21 min read

Chatbot Retention Playbook

Five Moments, One Permission Ladder, and a Stop Rule

Quick answer: A chatbot keeps customers at five moments: the order or sign-up confirmation, the first week, the first support contact, the survey, and the lapse. At the first three the customer has just spoken, so the bot may answer on any channel; at the last two the business has to speak first, and on the messaging channels that is where the rules bite. On WhatsApp a first message is an approved template, and if it is promotional it is a marketing template that Meta rations per recipient and, per its own page read 12 September 2026, "does not currently deliver" to US phone numbers (the page does not address the case of an open service window). On Messenger, Manychat's help article of 27 August 2026 describes the replacement for free Recurring Notifications as "a paid solution" billed to an ad account and "currently in beta" in "a limited number of countries" (20 markets including the US, not the EU, UK or Canada, per our Messenger tutorial), while the message tags that used to carry follow-ups "cannot be used for promotional content" and three of them stopped working in 2026. On a website widget the bot cannot speak first; Tidio's return trigger "starts when the visitor returns to your site." In-app and email are where a rule can start a sequence unprompted at no per-message cost, which is what Intercom's Series does with an Entry Rules block, wait blocks and exit rules.

So the retention programme on chat is built in the reverse order from email. Start with the moments where the customer has already spoken and the reply is free, and use them to earn the permissions (an opt-in, an email address, a Messenger List subscription) that the later moments need. Then order the paid and rationed channels by what a first message costs. Then set the stop rule before the first campaign runs, because on a metered channel the customer who does not answer is a block risk, not a lead.

What this guide is and is not

Every channel rule, setting name and vendor quotation is from the vendor's own documentation named in Sources, each read on the date given. Intercom, Tidio, Manychat, SendPulse and Blip have Chatbotscape reviews that carry affiliate links; Meta does not; see our affiliate disclosure. This guide owns the programme: which moments, which channel in which order, which trigger, which metric, and when to stop. It does not own the lapsed-customer message itself and its channel rules, which is the win-back campaign entry; the timed sequence as a mechanism, which is the drip campaign guide; the cart window, which is the abandoned cart recovery playbook; or the survey, which is the NPS guide for the recommend question and the survey design guide for anything longer. No platform is ranked here.

Retention on chat is a permissions problem before it is a copy problem

Email retention playbooks assume the send is free and the only constraint is the unsubscribe link. On the channels a chatbot lives on, that assumption fails in three different ways, and the programme has to be designed around which failure applies.

On WhatsApp the constraint is the customer service window: outside the 24 hours after the customer's last message, the only thing a business can send is an approved message template, priced per delivery by category. A retention message that reminds or informs (your renewal is due, your order shipped) is a utility template; one that promotes (come back, here is 15% off) is a marketing template, and marketing templates carry two further limits Meta documents on its per-user limits page: a per-recipient cap that "adapts automatically over time based on a person's recent engagement levels," and a flat exclusion for "WhatsApp users with United States phone numbers." A recipient's reply "starts a 24-hour customer service window" inside which marketing messages "do not count towards the limit," which is the single most useful fact in the whole programme: the reply is the permission.

On Messenger and Instagram the constraint is the tag regime. Manychat's compliance article, updated 27 August 2026, is blunt that message tags "cannot be used for promotional content. This includes product announcements, event invitations for users who haven't registered, cart abandonment reminders, or notifications about new blog posts," and, in a sentence still in the future tense under its 27 August 2026 stamp, that "Meta is deprecating the Message Tags feature on Messenger on Monday, February 9, 2026." Our Messenger chatbot tutorial follows Meta's own changelog instead, under which the three tags return error code 100 from 27 April 2026; the two sources give different dates and this guide does not reconcile them, but by either reading the tags were gone months before this page was written. What replaces them for promotional follow-up is Marketing Messages, which the same vendor's product article describes as "built on top of Meta's paid advertising capabilities," charged "through the business's credit card connected to the Ad account(s)," with campaigns on a daily or lifetime budget that "run for 30 days," and reachable only by contacts on an opted-in Messenger List. The tutorial lists the 20 markets it runs in (the US, Brazil, Mexico and India among them; the EU, the UK and Canada excluded) and Meta's caps: one marketing message per unique user per day, a 12-hour cooldown per subscription token, and frequency caps enforced since 10 April 2026 that block sending when marketing threads exceed 20 percent of organic threads over seven days or 40 percent over one day. Recurring Notifications, the free predecessor, was retired in early 2026 except for existing integrations in the EU, UK, Australia, Japan and Korea. Instagram is narrower still: "in February 2024, Meta paused allowlisting for the Instagram DM List Beta," and outside the 24-hour window only a human agent may write, for up to 7 days.

