Two chatbots on two competing websites answer the exact same question: "Do you offer annual billing?" The first says yes and stops. The second says yes, mentions annual saves two months, asks how many seats you need, and offers to send a quote. Same knowledge. Only one of them is trying to make a sale.
The difference isn't the model or the data so much as the instructions you wrote. Most bots answer well and sell terribly because nobody told them selling was part of the job.
A sales prompt is the bot's marching orders
When people say "chatbot prompt," they often mean the customer's question. Here I mean the opposite: the standing instructions you give the bot about how to behave. This is where you decide whether it's a passive encyclopedia or an active member of your sales team.
A knowledge-only prompt says, in effect, "answer questions using these docs." A sales prompt adds a goal: "help the visitor find the right product, and move them toward a demo, a quote, or a purchase." That one addition changes every reply, because now the bot is looking for a next step instead of just closing the conversation.
You're not writing a script. You're setting priorities the bot applies to whatever comes up.
Give it a goal, then the guardrails
Start the instructions with a clear objective and the honest limits around it. Something like: "Your goal is to help visitors choose the right plan and book a demo. Be genuinely helpful. Never pressure, never invent features, and if a question is outside what you know, offer to connect them with a person."
That last clause matters as much as the goal. A sales bot that oversells is worse than one that undersells, because a customer who buys the wrong thing becomes a refund and a bad review. Tell it plainly: recommend the plan that actually fits, even when that's the cheaper one.
Here's how the same underlying reply changes when the instructions carry a goal:
| Weak instruction | Result | Sales-oriented instruction | Result |
|---|---|---|---|
| "Answer questions about pricing." | "Our Pro plan is $49/month." | "After a pricing answer, ask what they need so you can recommend a plan." | "Pro is $49/month. How many people will be using it? I can point you to the plan that fits best." |
| "Help with product questions." | "Yes, it integrates with Shopify." | "Confirm the feature, then offer a next step." | "Yes, it works with Shopify. Want me to send setup steps or book a quick call to walk through it?" |
Ask questions that feel like help, not a form
The fastest way to make a bot pushy is to have it interrogate people. The fix is to have it ask one question at a time, and only when the answer lets it help better.
Tell the bot to qualify through curiosity. "Before recommending, find out their team size, their main goal, and their timeline, but ask naturally, one question per message, and only when it's relevant." A visitor asking about a feature gets a real answer plus a single follow-up, not a five-field questionnaire dropped on them mid-sentence.
Good qualifying questions double as helpful ones:
- "How many people are on your team?" helps size the plan.
- "What are you hoping to fix first?" surfaces the feature that matters to them.
- "Are you replacing another tool, or is this new?" tells you how to frame the value.
Each answer makes the bot's next recommendation sharper, and the visitor feels understood rather than processed.
There's a limit to respect, though. If someone just wants a straight answer and clearly isn't in the mood to chat, forcing a qualifying question makes you the pushy salesperson who won't let them browse. Tell the bot to read the room: when a visitor asks a direct question and shows no interest in more, give them the answer cleanly and leave the door open rather than pressing. A soft "happy to recommend something if you'd like" beats an interrogation every time.
Recommend, handle price, and don't stall
An answering bot ends conversations. A selling bot ends them with a suggestion. Instruct it to always close a helpful answer with a relevant next move: a recommended plan, a link to book, an offer to save the details.
Price objections deserve their own line in the instructions. Tell the bot how you actually handle "that's expensive" or "why should I pick you over X." Give it the honest angles: the time saved, the plan that fits a smaller budget, the free tier to start on. "If someone hesitates on price, mention the Free plan and the annual discount, and offer to show the demo so they can judge the value themselves."
Consider Brightline Studio, a small web-design shop that added a bot to their pricing page. At first it just quoted their packages, and most chats ended there. They rewrote the instructions to have it ask about the project's scope and timeline, then recommend a package and offer to book a call. A typical exchange now runs: visitor asks the cost of a five-page site, bot gives a range, asks whether they need e-commerce, learns they do, points to the right package, and offers a free consultation. The chat that used to end at a number now ends with a booked call.
Capture the lead before they wander off
Most visitors won't buy in the same session. If they leave without a way to reach them, the conversation was worth nothing. So the instructions should make the bot offer a reason to share contact details, at the right moment.
Timing is everything here. Asking for an email in the first message reads as a toll booth. Asking after the bot has helped feels natural: "I can email you this quote so you have it handy. What's the best address?" You're trading value for the address, not demanding it. In SpideyChat you'd wire this up by having the bot collect the email as part of a quote or demo flow, so the lead lands in your dashboard instead of vanishing.
Test on real transcripts, not your imagination
You will not guess the right instructions on the first try, and that's fine. The instructions are the one part you can change in minutes.
Read actual conversations weekly and hunt for two patterns. First, where does the bot answer and stop when it should have offered a next step? Add a nudge there. Second, where does it come across as pushy or ask for an email too soon? Soften or move that. Adjust one thing at a time so you can tell what actually moved the numbers.
Keep a small file of your best-performing instruction lines as you find them, along with the phrasings that flopped. Over a few months this becomes your own playbook, tuned to your customers rather than to generic advice. What works for a software product won't match what works for a wedding photographer, and the only way to learn the difference is to watch your own conversations.
A sales bot is never finished, and it shouldn't be. Your products change, your objections change, and the phrases customers use change with them. Treat the instructions as a living document, revisit them every few weeks, and the same traffic you already have starts turning into more booked calls and closed carts.