Test: How Amazon’s algorithms really work – myth and reality

The way Amazon calculates a sales rank for an ebook often seems mysterious. Does the price play a role in it, is it important to have a title enrolled in KDP Select? Nobody knows for sure, and I think even Amazon’s software engineers are not really sure about the whole process. Last week I called the readers of this blog to help in a large test, and about 90 of them responded.

To be able to actually prove or falsify a thesis, we need an experimental setup. The three primary theories to check were:

  • Price influences sales rank
  • Enrolling in KDP Select influences sales rank
  • The dynamics of sales influence sales rank

To create a concise test environment, I uploaded four new ebooks, generated from different parts of the Selfpublisherbibel. All titles went into the same category. The test was carried out on Amazon.de but I’m sure the results are valid for other marketplaces too. These are the four test titles:

  • Wissen“: KDP  Select yes, 0.99 Euro.
  • Neuigkeiten“: KDP Select no, 2.99 Euro.
  • Tipps“: KDP Select yes, 2.99 Euro.
  • Marketing“: KDP Select no, 0.99 Euro.

The experiment then went as follows:

  • 25. 11., 17:30: all ebooks uploaded.
  • 25.11., 19:45: all ebooks approved.
  • 25.11., 19:45: bought all ebooks once.
  • 26.11., 17:30: all titles have a sales rank.
  • 27.11. 08:40: message to all testers describing what they had to do.
  • 27.11., ca. 17 Uhr: 60 sales/borrows of “Tipps” and “Neuigkeiten”.
  • 28.11., ca. 10 Uhr: 20 sales/borrows of  “Wissen” and “Marketing”.
  • 28.11., ca. 17 Uhr: 20 sales/borrows of  “Wissen” and “Marketing”.
  • 28.11., ca. 17 Uhr: 30 sales/borrows of  “Tipps” and “Neuigkeiten”.
  • 29.11., ca. 10 Uhr: 20 sales/borrows of  “Wissen” and “Marketing”.
  • 30.11., ca. 10 Uhr: 20 sales/borrows of  “Wissen” and “Marketing”.
  • 01.12., ca. 10 Uhr: 10 sales/borrows of  “Wissen” and “Marketing”.

All titles were supposed to be bought resp. borrowed by 90 users. Kindle Unlimited users were asked to read at least upto 10 percent. The chart and the table below are showing the results.

Charts_Amazon_Test

Fact 1: Price does not influence sales rank.

Both 2.99 and 0.99 titles reached similar sales ranks with similar sales (see table).

Fact 2: Enrollment in KDP Select does not influence sales rank.

All titles are treated equal, whether they are enrolled in KDP Select or not.

Fact 3: Organic growth of sales number results in a higher ranking.

Yellow and green curves are based on less sales than red and blue curves but still reach higher rankings in the end.

Fact 4: Sales of (much) more than 24 hours are important for the sales rank.

One day without sales decreases the sales rank of a title in the same way as halving its sales numbers. That means the Amazon algorithms devalue a sale by half every 24 hours (approximately, they exact amount is impossible to calculate). Add to this that the sales numbers are increasing exponentially with high ranks and it gets clear how hard it is for a new title to climb into the charts.

A title with sales rank X gets counted by the algorithm with all its sales from today plus half its sales from yesterday plus a quarter of its sales from the day before yesterday plus… You get the idea. All these halfs and quarters add up to 1 so basically the title gets counted with double its actual sales numbers. That means a title with a lower rank (and less visibility) must reach at least twice the sales numbers of the higher ranked title to actually overtake it if it is brand new. This distance decreases from day to day because the newer, lower title also gets counted with its sales from yesterday etc.

This kind of dynamics explains strange phenomenons seen by many users. “My title 1 only sold x this day on rank y while my title 2 sold more on a lower rank”.

Fact 5: Borrows on KindleUnlimited influence the sales rank immediately.

I couldn’t believe this first. Borrows on KindleUnlimited are only paid when the user reads at least 10 percent. But for the sales rank all borrows count immediately. I saw a first hint of this in the chart above because the numbers for the KU titles are always leading in front of the Non-KU-Titles a bit. Therefore, I tested that with another title. The result was very clear: After four users borrowed it without ever opening, the title climbed from position 190,000 to 4,000. Sales, on the other hand, only count for the sales rank if they are in the sales report (being shown on the dashboard is not enough).

This clearly is a bonus to KindleUnlimited authors. You can reach higher visibility on Amazon without having to wait until a borrower actually reads your book. Even if a borrower never reads your book but only returns it later, you still have a higher sales rank (but you won’t get paid, of course).

Fact 5 also explains why there often is a difference between sales numbers and sales ranks as the author has no possibility to actually see these virtual borrows.

What does that mean for the author?

Be patient. If you are planning promotions, start with the least effective and increase their power in several steps. 20 sales each day for a week get you higher than 150 sales one day and none the other. KindleUnlimited is not as bad a deal as one might think – at least, you get an immediate boost if someone borrows your title.

