Every few weeks someone asks us why Mavrtr's ad copy reads so differently from ChatGPT's. The answer isn't that we have a better model. It's that we don't ask the model to write the copy from a prompt.
Here's the pattern we kept seeing before we built the pipeline, and the one that still trips up most teams using general-purpose LLMs for ads.
The convergence problem
Prompt a general-purpose LLM with "write a Facebook ad for [product]" enough times and you'll notice the outputs start to rhyme. Same hook patterns. Same emotional appeals. Same CTA structures. Different products, identical voice.
This isn't a model quality problem. It's a context problem. When the only information the model has is your prompt, it falls back on the most common patterns in its training data — which is exactly the marketing copy that already exists everywhere. The result reads like marketing because that's the source material.
The thing we noticed: the prompts that produced the least generic copy were the ones where the user had pasted in a real customer review or forum quote before asking for the rewrite. The output suddenly had vocabulary it couldn't have invented.
That observation is the entire premise behind how we built Mavrtr.
What grounding actually changes
When we shipped the first version of the research pipeline, we expected the segments to feel different. The thing that surprised us was how much the language inside the ad copy changed.
Three patterns we keep watching:
Specificity replaces hedging. Ungrounded copy defaults to safe abstractions — "feel confident," "elevate your routine," "experience the difference." Grounded copy uses the exact construction customers used in a review. We rarely see "feel confident" in real reviews. We see "I stopped second-guessing the outfit before I left the house."
Objections show up in the copy. Ungrounded copy almost never names the objection a buyer has. Grounded copy does, because the reviews and forum posts feeding the pipeline are full of buyers stating their objections explicitly. The result feels less like an ad and more like a conversation the buyer was already having.
Hook patterns get product-specific. Ungrounded copy reuses the same five hook templates across every category. Grounded copy pulls hooks from the actual market — whatever pattern competitors have been scaling in the Meta Ad Library plus whatever phrasing keeps repeating across reviews. The hooks become category-native instead of generic.
Where this breaks down
Grounded copy isn't strictly better. There are workflows where it's the wrong tool.
If you're brainstorming a creative concept before research, grounded copy will narrow you too early — the pipeline is anchored to what already exists in the market. A prompt-only LLM is genuinely more useful for exploring left-field angles.
If you're writing for a market with very little public review data, the grounding signal is too thin and you may as well prompt directly. We see this with brand-new categories or B2B niches where Trustpilot has 12 reviews and the relevant subreddit has 4,000 subscribers. Mavrtr will still produce a brief, but the segments will lean on weaker signal.
And if you're iterating on a specific variant — tightening one line of copy, testing a different CTA — a general-purpose LLM is fast and frictionless. We do this internally all the time. Creative intelligence isn't always the right altitude for the task.
The practical takeaway
The category split that actually matters is:
| You're doing | Use |
|---|---|
| First draft for a real product launch | Creative intelligence (Mavrtr or equivalent) |
| Brainstorming creative concepts | Prompt-only LLM |
| Iterating on existing copy | Prompt-only LLM |
| Writing for a new category with no public data | Prompt-only LLM, but expect more rewriting |
| Scaling research across many products or competitors | Creative intelligence |
Most teams we talk to default to prompt-only because it's faster and they've already paid for the subscription. The shift happens once they spend a Saturday manually researching a single Shopify product and realize the model was generating from the wrong context the whole time.
That Saturday is the reason Mavrtr exists.