On a website widget the constraint is physical: there is nobody to message. The bot is a set of triggers, and the retention trigger fires only when the customer is back on the page. Tidio's list, dated 19 November 2025, has two that matter here: Visitor returns to the site ("starts when the visitor returns to your site (is not a new visitor)") and The visitor hasn't contacted you for some time, which counts "a specified time (counted in days)" and "will work only if the visitor has left you an email address." That last clause is the whole reason the email capture guide matters to retention: without the address, the widget has no way to reach a lapsed customer between visits.

In-app messaging and email are the exception. Intercom's Series lets a rule start a sequence: an Entry Rules block holding "the filters a customer must match to enter the series," content and wait blocks ("Send an email → Wait for 3 days → Send a follow up chat message"), a goal with a time limit after which a match "won't be counted," and exit rules under which a customer who matches "will exit the series completely, and cannot re-enter." That is the email model, and it is available only where the customer is logged in to your product or has given an address.

The five moments

Each moment is defined by who has just spoken, because that decides which channel rules apply.

Moment	Who spoke last	What the bot may send, by channel	What retention work it does
1. Confirmation (order, booking, sign-up)	The customer, seconds ago	Anything, on any channel, for 24 hours on WhatsApp and Messenger; a utility template after that	Collects the permissions the later moments need: the opt-in to marketing messages, the email address, the Messenger List subscription. Ends with a question so the window stays open
2. First week	The customer, days ago	WhatsApp: utility templates for what they expect (delivery, activation, how-to); Messenger: nothing promotional; web: return trigger; in-app: a Series	The onboarding flow and, for software, the SaaS onboarding guide; the aim is one completed action, not a tour
3. First support contact	The customer, now	Anything, in the open window	Resolution first; then the one question that identifies a cooling customer ("did that fix it?") with buttons. The handoff design guide covers what the agent inherits
4. Survey	The business, by rule	WhatsApp: a feedback survey tied to a specific transaction is a utility template (the template entry has Meta's categorization rule); web and in-app: a trigger or Series; Messenger: only inside a window or via a paid marketing message	The NPS guide and CSAT entry cover timing; here the point is that a detractor's reply opens a window, and the follow-up inside it is the retention message
5. Lapse	Nobody, for longer than the customer's cycle	WhatsApp: a marketing template, rationed and not delivered to US numbers; Messenger: a paid Marketing Message to an opted-in list; web: return trigger; in-app or email: a Series entered by inactivity	The win-back campaign: one reason, a cheap answer, a stop

The order is the point. Moments 1 to 3 are free and the customer is present; a programme that skips them and starts at moment 5 has to pay for every message and has no permissions to spend. Our editorial rule, and it is a rule rather than a finding, is that a retention programme is not allowed a lapse campaign until moment 1 has been collecting opt-ins for at least one full customer cycle.

The permission ladder

When the business has to speak first (moments 4 and 5), order the channels by what a first message costs, and start at the bottom.