(This is the shortened English version of this article)

DatumUhrzeitNeuigkeiten (2,99)VerkäufeTipps (2,99)VerkäufeMarketing (0,99)VerkäufeWissen (0,99)Verkäufe
26.11.1417:30287561287551287571287541
27.11.1407:45221662439631439641439621
27.11.1409:30234882452411452421452401
27.11.1412:00262732478461478471478451
27.11.1413:30265302483321483331483311
27.11.1415:30268302488831488841488821
27.11.1417:00278952501901501911501891
27.11.1418:30288732513151513161513141
27.11.1422:00299932525851525861525841
27.11.1423:0091112106116 (7/9)539691539681
28.11.1402:008401260716 (7/9)554981554971
28.11.1403:0010061264516 (7/9)572641572631
28.11.1404:007191252616 (7/9)580281580271
28.11.1406:003984237540 (30/10)594351594341
28.11.1407:0041542397608141608131
28.11.1408:0041542397608141608131
28.11.1409:0041542397608141608131
28.11.1409:224844249162.523162.5221
28.11.1410:0048442491625231625221
28.11.1410:0048442491625231625221
28.11.1410:254844249162.523162.5221
28.11.1411:00500425038829188181
28.11.1411:25500425038.82918.8181
28.11.1411:455004350340 (30/10)8829288183 (2/1)
28.11.1412:00500435038829288183 (2/1)
28.11.1412:26500435038.82928.8183 (2/1)
28.11.1413:00523435259508294953 (2/1)
28.11.1413:39523435259.50829.4953 (2/1)
28.11.1414:005704354110995246993 (2/1)
28.11.1414:405704354110.99524.6993 (2/1)
28.11.1415:005704354110995246993 (2/1)
28.11.1415:395704354110.99524.6993 (2/1)
28.11.1416:00637435875232218683 (2/1)
28.11.1416:306374458742 (32/10)5232318686 (3/3)
28.11.1416:50637445875.23231.8686 (3/3)
28.11.1417:00637445875232318686 (3/3)
28.11.1417:50637445875.23231.8686 (3/3)
28.11.1418:00732446946204317766 (3/3)
28.11.1419:11732446946.20431.7766 (3/3)
28.11.1420:00828447883495318296 (3/3)
28.11.1420:308285278850 (38/12)3495618298 (4/4)
28.11.1420:31828527883.49561.8298 (4/4)
28.11.1422:003755936656 (44/12)2396814688 (4/4)
28.11.1422:00375593662396814688 (4/4)
28.11.1422:03375593662.39681.4688 (4/4)
28.11.1423:59327593182231816278 (4/4)
29.11.1401:00327593182231816278 (4/4)
29.11.1402:00404593962707819438 (4/4)
29.11.1403:00404593962707819438 (4/4)
29.11.1408:004387142666 (53/13)128618114618 (14/4)
29.11.1409:0047971483146818132918 (14/4)
29.11.1409:42479714831.468181.32918 (14/4)
29.11.1410:0047971483146818132918 (14/4)
29.11.1411:00498715051.555181.40418 (14/4)
29.11.1412:0049871505155518140418 (14/4)
29.11.1412:28498715051.555181.40418 (14/4)
29.11.1413:0051971544168118152318 (14/4)
29.11.1413:59519715441.681181.52318 (14/4)
29.11.1414:0056071585150918137018 (14/4)
29.11.1415:00560715851.509181.37018 (14/4)
29.11.1416:00607716416931866418 (14/4)
29.11.1416:456077164166 (53/13)6932566425 (19/6)
29.11.1417:00607716416932566425 (19/6)
29.11.1417:04607716416932566425 (19/6)
29.11.1418:00672717207472571425 (19/6)
29.11.1418:17672717207472571425 (19/6)
29.11.1419:00672717207472571425 (19/6)
29.11.1420:00779717358372579325 (19/6)
29.11.1421:00829717887402575925 (19/6)
29.11.1421:37779717358372579325 (19/6)
29.11.1422:00829717887402575925 (19/6)
29.11.1422:41829717887402575925 (19/6)
29.11.1423:00829717887402575925 (19/6)
29.11.1423:59839718897882580125 (19/6)
30.11.1401:009807110458962591025 (19/6)
30.11.1402:009807110458962591025 (19/6)
30.11.1403:009807110458962591025 (19/6)
30.11.1408:001.357711.4421.240251.42825 (19/6)
30.11.1409:001189711474130925148425 (19/6)
30.11.1409:00118973147466 (53/13)130933148431 (25/6)
30.11.1409:461.189731.4741.309331.48431 (25/6)
30.11.1410:001292731574143833159831 (25/6)
30.11.1411:001292731574143833159831 (25/6)
30.11.1412:001.292731.5471.438331.59831 (25/6)
30.11.1412:461.337731.5811.529331.68631 (25/6)
30.11.1413:001337731581152933168631 (25/6)
30.11.1414:001397731629136133146931 (25/6)
30.11.1415:001.397731.6291.361331.46931 (25/6)
30.11.1415:30139773162967 (53/14)136139146937 (29/8)
30.11.1416:00

