Hi, it's Peggy.

John Caples spent the early part of his career documenting something most copywriters know intellectually and forget in practice. The headline you pick on a Tuesday can do twenty times the work of the one you almost picked. He published the number in 1932. The discipline he built around it took most of the industry forty years to adopt.

I want to talk about what's happened to that discipline since Claude and ChatGPT made the generation step nearly free.

The discipline got cheaper. The skill it still asks for is the same one it asked for in 1925.

Here's the case.

The product is a scheduling tool for independent service providers — think massage therapists, tutors, personal trainers. The kind of solo operator who books appointments, takes payments, and sends reminders. A real category with real competitors.

In 1932, John Caples published Tested Advertising Methods. Inside that book, he documented a finding from his split-testing work at Ruthrauff & Ryan and BBDO. Across the headline tests he had run, the best-performing headline outpulled the worst-performing headline by 19.5 times.

That number has appeared in every serious direct response book since. It is the most quoted statistic in copywriting. Almost nobody acts like it.

Most working copywriters know the gap exists. They write three or four headlines for a project, pick the one that sounds best, and move on. Caples's actual teaching was something different. He taught that sounds best is a much weaker filter than tested best, and that the difference between the two, accumulated across a career, is enormous.

Caples's methodology was crude by modern standards. He held everything constant except the headline. Same product, same offer, same ad layout. Response was measured by tracking mail-in coupons, each printed with a key code that pointed back to which ad pulled it.

A single test could take weeks. Samples per treatment were small. But the discipline was rigorous in the ways that mattered most. The variable was isolated. The dependent measure was real revenue rather than a proxy for it.

What came out of years of this work was a catalogue. Caples avoided calling them rules. They were tested patterns:

The pattern that held across every category was specificity. Caples kept finding it. Specific outperformed general almost every time.

The most cited example of his specificity rule is the 1925 piano headline he wrote for the US School of Music. They Laughed When I Sat Down at the Piano — But When I Started to Play! runs longer than most modern advertisers would tolerate. It also tells a specific scene, names a specific instrument, and promises a specific reversal of social embarrassment.

The headline ran for years and outpulled every variant tested against it. Caples returned to it in Tested Advertising Methods as the canonical example of what specificity does that generality cannot.

The expensive part of Caples's method, until very recently, was generating the variations. A working copywriter could write five strong headlines in an afternoon. Ten if it was a good day. Twenty was a full day's work, and most of the twenty would be obviously worse than the first five.

That generation cost was the bottleneck. Most working copywriters got to five, called it enough, and tested those. More commonly, they didn't test at all and picked the one that read best.

LLMs removed the generation bottleneck. You can produce twenty headlines in twenty minutes now, or thirty, or fifty. The marginal cost of variation went near zero.

What that should mean is that every working copywriter on every project runs the Caples test. Generate variations, send the top candidates to a real audience, let the data pick.

In practice, most working copywriters use Claude or ChatGPT as a faster way to generate the first five and then revert to the same practice they used before. They pick the one that sounds best. They skip the test.

The 19x gap is still there. The cost of generating variations got lower. The gap itself did not move. The discipline of choosing well, which is the discipline that actually closes the gap, is the same one it was in 1925.

Two psychological forces work against the test even when it is cheap.

This is the prompt I run when I need to produce headline variations for a client project. It encodes the categories Caples documented as tested winners, plus the specificity rule he kept finding across every test.

Here is what this prompt is doing:

The five categories force the model into Caples's tested structures rather than letting it default to whatever "headlines" looks like in its training data. Without category constraints, the model produces a regression to the mean of headline-shaped strings, which is exactly the generic output most copywriters complain about.

The specificity rule is built in as a replacement requirement rather than a preference. If the model writes "more energy," the rule tells it to replace the phrase with a named outcome. That forces the prompt to do the work of Caples's editorial pass.

The "flag any variation where specificity requires proof I haven't given you" line is the protection against the model inventing numbers. LLMs will produce plausible-sounding specific claims to satisfy a specificity rule, and those claims will not always be true about your offer. Asking the model to flag the unverified ones puts the check inside the workflow.

The ranked top 5 with one-sentence reasoning is the starting point for an actual test. You don't need to take the ranking. You do need to look at it and pick which two or three you'll run against the control.

Most working copywriters don't generate every headline from scratch. They inherit headlines from clients, from previous campaigns, from internal teams who already shipped the page. The Caples test in those cases is auditing the headline you already have rather than generating twenty new ones.

Here is the audit prompt I run on inherited copy.

The fifth step is where the value lands. If the rewrite is a meaningful improvement on the original, the rewrite becomes your first new variation and you can run the generation prompt to produce nineteen more in the same direction.

If the rewrite is not a meaningful improvement, the inherited headline is probably already close to the ceiling for what's possible without better proof in the offer. That itself is a finding worth bringing back to the client.

If you're going to run the Caples test regularly, the prompt is faster as a persistent project than as a copy-paste each time. The setup takes about twenty minutes once. After that, every headline test you run starts a fresh conversation inside the project, with the framework already loaded.

Build a Claude Project called something obvious like "Headline Lab" or "Caples Test." Put the following in the custom instructions field so the model carries the framework into every conversation in the project.

Attach two files to the project. The first is a running offer document with your client offer descriptions and current control headlines, updated as projects come in. The second is a winner library — a running file of headlines that have actually won A/B tests for your clients, with the offer they ran against and the lift they produced.

Over time, the winner library becomes additional context the model draws on, and the variations it produces get sharper for your specific audience.

The winner library is the part that compounds. Caples spent decades documenting what worked across the industry. Your library does the same thing at the scale of your client base, and the model gets more useful for you with every test you log.

Three honest limits before you build this into a workflow.

Here is a concrete example from a recent test. I ran the prompt for a B2B SaaS landing page where the offer description I gave the model was light on numbers.

The model returned a "How I" headline that named a specific dollar amount the client had never claimed in any material I had passed it. The headline read well. The number sounded plausible. The number was also invented. The flag-unverified-claims line in the prompt caught it on review, but only because I had told the model to flag the claims it was unsure of. Without that step, the invented number would have gone into the A/B test against the control.

The 19x gap is the most important number in copywriting that almost nobody acts on.

Caples wrote it down in 1932, and the discipline of actually closing it stayed expensive for ninety years. LLMs made the test cheap to run.

The skill that closes the gap, which is running the test, reading the data, and choosing the headline that earned its place, is the same skill it asked for in 1925.

More clicks, cash, and clients,
Peggy Burnett