I read eight Amazon algorithm guides. Here are two top factors, and the short list Amazon actually publishes

The 30-second read

When I decided to try writing romance seriously, I went looking for the formula. I'm a data person, so I assumed there was one. I read eight guides, six of them in full, which was plenty, and reading about Amazon felt like progress. It wasn't. None of it put a single word on the page.

Here's what I found. An Amazon community manager on its seller forum says Amazon doesn't publish its search rules. The guides rebuild them from the outside, with different data, for different products. Four of the six I compared lean on sales velocity, three of them with a weight of 35 to 40%. Two put conversion first. None showed me their data. What Amazon does tell authors is short: keywords, categories, rank, and how long a change takes to show up. It was enough to start designing a test model of my own.

In this piece - The algorithm, in one line - What Amazon actually tells authors - How long a change takes to land - What "we modeled it" actually means - Six guides, side by side - The test model I'm building instead - Where this is soft

The algorithm, in one line

When a reader searches Amazon, a computer decides which books to show and in what order. That's "the algorithm". Amazon doesn't share the rules.

On Amazon's own seller forum, a community manager put it about as plainly as possible: "Amazon does not disclose information or release official documentation around search/discovery algorithms". That thread is about product search in general, not books, and I haven't found anything Amazon publishes for authors that says books work differently.

What Amazon actually tells authors

The quotations in this section are Amazon's own words, from its own help pages, read on September 28, 2026. The lines around them are mine.

What goes into search. "Along with factors like sales history and Amazon Best Sellers Rank, relevant keywords can boost your placement in search results on the Amazon Store". So Amazon names the factors. It doesn't say how much each one counts.

Keywords are mine to set. "Use up to seven keywords or short phrases." "Think like a reader." And my favorite, because it's an experiment hiding in a help page: "Before publishing, search using keywords you're considering on Amazon. If you get irrelevant or unsatisfying results, make some changes".

Categories are mine to set. "As an author, you can select 3 categories when creating a book in KDP." Amazon describes categories as "like digital shelves".

Sales rank sits under Product Details on a book's page. Amazon says it's "based on customer activity of your book relative to the activity of other books", that "Rankings reflect recent and historical activity, with recent activity weighted more heavily", and that "Rankings are relative, so your Sales Rank can change even when your book's level of activity stays the same". Each format "has its own independent Amazon Best Sellers Rank".

That recency weighting is probably why a launch-day burst looks dramatic and then fades. Amazon adds that "One sale of a very popular book may not influence its rank much at all, but one sale of a lower volume book may significantly improve that book's rank". For a new or quiet book, that's both comforting and a warning: one sale can move rank a long way, which means one jump tells me very little.

Star ratings aren't a plain average. Amazon says it uses "advanced models to calculate star ratings, not just a simple average", that "Newer reviews carry more weight than older ones", and that "reviews from customers who make verified purchases are weighted higher".

That's what I could find in Amazon's own words. No formula, but enough to plan a test.

How long a change takes to land

This is the detail that changed how I think about testing. Amazon publishes a timelines page, and it's more specific than the sales rank page. For a published eBook, updates to keywords, title and description appear within 72 hours, category and cover updates within 72 hours, and a list price change within 24 hours. Sales rank "is updated hourly", but can take 24 to 48 hours to show after the first sale, and after a category change, rank updates "can take up to 1 week".

Which means if I change a keyword on Monday and check on Wednesday, I may be judging a change that hasn't gone live yet. I am, I'll admit, exactly the person who would check on Wednesday.

What "we modeled it" actually means

Because Amazon keeps its rules private, people build models.

A model is an attempt to work out the rules from the outside by watching what happens. It's a bit like guessing a recipe from the taste. I might get close, but I'd never know for sure. So when a guide says something like "after analyzing millions of data points, patterns emerge", that's real work and often a fair attempt. It's also fair for them to keep their data private. It might belong to clients, or it might be the thing they sell. That's exactly what makes one model hard to check, and two hard to choose between.

These are the questions I ended up asking every guide.