Von Matthias Matting

Matthias Matting, geboren 1966, ist Physiker und Journalist und einer der erfolgreichsten deutschen Self-Publishing-Autoren. Er hat über 50 Bücher im Self-Publishing veröffentlicht und ist Autor des offiziellen Amazon-Bestsellers 2011. Für sein Buch “Reise nach Fukushima” erhielt Matthias Matting den 2011 erstmals ausgeschriebenen Buchpreis “derneuebuchpreis.de” in der Kategorie Sachbuch. Matting war als Programmleiter eBook bei der Münchner Verlagsgruppe tätig. Er arbeitet außerdem als Kolumnist für das Nachrichtenmagazin FOCUS und als Autor für SPACE, Federwelt und Telepolis. Schließlich gibt er auch Online-Kurse sowie Seminare an der Akademie der Bayerischen Presse.

26 Kommentare

  1. Interesting article. If the Kindle Unlimited borrows are influencing sales rank even if the title is unread, I can see that getting abused – ie "pay for borrows" services.
    Also, I wonder if an unread borrow still influences the recommendations? That could be another vector of abuse – borrows of your own and popular books in your niche to get on the "people also bought" lists.

  2. I don’t know if you are setting up future test ideas, but I recently had someone say that the number of reviews (and the type… verified purchase vs not-verified) factor into rankings.

    I highly doubt it, but that is what testing is for, right?

    Would love to see that test

    1. Hi, thanks, I doubt that as well. Would that be the pure number of reviews or their rating too? If it’s just the number, this could be easily tested with two books, one staying unrated and the other with ten 1-star-ratings. What I wouldn’t try for obvious reasons are the same number of 5-star-ratings…

    2. This would indeed be very interesting to test.
      Back in 2012 during the first 6 months after publishing my book, I’m pretty sure I saw this trend. I noticed several times that when my rank was stable and my daily sales rate fairly constant, a new 5-star review would correlate with an increase in rank within 2 hours of it being posted.
      (Yes, OK, I admit that back then I was following my rank and sales much more closely than was probably healthy, but it was all much more exciting back then 🙂

      I suppose however that what might have been happening was an increase in sales caused by the new review, but those sales not being reported as quickly as the change in rank. So instead of the causal chain being review -> rank ->sales, it may in fact have been review -> sales -> rank

      Would be great to find out though.

  3. Matthias Matting I understand what you're saying but I agree with Michelle Fox in that, to participate in KindleUnlimited a book needs to be in Select. As long as authors understand this point, your summary is excellent info. More to your point, if a book is in Select but goes unborrowed, then Select does not help the book. When our books were in Select were borrowed, we saw an increase in sales ranking. Outside of Select, the ranking dropped. Thanks for going to the trouble to test all this. Valuable information.

  4. Matthias Matting A lot of people had been reporting that Borrows were weighed heavier than sales in terms of weight, giving books in KU an unfair "finger on the scales" advantage. This was based on the fact that books with a lot of borrows were getting higher ranks than books with similar number of sales.

    I think most the people who doubted that this was the case assumed it was due to unread borrows counting towards rank. It's good to have some hard data to back that up.

    In general it's great to have this data to help combat all the missinformation about rank and select and borrows.

    Great stuff, thank you again.

  5. Enrollment in KU by itself definitely does not improve rank or sales. That's what the A/B test is about. A better visibility through KU is another matter though it is impossible to test in this scenario (with "dummy" titles and planned sales).

  6. How is rank influenced by the level the "author rank" is at? I'd like to see if that makes a difference. Fo rthis, you would need a few authors to participate. In the meantime, I'm off to ask people to use their KU borrows on one of my books.

  7. I feel like fact 5 sort of contradicts fact 2. Also, you're unable to track how much visibility KU is given vs. a non KU book which influences sales and rank. I can definitely say KU bumps up sales and rank. Don't know why your testing didn't see a correlation. Most every author who has tried KU sees a bump.

  8. I was going to ask the same question. They've explicitly stated (the obvious) that you'll only be credited with a borrow once per borrowing account, but that's for revenue.

    I suspect the answer is that it will only count once for rank as well, otherwise those pamphlets would be ranked even higher. It would be fairly easy to set up macros to automate borrow/read/return.

    So my guess would be that they're only achieving the ranks they do through a large organized networks of borrowers.

  9. Very interesting. Fact 5 has some troubling aspects to it. A KU subscriber can download and return a book as much as they wish. If Fact 5 holds true for each download then KU is rip for scamming, which would explain a lot of what we are seeing with 20-page books climbing to the top of the charts despite nothing but 1-star reviews. Do you see any evidence of a repeat borrow counting the same as a first borrow?

Kommentare sind geschlossen.