Methodology

Our goal here

Our goal with the taxonomy is simple.

To reveal and categorize the underlying linguistic patterns of persuasion.

Persuasion is a misunderstood concept. Most equate it with manipulation or convincing.

But when we look at the actual root of the word, we discover something else entirely.

persuasion

/pərˈswāZH(ə)n/

The word persuasion comes from the Latin persuadere, meaning “to advise, urge, or convince”.1

Word Origins

• Latin Roots: It combines per- (“thoroughly” or “strongly”) with suadere (“to urge” or “to persuade”).
• Deep Ancestry: The root suadere traces back to the Proto-Indo-European root *swād-, meaning “sweet” or “pleasant”.
• Literal Meaning: To persuade originally meant to make something “sweet” or agreeable to someone else so they would accept it.1

With that definition in mind, persuasion is a noble art, as well as a science.

A painter must know which colors to mix to create the hue they need. A chemist must know which molecules create which compounds. A musician must know which notes are available.

And marketers must know which techniques to combine to create the sweet message they intend to perform the role of selling their thing.

The problem is, no one ever took the time to give the marketer awareness of all the tools at their disposal.

Worse, every marketer who publishes a book or a course comes up with their own system of stringing together words to create a persuasive message. After reading 10 to 20 of these books, there's a hodgepodge of speculation, theories, case studies, frameworks and ideas all mixed up into a strange soup. Then it is up to the marketer's intuition to decide which techniques to use for any particular project.

While AI may never be able to write as well as a professional human, it can sure help with categorizing the field to provide better ideas and a deeper understanding of the art and craft.

And it also significantly improves the output of an LLM when it has structured data to work with for a systematic, scientific process of conversion optimization.

We did it different

In this marvelous age of AI, it's finally possible to conduct a linguistic study of persuasive writing and dissect it down to the individual building blocks of persuasion.

The genetic code of marketing messages.

And to be real, it had to be drawn from reality. Not hallucinated.

So we started on this study of all of marketing itself. Every ad we could find. Homepages, emails, radio commercials, infomercials, Facebook ads, VSL scripts, direct mail letters, newspaper ads… all of it.

We pulled in ads reaching back to 1890 up to modern day podcast ads.

And we pulled them apart, over months, looking for the individual techniques that get used and re-used.

Our analysis of techniques had four criteria.

1. It's a mechanism. A mechanism that makes a specific move inside of a reader's mind.
2. It is portable. The same mechanism has to occur across unrelated industries, and across time as well. Something which only appears in supplement marketing is a habit of that industry, not a mechanism of persuasion as a whole.
3. It is distinct. Before anything is added, it is checked against the entries closest to it, by name. If an existing entry captures the same reader-mechanism, the candidate merges into it. The discriminator between two entries has to be a real difference in the mechanism, not a new industry or a fresher phrase.
4. It is demonstrable. We needed real, specific examples where the quoted words perform the move. A technique that cannot be shown operating in the wild does not get published.

The default was no

Through our "mining" process we generated thousands of proposed techniques.

The common outcome for a proposed technique is a merge into an existing entry; the next most common is retirement, for candidates that turn out to be features, formats, or generic properties of copy rather than mechanisms. Coining a genuinely new entry is the rare outcome.

The goal was not to have a huge taxonomy with weak examples, but a smaller registry of techniques with enough evidence to say "Okay, yes, this is a real thing."

Because that is how the scientific process works. By hypothesis and elimination.

And so we whittled and carved down the data set, until we found the actual real techniques that make up the building blocks of persuasion.

What the mining looked like

The marketing world loves swipe files. Every good marketer has one, and regularly looks through it to find new ideas to use in a campaign.

We treated the entire internet as our swipe file.

The examples come from a mined corpus of real persuasion, collected continuously: public ad libraries, archived web copy going back to brands' founding eras, live brand surfaces, video sales material, and email programs.

Excerpts are captured verbatim, attributed, and tagged against controlled vocabularies for industry, channel, era, and voice.

All of this went through a first pass with an LLM to extract potential techniques.

These potential techniques then went through an evaluation round, where three different LLM "judges" voted on whether the technique was a real technique or should be merged or retired.

We captured verbatim examples, and set a threshold floor of four distinct examples across multiple industries for a technique to be verified as publishable.

Finally, each technique, and every example of it, was read by a real human (yes, every single one) and judged as worthy of inclusion or not.

This process took many months, and without LLM assistance would not have been possible.

How examples are chosen

An example has to pass two binary gates before it is even scored.

The move must be executed in the quoted words. Strip away the surrounding context and the technique must still be visible in the excerpt itself. If the move only exists in the description around the quote, the example fails.

The example must be a clean paradigm case. The page's technique has to be the dominant, unmistakable move in the excerpt. Other techniques can be present and are cross-tagged, but if a careful reader cannot decide which move dominates, the example fails. Ambiguity is a failing grade, not a judgment call.

Survivors are scored on a fixed rubric: how textbook the case is, whether it is legible on its own, verbatim fidelity to the source, what it adds to the diversity of the page's example set, fair-use discipline, and the quality of the analytical commentary attached to it. A page publishes only when enough examples clear the bar, at least four; entries that cannot field four clean examples wait in a research queue rather than publishing thin.

Fair use and attribution

The catalog quotes other people's copy to analyze it, which makes the excerpt discipline load-bearing.

Headlines and short lines are quoted verbatim. Body copy is described, not transcribed. When a longer work must be referenced, the shortest excerpt that demonstrates the mechanism is used, and deliberately not the most expressive passage available. No page draws more than one example from the same source. Every example carries a note stating what it is cited as evidence of, and every page points readers toward the source material instead of substituting for it.

