Quick Answer: Claude Hopkins' 1923 Scientific Advertising made measurement, specificity and complete answers the laws of effective selling. A century later, Princeton's GEO paper (Aggarwal et al., 2023) tested content tactics against AI engines and found the same rules win: quotations, statistics and cited sources lifted visibility in AI answers by up to 40 percent.

Two documents, published one hundred years apart, describe the same discipline. The first is a slim book by a Chicago copywriter who tracked coupon returns down to the fraction of a penny. The second is a peer-reviewed computer science paper, presented at KDD 2024, that measured how content changes move visibility inside answers generated by large language models. Neither cites the other. They did not need to. Both started from the same question, what does the evidence say actually works, and both arrived at the same answers: be specific, cite your sources, answer the whole question, and measure everything.

This piece reads the two texts side by side, principle by principle. Hopkins is quoted at length because his 1923 text is public domain and because his sentences have not been improved on. The GEO paper's findings are paraphrased with attribution; a full walk-through of its method and numbers lives in our companion piece, the GEO paper, explained.

Two texts, one discipline, a century apart

Hopkins opens Scientific Advertising with a claim that must have sounded arrogant in 1923: "The time has come when advertising has in some hands reached the status of a science. It is based on fixed principles and is reasonably exact" (Chapter 1). His evidence was keyed advertising: coupons coded so every reply traced back to the exact ad, headline and publication that produced it. "Mail order advertising is traced down to the fraction of a penny," he wrote in the same chapter. "The cost per reply and cost per dollar of sale show up with utter exactness."

The GEO paper runs the identical experiment against a new machine. Aggarwal and colleagues built GEO-bench, a benchmark of 10,000 queries across 25 domains, applied nine different content modifications to source websites, and measured how each modification changed the site's share of the AI-generated answer. Same logic as the coupon: change one variable, trace the return, keep what wins. The paper reports that the best methods improved visibility by up to 40 percent, and that the winning methods were not tricks of the ranking system but improvements in the evidence the content carried.

That is the thesis of this article in one sentence: generative engines have rebuilt the traced-return environment Hopkins worked in, and the tactics his tracing proved are the tactics the Princeton team's tracing proved. Everything below is the pairing, one Hopkins principle at a time.

Just salesmanship becomes just citability

Hopkins' second chapter contains the most quoted lines he ever wrote: "Advertising is salesmanship. Its principles are the principles of salesmanship. Successes and failures in both lines are due to like causes." And the corollary that disciplined every word he published: "The only purpose of advertising is to make sales. It is profitable or unprofitable according to its actual sales. It is not for general effect. It is not to keep your name before the people" (Chapter 2).

Swap one noun and the paragraph is current. The only purpose of web content, in an AI-mediated search world, is to be cited. It is profitable or unprofitable according to whether a generative engine selects it as a source when a buyer asks a question. It is not for general effect. It is not to keep your name in a blog feed. The GEO authors formalize this by replacing the old visibility measure, average ranking position, with impression metrics built for generated answers: how many words of the answer trace to your citation, how early it appears, how much the response relies on it. A page can rank and still contribute nothing to the answer a prospect actually reads.

Hopkins told advertisers to interrogate every line: "Would this help a salesman sell the goods?" The current form of the question is just as blunt. Would this passage help an engine answer the buyer's query? If not, it is decoration, and Hopkins had a verdict on decoration too: "Ads are not written to entertain" (Chapter 2).

Offer Service: the answer comes before the pitch

Chapter 3 of Scientific Advertising is four pages long and could be republished today as a GEO briefing. "The best ads ask no one to buy," Hopkins wrote. "The ads are based entirely on service. They offer wanted information." His examples are all structured the same way: the brush salesman who opens by giving a brush away, the coffee wagon that leaves a half-pound on trial, the motor maker whose ad said "Let us help you for a week without cost or obligation" and converted about nine trials in ten.

A generative engine enforces this principle mechanically. When someone asks ChatGPT or Perplexity how to choose between two products, the engine retrieves pages and lifts the passages that carry the answer. A page that opens with the company's own praises gives the model nothing to lift. A page that answers the question first, plainly and early, is the raw material the answer gets built from. This is Vector 4, Embed, in our 12-Vector framework: write the answers the engines extract, then let the commercial invitation follow. The 12 Vectors overview maps the whole sequence.

