where the bidding for the project got up to $55 for 50 reviews. The developer posting the job was careful to make sure that the person who took the job posted exactly forty 5-star reviews and ten 4-star reviews.
What We Can Learn From The Most Prolific Reviewers
Is it really possible to tell if a particular user ID is posting fake reviews? That is the question that we asked ourselves and we decided to dive into our plethora of data to see what we could uncover. We took a look at the top 25 users with the most reviews to see if we could find any patterns that would help us answer this question.
We want to use the users with the most reviews because it gives us more data points from which to see trends and draw conclusions. To protect the identities of the actual users, we will simply refer to them by their review count rank, with #1 being the user with the most reviews.
There is an average of 765 reviews per user on this list. Here is a graph of these users, showing their average rating and number of reviews posted.
If a user account is fake, the theory should hold that they will have a high average rating and there should be some discernible pattern that shows that the reviews were posted by an automated program or people hired just to write reviews.
However, instead of starting off by looking for fake accounts, let’s instead, first look at the user with the lowest average rating. This should be the best way to profile what a legitimate user looks like and having this information will help us to better identify a fake user.
What A Real User Looks Like
Since leaving low ratings does not benefit the apps that are reviewed, reviewers with a low average rating should be giving their honest opinion and not trying to artificially boost the apps that they rate. When we take this criteria into account and examine other characteristics of the reviews of such a user, the picture of a real person who leaves honest opinions starts to become clearer.
Let’s examine the username thespaciousmind (571 reviews), since this person had the lowest overall average rating of 1.59. He or she obviously loves to tell developers how bad their apps are.
Here is an example review:
“Lame controls. The game is designed to force you to buy upgrades after the 1st 5 levels. Or, just replay the first 5 levels 1000000 times to earn those credits. So lame.”
People are certainly entitled to their opinion and negative feedback can actually be useful to a developer because it shows how the app can be improved. This person might also only leave bad reviews and doesn’t leave a review when an app is good. Whatever the case may be, this very low average rating should be a good indicator of review authenticity, but let’s see what the rest of the data says.
Review Character Count
Negativity aside, there are other clues that show that these reviews are real. We noticed that the reviews are of widely varying lengths. They range from 3 characters to 2559 characters and are pretty evenly distributed. This demonstrates that the reviews were not created according to a formula and is what we would expect from a person who is simply passionate about apps.
Review Sentence Count
Closely related to the character count is the number of sentences that are written in each review. In the context of app reviews, we defined a sentence as any word or strings of words that end in a punctuation mark, regardless of how much sense it makes.
The most frequent number of sentences in this person’s reviews is two, but there are also as many as 51, with varying lengths in between. Logically, this is something that we would expect because a real person would have more to say about certain apps and less to say about others. Get them on a rant and 51 sentences is completely reasonable.
Spelling And Grammar
There are also numerous spelling errors in these reviews. This fact does not guarantee thatthespaciousmind is a real person because the reviews could just be coming from a comment farm where the workers are not fluent English speakers.
Here is an example (errors in bold):
“Nice **looling** card game. Take time to learn rules. Unfortunately **theres** no stats, no records, no **scoretable**. So not much reason to play it. Cross my fingers for an update. Until then **dont** bother. Lots of other/better card games that HAVE stats, records, etc.”
However, as we look through these reviews, it is obvious that this person speaks English pretty well but just does not proofread before submitting. In addition, the reviews are in different formats. There are multiple words that are all in upper case and some of the reviews are in paragraph form, while others are in list format. This finding goes against several posts that we have read that say that review misspellings are a sure sign of a fake review.
The reviews also tell a story of this user’s app usage patterns and personal life. This person loves games and plays a wide variety of them. He or she also gives very specific examples of what is good or bad about these apps and makes occasional references to his/her child.
You can actually see all of the apps that any individual user has reviewed by finding one of their reviews on iTunes. When you click on the username, you will see all of the apps that the user has reviewed. It is the easiest way for anyone to start examining individual user accounts for review patterns. If you have a little time to burn one day and this interests you, take a look at a few user and see what you find.
Finally, it is important to note that a large number of reviews doesn’t necessarily mean that a user is not a real person. You may assume that nobody has the time to leave hundreds of reviews. But asthespaciousmind has shown, there are people who are very passionate about their apps and do take the time to leave feedback for app developers.
What A Fake Reviewer Looks Like
Now let’s look at users that could be fakes. We say “could be” because there is no way to say for sure without actually personally knowing the owner of the account, but as we will show, it is very likely, given the data we have.
To get some insight into what a fake user looks like we will examine the top two most suspicious accounts,MrCooper5 and @LoveFunApps <3<3<3. We will get into why these two are the most likely to be fake in a minute, but the initial thing that first drew us to @LoveFunApps <3<3<3 was the username.
It looks suspicious because of all of the “3” and “<” signs in it. It was the only username in the top 25 that used symbols excessively. Yes, it could just be a teenage girl…but as we dig deeper, it is more likely to be a crooked account that is using the symbols to attract attention to the reviews this user leaves.
The biggest tip off that a user is fake is a high average rating. Both of these users have an average rating of 4.99. When was the last time that you left reviews on Yelp, Amazon or Netflix and gave everything five stars?
OK, that might be the case if you only leave a handful of reviews. But it is highly unlikely that almost all your ratings would be five stars, even if you left as few as 30 or 40 reviews.