Attribution is never invented. Where the brand, agency, or writer is documented, they are named; where the source is real but the authorship is not verifiable, the example says so plainly. When a technique is associated with a named practitioner, the practitioner is credited and cited. That is nominative use, not affiliation.

If you hold rights in something quoted here and believe a line has been crossed, the use policy describes exactly how to reach us and what happens next.

Classification

Every entry has a primary role: which of the nine reader-questions the technique most directly answers. Some entries carry a secondary role. Within roles, entries are grouped into families of related mechanisms, and the classification question is always asked in that order: which role, which family, which entry.

Sub-techniques are variants of a parent with the same mechanism in a different setting. A hard pattern interrupt and a soft pattern interrupt are sub-techniques of one entry. A pattern interrupt and an absurdist escalation are separate entries, even though both serve DISRUPT, because they pull different cognitive levers.

Verification and maintenance

The registry is audited, on a schedule, against its own corpus.

Referential integrity. Every technique tag on every example in the corpus must resolve to a registered entry. The audit that enforces this runs as a standing rule, and tagging always checks the existing registry before proposing anything new.

Reconciliation. On a recurring cycle, the registry is reconciled: near-synonym entries are merged, instance-named entries are generalized, and entries that turn out not to be mechanisms are retired. The classification axes themselves (industry, channel, era) are held to fixed canonical sets so the corpus stays queryable rather than fragmenting.

Pattern analysis. The same cycle looks at the corpus statistically: which techniques recur together, which are genuinely cross-industry, which candidate entries have earned promotion from one observed instance to a registered mechanism. The registry grows by evidence, not by brainstorm.

The role of AI

AI is used heavily in building this catalog, and it would be dishonest to pretend otherwise. It is used narrowly, and the narrowness is the method.

Machine agents do the high-volume work: pulling from ad libraries and archives, extracting verbatim excerpts, and proposing first-pass classifications. All of that work runs under the binding rules above, including the standing requirement to match against the existing registry before proposing any new entry. Deterministic tools handle retrieval and storage; models handle pattern recognition; and the judgment calls, every coinage, every merge, every retirement, every published page, are adjudicated by a person.

The analytical commentary on published pages is original editorial writing, accountable to a named author, not generated filler. The rubric-driven pipeline exists precisely so that machine-scale work stays checkable by humans, and so that any entry can be traced back to the real copy that justified it.

Identifiers and citation

Each entry has a stable ID, separate from its URL, in the format PT-{role}-{number}. PT-DSR-0042 is the 42nd entry filed under DISRUPT. IDs do not change. If you cite PT-DSR-0042 and check back in five years, it will point to the same conceptual entity. Slugs can change as usage settles on better names; old slugs redirect.

The general citation for the work as a whole:

Bowes, Jonathan Elijah. (2026). The Persuasion Taxonomy. Coppica. https://taxonomy.coppica.com

Citation formats for the whole work, individual roles, and individual entries live at /cite.

Versioning and corrections

The catalog ships in versions. Major versions mark structural change: a role added or removed, a real revision to the architecture. Minor versions roll as entries are added or revised. The changelog records what changed and when.

The catalog will get things wrong: misattributions, definitions that drift, classifications that deserve a second look. Corrections are part of the method, not an embarrassment to it. The errata page explains what to report and what happens when you do; material corrections are recorded in the changelog rather than silently patched.

This is also the page's honesty clause. The taxonomy is a soft-science classification. There is no atomic weight to verify. Some entries will turn out to be one technique under two names, and some will need to split. The essays mark which of their claims are established, which are inferential, and which are speculative, and the catalog inherits that posture: confidence is stated, not implied.

License

The catalog is released under Creative Commons Attribution 4.0 International (CC BY 4.0). Read it, quote it, teach it, build on it, translate it. The one requirement is attribution: credit The Persuasion Taxonomy with a link, the way you would credit any reference work.

CC BY is intentionally permissive. A reference work that is free and citable does more for the field than one that is locked down.

For machines

This catalog is built to be read by software as well as people, and machine citation is welcome on the same terms as human citation: use it, with attribution.

A curated, machine-readable index lives at /llms.txt. Entries and roles publish structured data (schema.org DefinedTerm and DefinedTermSet) alongside the human-readable pages. PT-IDs are the stable handles to cite or store; URLs are the canonical retrieval points; the use policy states what automated reuse the license permits. If you are building an agent, an assistant, or a retrieval system that needs a reliable reference for persuasion techniques, this is what the catalog is for.

What this isn't

It isn't exhaustive. No catalog of persuasion ever will be.

It isn't the only valid way to organize the field. By industry, by era, by psychological mechanism, by linguistic structure: each cut would produce a different useful catalog. This one is organized by function in the reader because that is the angle the underlying software needs.

It isn't a substitute for primary research. The catalog points at sources. Read Schwartz. Read Cialdini. Read the papers cited in the Green Desert essay. A reference work is a map, not the territory.

It isn't neutral on persuasion's ethics. The 18 Theses argue a position: legitimate persuasion is the kind that stays true across all nine dimensions. You can disagree with the position and still use the catalog.

Who stands behind it

The taxonomy is researched, adjudicated, and published by Jonathan Elijah Bowes, the founder of Coppica, with AI doing the volume work under the rules described above and a human owning every judgment call. It is published openly because a reference work is more useful in public than in a private backend, and because the field's real intellectual tradition deserves a working map that anyone can read, cite, and argue with.

Argue with it. That is what the errata page is for.