The intuition behind proactive chat is simple. A visitor on a pricing page who has scrolled to the bottom, hovered the "Pro" tier twice, and is now staring at the screen is probably one nudge away from buying. A reactive chat widget — the green dot in the corner — has no idea any of that is happening. A good salesperson watching the same person would step in. Proactive chat is the software version of that.
The intuition is right. The execution is where most teams fall down.
Baymard Institute, 2026
What proactive chat is, and isn't
Proactive chat is system-initiated. The system decides to open the chat surface, send a greeting, or surface an offer. The visitor did not click anything to get there.
Reactive chat is visitor-initiated. The visitor clicks the chat icon and waits for a reply. Most "live chat" products from 2010-2020 were primarily reactive.
Proactive chat is not a popup. A popup is a static template that fires on one condition (typically time-on-page or exit intent). Proactive chat picks the message based on the current behavior state. A popup can be proactive in spirit; calling it "proactive chat" is misleading.
Proactive chat is not the same as an AI sales agent. An AI sales agent typically acts asynchronously — it sends an email after the cart is abandoned, it follows up on a demo request the next morning. Proactive chat acts synchronously, in-session, while the visitor is still on the page.
The behavior triggers that actually work
Across the industry, the list of signals that reliably predict conversion is short.
The pattern holds up well. Triggers anchored in cart state are reliable. Triggers anchored in time-on-page alone are noisy. Triggers on category and home pages are barely worth firing — the lift is real but small, and the cost in annoyance and trained dismissal rarely justifies it.
How to time it without annoying people
These principles are a good starting point. Tune them against your own traffic.
Wait for a real signal. Nothing should fire until a visitor has shown genuine interest or hesitation — reading carefully, comparing options, returning for a second visit, or changing what is in their cart. A greeting on arrival is not proactive; it is a popup in costume.
Throttle hard. Limit how often the agent can speak, per visitor and per session. Someone who waves off a couple of offers has told you they want to be left alone. Listen.
Give it a cooldown. After a dismissal, stay quiet for a while before trying anything else. Too short and you nag; too long and you miss the moment. Err toward patience.
Never re-greet a visitor who already converted. The job is done. A "thanks for ordering" agent that keeps talking only teaches people to close it.
Filter out bots before deciding anything. Automated traffic should never trigger an intervention; it wastes money and pollutes your numbers.
And keep a holdout. Hold back a slice of traffic that sees nothing, and measure the difference against it. It is the only honest way to know the agent is working — without it, seasonality and regression to the mean will fool you inside a couple of weeks.
What does proactive chat cost to run?
Proactive chat runs on someone else's model API, so the cost comes down to how often you call that model. The expensive way is to route every visitor and every signal through a frontier model and pay for all of it. The cheaper, saner way is to let a lightweight layer decide whether and when to step in, and only call the model when there is actually a message to write.
That difference is the whole ballgame on cost. A cost-controlled, hybrid approach keeps the per-visit cost small enough that the agent can pay for itself on a single recovered cart, while still sounding human when it does speak. For plan limits, see pricing.
How to roll proactive chat out
Not all at once. Roll it out in phases, and keep part of your traffic aside as a holdout the whole time so you can prove it is working.
Start with cart abandonment. One trigger, one offer: when someone with items in their cart looks like they are leaving, offer a small discount they can apply in a tap. It is the highest-intent moment on the site and the easiest win to measure. Give it a week or so before you read the result.
Then add pricing and product-page help. Once the first trigger is paying off, add a comparison card for visitors weighing two options, and a "common questions" prompt on product pages. Keep the holdout going.
Then add return visits and lead capture. Welcome back a returning visitor who still has items in their cart, and offer an email capture on long-form pages where that is the real conversion goal.
By the end, you have a holdout-measured lift you can trust, the agent has been live for the whole audience through at least one phase, and you have a rhythm for reviewing what it says.
Where proactive chat goes wrong
Three failure modes that are hard to spot until they have done damage.
1. The "everything chatbot" failure. The agent has been configured to handle every kind of question — pricing, returns, shipping, tracking, account help, technical support. CVR is fine, but visitor satisfaction has tanked. The fix: scope the agent narrowly. Pre-purchase questions only. Returns and account help should route to email or a real support flow.
2. The "ghost greeter" failure. The agent fires a greeting on every page, on every session, unconditionally. Visitors learn to dismiss the agent. The fix: wait for a real signal before saying anything.
3. The "dead handoff" failure. The agent collects a lead, says "Someone from our team will be in touch", and then nothing happens for days. The lead goes cold, the visitor loses trust, and the next visit is harder, not easier. The fix: connect lead capture to a real follow-up flow (email automation, a sales-rep alert) before you turn the agent on.
Where is proactive chat heading?
Two trends are worth watching.
The cost of always-on model calls keeps falling. As it does, the price gap between calling a model on every event and calling it only when needed will narrow — though the speed advantage of a fast, local decision will not go away.
Behavior signals are getting richer. Client-side machine learning is moving from experiment to production, and signals that are lab-stage today (webcam eye-tracking, gesture, tone of voice) will become usable over the next couple of years. The category will absorb them as new triggers, not as new categories.
The discipline does not change. Whatever new signals arrive, the principles hold: wait for a real signal, throttle hard, keep a holdout, and never re-greet someone who already converted.
Further reading
Frequently asked questions
Proactive chat is any chat intervention initiated by the system rather than by the visitor — the system decides to open the chat, send a greeting, or surface a coupon. The opposite is reactive chat, which only opens when the visitor clicks the chat icon.
Last updated May 31, 2026.