How much data is it, really? A data point is one recorded fact. I opened the Kindle page for one romance, Emily Henry's Funny Story, and counted 26 separate facts before I reached a single customer review: the title, author and format, four prices, the star rating, the number of ratings, twelve product details, four best seller ranks and the description. Record all 26 for 500 books every day for a year:

500 books × 365 days × 26 facts = 4,745,000 data points

It sounds enormous. It's still 500 books.

Did it only look at the books that survived? A model that looks at Amazon on one day only sees the books still there, usually the ones doing well. A model that tracks the same books for months can see the ones that sank, but only if it started watching before they did. One rank tracker I looked at takes a daily snapshot per product and builds 30-, 90- and 180-day histories, which is the second kind. Either way, I want to know: were the books that did the same things and failed ever in the data? Data people call this survivorship bias. It doesn't make a model useless. It limits what the model can say about why a book worked.

Which books, and when? Romance, or every genre? Books, or kitchen gadgets? Of the six guides in my table, half are written for sellers of physical products. Nova's worked example is a set of mixing bowls. A model built on mixing bowls may not describe my book, and a model built on last year describes last year.

What was it measuring? Search order, sales rank and recommendations are different things. One author-facing guide puts sales velocity at 35% and keyword relevance at 15% in its main table, then further down the same page puts keyword relevance at 60% for search ranking specifically. I don't think either number is wrong. They're answers to two different questions.

Was it checked on new books? Say a guide studies 500 romances and notices the best sellers all had long descriptions. The fair test is to take a fresh batch it never looked at and see whether long descriptions still pick out the best sellers. A model only ever checked against the books it was built from will look better than it deserves to, because it was shaped to fit them.

Cause, or coincidence? Say a model finds that books with more reviews rank higher. From outside, I can't tell whether reviews lift rank, or higher rank brings in readers who then leave reviews. Most outside models can only see things that happen together.

Can anyone check it? If the data and method aren't shown, no. The model might still be right, but nobody outside can tell.

Six guides, side by side

Here is what the guides I read in full in September 2026 say about the single factor that matters most, and how much it counts.

Lined up, four of the six lean on sales velocity, three of them at 35 to 40%, and two put conversion first. That looks like agreement until I check what each one measured. The six numbers aren't measuring the same thing. Some are about products, some about books, and some aren't weights at all, so they can't be added up into one formula for ranking a book. So I can't tell which guide is right. I can only see that none of them shows enough to check.

If a publisher has a method or data I missed, I'll add it.

The test model I'm building instead

It's a small plan, and a bit boring. That's fine for a first try. Here's how each round is meant to work.

• I pick one thing. A keyword, a category, the cover, the description or the price. Keywords and categories are the ones Amazon explains. The others will be mine to try.
• I write it down. The date, what changed, the format and marketplace, and where sales stood before. An eBook on Amazon.com is a different test from a paperback, or another storefront.
• I wait for it to land before counting. Up to 72 hours for keywords, categories, description and cover, up to 24 for price, and up to a week for rank after a category change.
• I judge by sales reports, not rank. Orders, and pages read for books in Kindle Unlimited. Amazon says rank can give "general insight" into marketing, but that it "is not an accurate way to track your book's activity", and for a book that isn't selling many copies, one sale can move rank a long way.
• Then I change the next thing. One at a time, so there's a fair chance of telling which change did what, with a note of anything else that happened that week: a newsletter, a promo, a holiday.

It won't turn any book into a hit, and it isn't meant to. After a few rounds it would tell me something no guide can: how a book of mine responds. Mostly, I want my Amazon reading to stop eating my writing time.

Where this is soft

• I can't say how much each search factor counts, because Amazon doesn't say.
• The 500-book example is my own arithmetic to show how data points add up, not any guide's actual data.
• The 26 facts come from one Kindle page I counted on September 28, 2026, not counting links like "Accessibility: Learn more". Other books' pages may show a few more or fewer.
• Amazon's timelines page calls its figures "estimates only", so I treat the waiting times as rough guides, not promises.
• The table covers six guides I read in full in September 2026. Two more guides I read that month, and a LinkedIn post from HMD, are listed in the receipts.
• Where the table says "no method" or "no data", I mean none on the page I read. A longer page or a different version may have more.
• Every quote from Amazon here is copied exactly from Amazon's own pages and saved as a dated screenshot.

Corrections to this piece are logged on the Corrections page. Want the next one? Get updates.