Stu Sjouwerman is co-founder and CEO of ReadingMinds, a pioneering AI-moderated interview platform for conducting sentiment analysis.
getty
Read enough AI-generated marketing content and you’ll start to see a pretty consistent pattern. The vocabulary is right, the structure is clean, the argument is coherent and yet it just doesn’t sound like it came from anyone in particular. It sounds disembodied.
That’s not a coincidence. AI language models are trained largely on publicly available text, including an enormous volume of existing, often average marketing content. So, when a model learns to write marketing copy, it learns from the full repository of everything that has already been written—including every forgettable blog post, every bland value proposition and every landing page that says "streamline your operations" without explaining what that means. It is designed to optimize for outputs that are broadly acceptable across the widest possible range of readers.
That generic output doesn’t represent a failure. The model is performing exactly as designed.
AI prompts are typically built on standard industry terminology, rehashed competitive positioning and category language. In short, these prompts ask AI to use the same vocabulary the same way it’s already used—the way “everybody” uses it.
Generic prompts cannot give these models something they don’t already have, and what they don’t have is the specific, unscripted language your customers use when they describe the problem you solve—in their own words, in their own voice and before anyone has coached them into the language of your category.
There’s a certain kind of language that doesn’t exist on the internet: bona fide customer conversations. These are the unscripted moments when customers explain, unprompted, why they almost didn’t buy your product or service, or the specific phrases a prospect uses to describe the problem they’ve been having for two years, or the exact words a longtime user chooses when they describe what changed the day your product stopped working for them.
That language is proprietary. So, when you feed it into a content brief, the model produces something it couldn’t otherwise produce. The output is genuinely specific because the input was genuinely specific.
This is why many of the teams producing the most distinctive AI-assisted marketing content aren’t the ones with the best prompt engineers. They’re the ones doing the best customer research.
Text-based research captures only what customers are willing to type—a narrow slice of their actual experiences, perceptions or feelings. Survey responses are shaped by the questions being asked, and written feedback is self-edited.
Voice helps to remove those filters. When a customer speaks, they use different words than they would write or check off a script. They backtrack, amplify, choose metaphors that reveal how they actually feel about the problem and say things they would never jot in a survey. This is because the conversation naturally carried them there.
The output isn’t a transcript; it’s a library of specific, emotionally grounded customer language that has never appeared in any model's training data.
If your AI-generated marketing content still sounds like everyone else's, the place to start is what happens before the prompt. Stop writing for your category and start writing from your customer’s point of view. Category language is what every competitor also has access to. Customer language, captured in the customers' voices before anyone has polished it, is what only you can collect. That gap is where differentiation lives.
Treat spoken language as a content input, not a research output. Most teams use customer interviews to generate reports. The specific phrases, unexpected metaphors and precise words a customer chooses represent raw material.
Additionally, make sure to evaluate your content. Could anyone have written it? If a competitor could publish the same piece with their logo swapped in and nothing else changed, the content isn’t doing its job. Specificity is the only defense against generic content, and the specificity that matters most is customer language that only you have collected.
With that said, collecting proprietary language comes with responsibilities. Consent and disclosure are the starting point. At a minimum, for every interview or voice-based research session, let the participant know their conversation is recorded, transcribed pseudonymously and used to inform the marketing campaign.
Audio should not be stored indefinitely, and systems should be designed to avoid generating biometric identifiers or voiceprints unless explicit consent has been obtained. This is especially important for compliance with regulations such as the Illinois Biometric Information Privacy Act (BIPA), which restricts the creation and storage of biometric identifiers, including voiceprints, without informed consent.
Using a customer’s exact words and phrases is also very different from publishing a customer’s identity, personal story or attributed testimonial. Organizations can and should use the vocabulary, metaphors and emotional framing surfaced in customer conversations to strengthen AI-assisted content briefs.
However, the operational standard should be clear: by the time customer language reaches a content brief, it should already be aggregated and anonymized. No individual should be identifiable in the prompt, no quote should be reproduced verbatim without confirmed usage rights and no sensitive personal context shared during an interview should appear in public-facing content.
Building this into the content workflow as a documented operational step is how responsible teams protect both their customers and the integrity of their research. Ultimately, competitive positioning depends on what you understand about your customers that nobody else does, and whether that understanding is reflected in your content.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。