Hopkins closed the chapter with a warning that reads like a note to every agency still publishing pitch-first content: "People can be coaxed but not driven. Whatever they do they do to please themselves" (Chapter 3). The engine, acting as the buyer's proxy, cannot be driven at all. It takes the most serviceable source available and ignores the rest.

Being Specific: the claim an engine can verify

If one Hopkins chapter predicts the GEO paper's results table, it is Chapter 7. "Platitudes and generalities roll off the human understanding like water from a duck. They leave no impression whatever," he wrote. "But a man who makes a specific claim is either telling the truth or a lie. People do not expect an advertiser to lie... So a definite statement is usually accepted. Actual figures are not generally discounted. Specific facts, when stated, have their full weight and effect."

His examples are wonderfully concrete. "Say that a tungsten lamp gives more light than a carbon and you leave some doubt. Say it gives three and one-third times the light and people realize that you have made tests and comparisons." The retailer whose "Lowest prices in America" slogan had been copied into meaninglessness switched to "Our net profit is 3 percent" and grew sensationally. The shaving-soap maker who replaced "Abundant lather" with "Multiplies itself in lather 250 times" and "Softens the beard in one minute" produced, in Hopkins' telling, one of the fastest wins ever recorded in a difficult field.

Now set the 2023 results beside that. Of the nine methods the GEO team tested, the top performers were Statistics Addition (converting qualitative talk into quantitative claims), Quotation Addition (adding credible, attributed quotes) and Cite Sources (naming where claims come from). Those three delivered roughly 30 to 40 percent relative improvement on the position-adjusted word count metric. The paper's own qualitative examples read like Hopkins exhibits: a source about workplace robotics gained visibility after a vague trend statement became a dated percentage; a chocolate-consumption page gained 132.4 percent after its central claim was attributed to a named research group. The mechanism a century ago was reader trust. The mechanism now is that a language model, asked to ground an answer, weighs a verifiable, attributed figure more heavily than an adjective. Same law, new enforcer.

Tell Your Full Story: the complete, extractable answer

Brevity was the fashionable advice in Hopkins' day, and he spent Chapter 8 demolishing it. "When you once get a person's attention, then is the time to accomplish all you can ever hope with him. Bring all your good arguments to bear. Cover every phase of your subject. One fact appeals to some, one to another. Omit any one and a certain percentage will lose the fact which might convince." His traced returns settled the argument: "The motto there is, 'The more you tell the more you sell.' And it has never failed to prove out so in any test we know" (Chapter 4, on mail order copy).

His best illustration is a buyer, not a seller. A wealthy man shopping for a personal car leaned toward a Rolls-Royce, but the prestige marques ran short, dignified ads that told him nothing. "The Marmon, on the contrary, told a complete story. He read columns and books about it. So he bought a Marmon" (Chapter 8). The complete story won the sale from brands with better names.

Retrieval works the same way. An engine assembling an answer favours sources that cover the question fully in one place: the page that states the answer, quantifies it, handles the objections and defines the terms gives the model everything it needs without a second retrieval. Thin pages that each hold a fragment of the story lose to the single page that holds all of it. In Matt's testing across Ontario service categories, the pages the engines quote are consistently the long, complete treatments, not the five-hundred-word summaries that outrank them for some keywords. The reader who arrives by AI answer is Hopkins' one-time reader: "That reader, if you lose him now, may never again be a reader" (Chapter 8).

Headlines were targeting; quick answers are extraction

Hopkins treated the headline as a sorting device, not a flourish. "The purpose of a headline is to pick out people you can interest. You wish to talk to someone in a crowd. So the first thing you say is, 'Hey there, Bill Jones' to get the right person's attention" (Chapter 5). He reported keyed returns on nearly two thousand headlines for a single product and noted, "It is not uncommon for a change in headlines to multiply returns from five to ten times over." The body copy stayed nearly identical; the address on the envelope did the work.

