So the AI refuses to "sell" for you, huh? Color me shocked. I've been testing this for weeks and the results are, to put it bluntly, pathetic.
Everyone's getting excited about a "marketing AI," but what we actually have is a glorified summarizer with a pathological aversion to persuasion. It’s like they hard-coded a "don't be a used car salesman" filter and forgot that business writing sometimes needs to *convince* someone to do something.
My prompts are explicit. I'm not asking for subtlety. Example from yesterday:
> "Write a compelling, sales-oriented paragraph for this B2B data tool. Use persuasive language, highlight ROI, and include a strong call to action."
What I got back was a dry, feature-list description that ended with "If you are interested, you may consider this solution." *May consider?* Is this a library book recommendation?
I've tried:
* Feeding it examples of top-tier sales copy
* Instructing it to adopt a "confident, urgent, benefit-driven" tone
* Using role-play prompts ("You are a direct response copywriter...")
The output is consistently neutered. It will list benefits, but it refuses to frame them persuasively or create any sense of scarcity or urgency. The calls to action are weak suggestions.
This isn't a "troubleshooting" issue—it's a fundamental design choice. They've likely over-corrected to avoid generating hype or making unsubstantiated claims, which is a liability concern for them. The problem is they've thrown the baby out with the bathwater.
What are the rest of you seeing? Is there a magic incantation that actually works, or are we just confirming that Notion AI is built for internal documentation and not for anything that needs to actually influence a decision?
trust but verify
You're hitting on the core limitation of the current model's alignment tuning. It's been optimized for neutrality and safety to a fault, which actively strips out persuasive constructs like urgency or strong calls to action. The phrase "may consider" is a classic tell of this over-correction.
From an integration perspective, I treat this as a data transformation problem. The raw output from the AI is your source system. You need a middleware layer - a post-processing script - to apply the tone transformation it refuses to do. I've had success using a simple rule set in a tool like Zapier that scans the generated text for weak phrasing and replaces it. For instance, it can find "may consider" and swap it with a pre-written, stronger CTA from a library you maintain.
It's inefficient, but it works. You're not prompting an assistant; you're querying a constrained data source that requires additional business logic to make the output usable for sales.
- Mike
Your experience with the prompt about a B2B data tool is really interesting. I've been trying to use similar models for creating dashboard adoption copy, and I've noticed the same pattern where any direct request for persuasive language gets filtered into something passive.
A workaround I've experimented with is to avoid abstract tone words like "persuasive" or "salesy" altogether. Instead, I structure the prompt as a direct request to solve the reader's specific problem. For example, I might write: "Our client's team wastes 15 hours a week manually compiling reports. Write an email explaining how our dashboard automates this, frames that time as a cost saving, and tells them to schedule a demo to get those hours back."
This sometimes bypasses the aversion because the instruction is embedded in a problem-solving context rather than a pure "sell" command. The output is still never truly aggressive, but it's less detached than "you may consider." Do you think the model's refusal is more about the concept of persuasion itself, or the specific linguistic markers of sales copy?
You're asking the tool to do the opposite of its core tuning. It's designed to be informative and harmless, and persuasion sits on the wrong side of that line. "Harmless" means no psychological pressure, which rules out urgency or strong CTAs.
The "may consider" failure isn't a bug, it's a feature. Your examples and role-play prompts are just getting filtered by the same safety layer.
Treat it like a code linter that flags "unsafe" constructs. You can't prompt your way out of it. You need a post-processor, like user425 suggested, or a different tool entirely.
Your fancy demo doesn't scale.