How can marketers use GPT-3 without sounding generic?

Marketers who try GPT-3 often have the same first reaction: it writes fluently, quickly and with no personality at all. The drafts lean on phrases like "the next level" and "perfect for". Readers have seen those lines many times, and they tend to skim past them.

The blandness is not a fault in the model so much as a mirror of the prompt. Ask for "an ad for a coffee mug" and GPT-3 writes the most typical ad for a coffee mug, built from the phrases that show up most often in ads. Give it something specific to work from and the writing becomes specific too.

Why the first draft sounds like everyone else

GPT-3 writes by predicting likely text. Researchers Laria Reynolds and Kyle McDonell describe the output as spread across all the ways anyone might continue a prompt. When the prompt gives little to go on, the likeliest continuation is the most common one, which in marketing means stock phrases.

OpenAI's own documentation shows the effect. One of its example prompts asks for a creative Facebook ad aimed at parents for a learning product, and the sample reply leans on lines such as "Look no further" and a promise to help the child get ahead. It is correct, on topic and forgettable. Nothing is wrong with the model; the prompt simply described a category, not a brand.

The worked example: one product, two prompts

Take an invented small business: a pottery studio selling a hand-thrown mug. The thin prompt:

Write a product description for a handmade coffee mug.

Expect warmth, craftsmanship and "the perfect gift". Now add four ingredients: the audience, real facts, a sample of the brand's voice, and words to avoid.

Write a product description for an online shop.

Audience: people who work from home and drink a lot of coffee; they care about objects that last.
Product facts: hand-thrown stoneware mug, holds 350 ml, matte speckled glaze, thick wall that keeps coffee hot longer, dishwasher safe, each one slightly different, made in our two-person studio.
Our voice (sample from our site): "We make plain things for daily use. No slogans, just good clay and a steady hand."
Avoid these words: perfect, elevate, game-changer, unique, luxury, look no further.
Length: 60 to 80 words. End with one practical detail, not a call to action.

Description:

The second prompt gives the model facts it cannot invent, a rhythm to imitate and a fence around the clichés. A draft from it should mention the thick wall, the speckled glaze and the small studio, because those are now the most relevant words available.

Then refine. Read the draft aloud, mark one sentence that sounds like anyone's shop, and ask for a revision with that note:

Rewrite the description. Keep the facts. The second sentence sounds like a generic shop; replace it with something only our studio would say. Use shorter sentences.

It is usually easier to improve a draft over two or three rounds than to write one flawless prompt.

Settings that help

• Temperature. OpenAI's docs explain that lower values make output less random and values near zero make it repetitive. For product copy, a middle setting gives some variety without wandering.
• Several drafts. Ask for three versions and combine the best lines. Variety is cheap; a good sentence is not.
• Facts check. GPT-3 can add features you never mentioned. Every claim in the final copy should come from your product sheet.

How do you give it your brand voice?

Start by describing the voice in contrasts, because single adjectives are vague. "Friendly" means one thing to a bank and another to a skateboard shop. "Plain, not folksy. Confident, not pushy. Short sentences, no exclamation marks" gives the model, and any human writer, something to follow.

Then show, don't just tell. Two or three short passages of your best existing copy, placed in the prompt before the request, set the rhythm and vocabulary more reliably than a description. Reynolds and McDonell make a related point: framing a prompt with a familiar situation or example carries a lot of intent that would take many words to spell out.

When you need the voice across hundreds of pieces, consider fine-tuning. Since December, OpenAI has let every API customer train a custom version of GPT-3 on their own examples. The company says gains appear with fewer than 100 examples and keep improving as the dataset grows, and the shorter prompts that follow cut cost per request. A few hundred of your past product descriptions, each paired with its product notes, is a natural starting set.

Whatever route you choose, have a person read every draft against the product sheet before it goes live, since only you know which claims your business can stand behind.