The modern equivalent of the headline's job has split in two. Half of it still belongs to titles and H2s, which tell both readers and retrieval systems what a passage answers. The other half now belongs to the answer-first block: the short, definitive statement at the top of a page that an engine can lift whole. The GEO paper's position-adjusted metric explains why placement pays; sentences early in a response, and passages early in a source, carry more weight, mirroring the click-through power law the authors cite from search-era studies. Structure a page so its opening fifty words answer the query outright and you have written the modern "Hey there, Bill Jones": you have hailed the machine that hails the buyer.

Test Campaigns: measurement is the whole method

Chapter 15 is the book's engine room. "Almost any questions can be answered, cheaply, quickly and finally, by a test campaign. And that's the way to answer them, not by arguments around a table. Go to the court of last resort, the buyers of your product." And the line that defines scaled testing to this day: "Now we let the thousands decide what the millions will do. We make a small venture, and watch cost and result. When we learn what a thousand customers cost, we know almost exactly what a million will cost."

The GEO paper is a test campaign in Hopkins' exact sense, run at a scale he would have envied: nine treatments, 10,000 queries, five random seeds per experiment, results averaged and compared against an unmodified control. And its findings argue for continued testing rather than settled doctrine, because the winning tactic varied by domain. Statistics carried the most weight in law, government and debate-style queries; quotations won in people, society and history topics; citations won on factual questions. The authors conclude that content owners should make domain-specific adjustments, which is Hopkins' "We learn on each line by experiment" (Chapter 5) restated in benchmark form.

There is one modern complication Hopkins never faced: the buyers now sit behind four different machines. ChatGPT, Gemini, Claude and Perplexity retrieve differently and cite different sources, so a single aggregate number hides which engine your work moved. This is why Vector 11, Measure, tracks citations per engine rather than one blended score. The GEO team demonstrated the transfer themselves by re-running their methods on Perplexity.ai, where quotation addition improved visibility 22 percent over baseline and statistics addition improved the subjective impression metric by 37 percent. Different engine, same winning tactics, separately measured. Hopkins would have keyed each one.

What fails in both eras: stuffed keywords and empty superlatives

The negative results agree as neatly as the positive ones. The GEO team tested keyword stuffing, the classic move of repeating the query's words through the content, and found it performed at or below the do-nothing baseline; on Perplexity it came in about 10 percent worse than leaving the page alone. The authors' reading is that a generative model is "not limited to keyword matching," so tactics aimed at lexical overlap simply miss the mechanism. They found the same for injected persuasive tone: making the text more authoritative-sounding produced no significant visibility gain in their aggregate results, because the engines largely shrug off tone.

Hopkins ran the 1923 version of that experiment and got the 1923 version of that result. "To say, 'Best in the world,' 'Lowest prices in existence,' etc. are at best simply claiming the expected. But superlatives of that sort are usually damaging. They suggest looseness of expression, a tendency to exaggerate, a carelessness of truth" (Chapter 7). Empty emphasis did not merely fail to help; it taxed the credibility of every true statement on the page. Any business still buying content stuffed with its target phrase, or padded with unverifiable boasts, is paying for the tactic both centuries have measured and rejected. It is one of the patterns worth checking for before you sign with any provider; our guide to choosing an SEO agency lists the rest.

The finding Hopkins never had: the machine favours the underdog

One result in the GEO paper has no 1923 parallel, and it matters most to the businesses we work with. When the team optimized all sources at once, the gains concentrated at the bottom of the rankings: the fifth-ranked source gained 115.1 percent visibility from adding citations while the top-ranked source lost ground. Their explanation is structural. Traditional rankings lean on backlinks and domain authority, signals that take years and budgets to accumulate. A generative model conditions on the content itself, so a small site with better evidence can out-argue a large site with better links. The authors describe GEO as "a tool to democratize the digital space" for small creators and independent businesses.

Hopkins believed disciplined method made advertising safe for the modest budget; "Advertising, once a gamble, has thus become, under able direction, one of the safest business ventures" (Chapter 1). But he never claimed the method favoured the small advertiser over the giant. The evidence now says the new surface does, provided the small player does the evidentiary work. That finding is the research backbone of our own positioning: the independents we serve cannot outspend national chains on links, but they can out-evidence them, and the engines reward exactly that. It is what our GEO service is built to execute, vector by vector.

The Hopkins test for your pages, 2026 edition

Fold the two texts together and you get a four-question audit any owner can run on a key page this afternoon.