Review Character Count
Next, when we look a the character count of the reviews from @LoveFunApps <3<3<3, there is an inordinately high number of reviews that are exactly the same. Of the total 686 reviews left, 205 of them, or about 30%, say: “Fun!”. I guess they gave themselves that username for a reason.
This can be seen in the graph below, as pointed out by the red arrow. There are a few reviews that have higher character counts, but nowhere near the 2,000+ that we saw in the real reviewer profile.
The character count of the reviews from MrCooper5 shows us a different type of pattern, but one that is suspicious nonetheless. In this case, the sentences in these reviews were written in a way that gave us the impression that the writer had never actually used the app.
As we kept on read through the reviews, it almost felt like there was a minimum number of characters that the review writer had to fulfill, similar to how an article spinning software program would operate. This can be seen on the graph below, where the lowest number of characters in a review was 15, instead of the much lower number seen the real user profile.
Review Sentence Count
Now let’s turn to the number of sentences per review. In the real user profile, we learned that a real user should leave a wide range of results when it comes to sentence count.
The first graph is from the reviews of @LoveFunApps <3<3<3. As you might expect, the distribution of sentence counts are nowhere near the variation that we saw with thespaciousmind, in spite of@LoveFunApps <3<3<3 having 100+ reviews more.
MrCooper5 also has a noticeable formula when it comes writing reviews. It is almost as if there was a manual that told the writer to write about 5 sentences per review. Like @LoveFunApps <3<3<3, the maximum number of sentences also tops out at 10 sentences. Coincidence?
Spelling And Grammar
The spelling and grammar in these reviews would make any English teacher proud. When we scrolled through the reviews in our text editor, there were no spelling mistakes (AKA red underlines), with the exception of an occasional “sooooo,” as in: “OMG, this app is sooooo fun!”
The grammar was equally clean, with excessive punctuation being the only real violations. But then again, if you are writing “Fun!” most of the time, it is kind of hard to mess that up.
Finally, as we read through these reviews, there were no real personal experiences or personality traits that created a common thread that led us to believe that this was a real person writing these reviews. A real user (like thespaciousmind) will usually share some sort of personal experience or reference family or pet peeves. This was not the case in these reviews and they had a more distant and robotic quality to them.
Analyzing Other Top 25 Users
As we looked through the other reviewer profiles, the users who have a very high average ratings, are Bad Watch (4.96), Yinismss (4.99) and jaryre (4.84). They all share very similar characteristics with MrCooper5and @LoveFunApps <3<3<3.
The other users in the top 25 that have average ratings between 4.80 and 1.59 are also interesting. In line with what we have outlined above, the higher the average rating, the more the account looked like a fake and the lower the average rating, the more likely it was a real person. But there were a couple of users where it was a little tough to tell and deeper analysis would have to be done.
One Hit Wonders
Another possible way to fake reviews is to setup accounts that only leave one review. This is much more time consuming, but since there is no trend, that makes it almost impossible to tell if a user is fake or not.
When a user has only one review, this does not necessarily indicate that it is a fake account. However, it might mean that they are less credible because they may have been asked to review a friend or employer’s app, when they could otherwise care less about leaving a review.
Out of all the reviewers on the US App Store, 9,074,360 users have only left one review, accounting for 57.7% of all users who have left a review. When we looked at the data, only 1.8% left 10 reviews or more. A more detailed inspection of the difference between users who leave one review and users who leave two or more may yield further clues into how to profile a fake account, or at least an account that should not be as influential.
The Top 250
If you want to take a look at more than just the top 25 reviewers, here is an interactive graph of the top 250 reviewer accounts. You can see their review counts, average rating and average characters per review. What are some interesting patterns that you see?
Can the characteristics outlined in this post be used to profile and stop review spam? We believe that it can certainly help, but a lot more analysis and testing would have to be done to create a solid spammer profile. Posting patterns, correlations between apps reviewed and data that only Apple has, like IP addresses that the reviews were posted from would have to be thrown into the mix.
But the bigger question is: Does stopping fake downloads and reviews benefit Apple significantly enough to make it worthwhile for them to figure it out and implement a remedy? After all, as long as an app works, developers are happy because they have an app on the App Store, Apple makes money through fees and they get to add another app to their total app count.
We believe that they do have to figure it out to ensure the long term success of the App Store. As more and more apps hit the App Store every day, it will be more important that quality apps rise to the top in both search and the Top Charts. But as we have shown, it would be a complex task and it would probably be better to lean towards being too liberal than too strict.
One thing that might help is if Apple implemented a minimum character count for each review like Yelpdoes, in order to leave a review and rating. This obviously would not stop the more savvy scammers, but it would start to weed out the less complex ones because it would make the process more time consuming.
This step would also make the reviews more beneficial to potential users and possibly eliminate a large proportion of people who leave poor reviews out of haste, which would be better for developers. In addition, since we have shown that the vast majority of the reviewers out there leave less than 10 reviews, that could be a threshold for the rating actually counting towards the average rating for the app.
The review would still be visible, but the average rating of the app would not be affected unless that user has already reviewed 10 apps. A similar tactic has been used on forums for a long time, where users who do not have a minimum number of posts cannot access certain forum benefits.
There are a lot of articles out there that give opinions on what a fake App Store user profile looks like, based on some cursory personal experience, but as far as we know, this is the first post to put some actual statistics behind them. We hope that we have given you some insight into how fake reviewers operate, how to spot them and what may be able to be done about it.
To learn more about how to do a deep analysis on your reviews to make your app better and find more keywords, read our post on Review Analysis.