Every chat widget waits. Someone arrives, hesitates, leaves, and the bubble sat in the corner saying nothing. A proactive agent does the opposite: it spots a specific moment of hesitation and opens the conversation. Whether that works or turns into an annoying pop-up depends entirely on which moment you pick.
A proactive trigger only converts when the message answers the question the behaviour revealed. Time on the pricing page calls for a plan comparison; exit intent calls for an open question; a stalled cart calls for removing an objection. The generic trigger ("can I help?") has the worst rate of all because it answers nothing.
1. Dwell time on the pricing page
This is the highest-return trigger, for a simple reason: someone who has spent ninety seconds on the pricing page is not reading, they are calculating. The question is almost always the same, which plan do I land in, and what happens when I grow.
The message that works names the question instead of offering generic help: "want me to compare Growth and Enterprise for your volume?". It converts because the person does not have to formulate the question, only accept it.
2. Exit intent with context
Exit intent has a bad reputation because it spent a decade pushing coupons in full-screen modals. The trigger is not the problem; the message is. An agent that detects the exit and asks "before you go: is there anything I did not answer?" recovers the conversation without promising a discount that eats margin.
The practical rule: exit intent should only fire if the person already showed intent. On someone who arrived ten seconds ago it is an interruption. On someone who read three pages it is service.
3. A stalled cart with hesitation between options
In e-commerce, the classic abandoned cart is handled by email hours later. The real moment of loss is earlier: the item is in the cart and the person is flipping between two product pages. The question there is comparative (size, compatibility, delivery) and an immediate answer settles it.
This trigger depends on the agent knowing the catalogue, not just the site. That is the difference between "can I help with your order?" and "both models fit; the second one delivers in two days in your area".
4. Returning visitor who has not converted
A third visit without a purchase is the most underrated signal in the funnel. The person has already decided they need the solution and is stuck on one specific detail. An agent that recognises the return and asks what is missing usually gets the real objection stated plainly: which is exactly the data a sales team never has.
The message here should acknowledge the history without sounding like surveillance. "You have been here before: anything I can clear up?" works. Listing the pages they visited does not.
5. Documentation page with no exit
In technical products, someone who spends a long time in the docs and does not move on has usually hit a prerequisite. This is a retention trigger, not an acquisition one, and it gets ignored because it does not show up in the sales funnel, but it is where churn starts.
The three triggers that only annoy
Worth recording the inverse, because getting this wrong costs more than doing nothing:
- Firing on absolute time on the home page. Time on the home page is not intent; it means the person is orienting themselves.
- A message on the first scroll. It interrupts before there is enough context for any useful answer.
- Repeating the trigger in the same session after it was dismissed. A refusal is an answer; insisting turns the agent into a pop-up.
4
How to calibrate without guessing
Run every trigger switched off for two weeks and record only the behaviours. Then turn them on one at a time, with a specific message, and compare the response rate against that baseline. A trigger below the average acceptance rate after two hundred impressions should be switched off, not rewritten.
The common mistake is turning all five on the first day. You lose the ability to know which one worked and you hand the visitor a noisy experience. If you want to see the triggers running before configuring anything, they are described in features, and which ones to pick changes a lot by industry.