Then measure. Ask the four major engines the questions your buyers ask, record which sources they cite, change one thing, and ask again. That loop, Diagnose through Iterate in our framework, is nothing but Hopkins' keyed coupon rebuilt for a machine reader.

Matt Griffin, Formative Digital: "I keep a copy of Scientific Advertising on the shelf behind my desk, and I re-read Chapter 7 before every major content engagement. What struck me when the Princeton paper circulated was not that it contradicted the old direct-response canon, it was that a team with GPUs and a 10,000-query benchmark had independently rediscovered a book written before my grandparents were born. We named our operating principle 'Engineering Principles, not magic' before I made that connection explicit. Hopkins got there first: every claim tested, every result traced, no guesswork honoured. The machines changed. The discipline did not."

Run the Hopkins test on your own site

Request a no-charge AI visibility audit and we will show you, engine by engine, where your pages stand in ChatGPT, Perplexity, Gemini and Google AI Overviews, and which of the four checks above they currently fail. Reply within one business day.

Frequently asked questions

What is Scientific Advertising and why does it still matter for AI search?

Scientific Advertising is Claude C. Hopkins' 1923 book arguing that advertising should run on measured results rather than opinion. He built the case from keyed coupons and traced returns: compare versions, keep what the numbers prove, discard what they refute. AI search rewards exactly that discipline. The 2023 Princeton GEO paper tested content changes against generative engines the same way Hopkins tested headlines, and the tactics that won were his: specific statistics, credible quotations, cited sources and clear, complete answers.

What did the Princeton GEO paper actually find?

Aggarwal et al. (2023) tested nine content-optimization methods across 10,000 queries in their GEO-bench benchmark, measuring how much visibility each method earned inside AI-generated answers. Adding quotations, statistics and source citations improved visibility by roughly 30 to 40 percent on their position-adjusted word count metric, while keyword stuffing performed at or below the unoptimized baseline. They confirmed the pattern on Perplexity.ai, a live commercial engine, and found lower-ranked sites gained the most, in one case a 115.1 percent visibility lift for a fifth-ranked source.

Which Hopkins principles map most directly to Generative Engine Optimization?

Five map almost one to one. Just Salesmanship becomes citability: content is judged by whether an engine uses it, not by how it reads. Offer Service becomes answering the query before pitching anything. Being Specific becomes verifiable numbers and named sources, the exact tactics the GEO paper measured as top performers. Tell Your Full Story becomes the complete extractable answer on one page. Test Campaigns becomes per-engine measurement, since each AI engine cites different sources and only tracked results tell you which changes worked.

Does keyword stuffing still work in AI search?

No. In the GEO paper's benchmark tests, keyword stuffing scored below the unoptimized baseline on the position-adjusted word count metric, and on Perplexity.ai it performed about 10 percent worse than doing nothing. Hopkins would not have been surprised: he wrote in 1923 that platitudes and generalities leave no impression whatever. A generative model reads for meaning, weighs evidence and lifts the passage that answers the question best, so repeating the query's words adds nothing an engine values.

How do I apply these two texts to my own website?

Run each key page through the paired test. Does it answer the visitor's actual question before asking for anything, or does it open with a pitch? Does every claim carry a number, a date or a named source an engine could verify? Does the page tell the complete story, or does it scatter the answer across several thin pages? And are you measuring which AI engines cite you, per engine, the way Hopkins keyed every coupon? Results depend on your industry, competition and existing presence, but pages that fail those four checks rarely get cited anywhere.

Sources

  1. Hopkins, Claude C. (1923). Scientific Advertising. Public domain; full text via the Internet Archive: archive.org/details/scientific-advertising-by-calude-hopkins. Chapter references in this article follow the original chapter numbering.
  2. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2023). "GEO: Generative Engine Optimization." arXiv:2311.09735; presented at KDD 2024. arxiv.org/abs/2311.09735
  3. Dean, B. (2023). "We Analyzed 4 Million Google Search Results. Here's What We Learned About Organic Click Through Rate." Backlinko. Cited in the GEO paper as the basis for its position-weighted metric. backlinko.com/google-ctr-stats

If the pairing in this article describes work you want done on your own domain, talk to us. We will walk you through what the evidence says your market rewards, with the numbers on the table, in the spirit of the man who keyed every coupon.