Notes from the Field

Occasional writing on metadata, discoverability, and the economics of publishing online.

Sent a few times a year. Read on site, or have it sent to you.

Earlier pieces in this archive date from 2015 to 2018, when Kadaxis was active in industry forums and conferences. New writing resumes in 2026.

Diagram showing launch metadata forming a substitute neighborhood that supplies warm-start ranking signal. RESEARCH

Amazon's warm-start research makes launch metadata a ranking decision

Amazon has described a production search-ranking system that uses substitute products to give new products a warm start before they have sales history of their own. For publishers, the implication is direct: cover, title, description, comp-title language, and category placement help determine the behavioral neighborhood a new book enters at launch. That neighborhood can shape early visibility, sales velocity, and the book's ability to build its own ranking signals.

Vertical discoverability funnel showing indexing, eligibility, rankability, and convertibility narrowing toward a sale. METHOD

The discoverability funnel

Amazon accounts for roughly half of US print book sales and two-thirds of US ebook sales, and most US consumers now start their online product searches on Amazon rather than Google. The path from a book's metadata to a reader's purchase runs through a funnel with four sequential filters. Most candidate books drop out at one of them.

Four-set overlap diagram showing searched, relevant, additive, and rankable keyword tests meeting in a small working-keyword space. METHOD

The narrow space where keywords actually work

For a book keyword to do anything useful, it has to clear four independent tests: someone has to be searching for it, Amazon has to consider the book relevant to it, it has to add signal the rest of the metadata does not already supply, and the book has to be able to rank for it. Most keyword candidates fail at least one.

Diagram contrasting duplicated metadata signals with connective keywords that add missing relevance. METHOD

Not every keyword is doing work

In our recent analysis, only about eight percent of publisher-assigned book keywords produced a visible match in retailer search. The other ninety-two percent are not wrong; they are mostly repeating signals the rest of the metadata already supplies.

Diagram showing semantic eligibility as Gate 1 and ranking likelihood as Gate 2 in search visibility. METHOD

Relevance is not the same thing as ranking

Most keyword strategies treat search visibility as a single problem. It is two problems, in sequence, and confusing them is why so much keyword work fails to move performance.

Infographic showing keyword search hits before and after the 250-character mark. FIELD NOTE

Amazon Keywords Don't Stop at 250 Characters. Publishers Do.

A direct look at whether Amazon ignores ONIX keyword phrases after 250 characters, and what publishers should actually do with the field.

Diagram showing catalogue language and reader review language moving toward retailer search. COMMENTARY

What reader reviews know that catalogues don't

Catalogues describe books in the trade's language. Readers search in their own. The gap between those two registers is where most of the lost sales live.

Diagram contrasting book-text surface terms with reader-search keyword language. METHOD

Why book keywords don't live inside the book

Most keyword tools start with the text of the book. The text matters, but reader search often lives in a different vocabulary: need, outcome, comparison, emotional texture, and intent.

Diagram showing manuscript text as one input among missing commercial variables. COMMENTARY

What we learned trying to predict bestsellers

A year of work on bestseller prediction from manuscript text produced a working system and a clear lesson: the model was answering the wrong question, and the right question requires variables the manuscript cannot supply.

Diagram contrasting text-based comparable titles with review-based reading-experience comps. METHOD

Why comparable titles fail when derived from book text

Most automated comp-title tools operate on the text of the book. The comparisons readers and the trade actually make operate on something the text cannot fully capture: the reading experience.

Diagram showing four format editions feeding keyword signals into one Amazon work-level association layer. RESEARCH

Different keywords on different formats of the same book

A common piece of advice is to assign different keyword sets to the hardcover, paperback, ebook, and audiobook editions of the same title to expand the work's reachable search surface. We tested whether this works at scale. The phenomenon exists in specific cases, but our data does not support recommending it as default practice.

RESEARCH

How Top Publishers Use Keywords

A study of 150,000 publishers, conducted with Bowker and Firebrand, on how the industry was actually using metadata keywords and where the practice was still falling short.

METHOD

Why a 500-Character Keyword Limit Is Costing You Book Sales

The widely accepted ceiling for keyword metadata is wrong, and the cost compounds across a list.

FOUNDATIONS

How Amazon Search Really Works

A working explanation of the algorithm that determines which books are surfaced and which are not.

METHOD

Measuring Keyword Effectiveness on Amazon

A practical framework for assessing whether metadata keywords are working once they are deployed.

RESEARCH

How Do Keywords Impact Sales?

Why search visibility, sales history, and metadata quality reinforce each other over time.

RESEARCH

Who Uses the Keywords in Metadata?

A look at how publishers and retailers were using ONIX keywords, and why Amazon was the outlier.

COMMENTARY

BISG 2017 Annual Meeting - Rights, Metadata and Marketing Panel

A panel response on rights, metadata, marketing, and the infrastructure around book discovery.

FOUNDATIONS

What are off-page keywords?

An evergreen definition of off-page keywords and how retailer search engines use them.

COMMENTARY

Machine Learning and Bestseller Prediction: More Than Words Can Say

Why book sales cannot be predicted by text alone, and what machine reading can and cannot explain.

METHOD

How A Keyword Sells A Book On Amazon

A simple model for how one metadata keyword can influence retailer search and sales.