A window the customer opened. Free on every channel. If a customer wrote yesterday, the retention message goes in today, before the window closes. Moment 3 is the richest source: a support conversation that ends "did that fix it?" and gets a "no" is a retention conversation already in progress.
A utility template the customer expects. On WhatsApp, a delivery note, a renewal reminder or a booking confirmation is a utility template, priced at a fraction of the marketing rate in most markets (the WhatsApp channel guide carries the per-market rates; the template entry has the category rules) and outside the per-user marketing cap. It may end with a question. It may not carry an offer, or Meta's categorization moves it to marketing; the appointment booking guide works through that reprice risk for reminders.
A return trigger. Free, but the customer has to come to you. Tidio's return trigger and the days-since-contact trigger; Intercom's in-app messages on next login.
Email, where you have the address. Free to send, subject only to consent and the unsubscribe link, which is why moment 1 collects it. Intercom's Series and SendPulse's bundled email are the corpus examples; the SendPulse review recorded one subscription covering "email broadcasts, SMS, web push, CRM" and chatbots, which is the shape a small business needs for this rung.
A WhatsApp marketing template. Paid per delivery, rationed per recipient, not delivered to US numbers, and each undelivered attempt (error 131049) is a signal to wait "at least 24 hours" and, on our rule, to stop for that recipient. The WhatsApp broadcast guide covers what a badly targeted marketing send does to a quality rating.
A Messenger Marketing Message. Paid through an ad account, opted-in Messenger List only, and capped by Meta at one marketing message per user per day with the frequency limits above. Available in 20 markets including the US, which makes it the moment-5 channel for a US business that WhatsApp is not; excluded in the EU, UK and Canada. Last on the ladder because the permission is the most expensive to acquire (a list opt-in plus an ad account with a payment method), not because it is rare where it exists.

The ladder is why the same customer might get a utility reminder on WhatsApp, an in-app message on next login, and a marketing template only if both were ignored: each rung is tried before the next costs more. The lead nurturing automation guide works through the same economics for the pre-purchase case, and the drip campaign guide sets out where a sequence stops.

What the platforms expose

Platform	The retention primitive	What the documentation says (read 12 September 2026)	Chatbotscape review
Intercom	Series: entry rules, wait blocks, rule blocks, goal, exit rules	Entry rules are "the set of filters a customer must match"; a rule block can "Try to match for 3d"; an event rule not triggered "within 30 days of reaching the block" sends the customer down the "did not match" path; customers "can re-enter the same Series multiple times if you enable 'Re-enter Series'"; exit is by rule, path end, goal, or "don't engage with messages for a set time"	The Intercom review captured a template library organized "announcement / NPS survey / onboarding / re-engagement" and a "churn risk" customer event among workflow triggers
Tidio	Flow triggers	Visitor returns to the site; The visitor hasn't contacted you for some time (days; "only if the visitor has left you an email address"); Agent marks conversation as solved (the moment-3 hook); a per-trigger option to "decide how many times they should be fired per person"	Tidio review
Manychat	Messenger Lists plus Marketing campaigns; Broadcasts and Sequences on Pro	Marketing Messages are selected as a campaign inside a "Send a message" step or as a new broadcast type; "an estimated reach based on the campaign budget" is shown; campaigns "are reviewed and approved by Meta" with "a 10-minute waiting period" after approval	The Manychat review recorded "Broadcasts target segments with one-off messages" under Pro Tools
SendPulse	Bundled email, SMS, web push and chatbot from one subscription	Not re-read for this guide; the bundling claim is from our review	SendPulse review
Blip	A 24-hour recurrence rate on the dashboard	Not re-read for this guide	The Blip review recorded Taxa de recorrência (contacts who interacted twice or more in 24 hours) as a metric "Manychat and Tidio do not surface in their default dashboards"

None of the fifteen reviewed platforms was tested on a retention sequence end to end; the table records what the documentation and our dated screenshots state, not a measurement.

The four metrics and the holdout

The chatbot metrics guide has the definitions; these are the four a retention programme reports, in the order they become available.

Permission rate. Of customers who reached moment 1, how many left an opt-in, an address or a list subscription. This is the leading indicator, because every later rung depends on it, and it is the one number a programme can move in its first week.

Reply rate. Of business-initiated messages at moments 4 and 5, how many got any reply. On WhatsApp and Messenger the reply is what opens the window and stops the meter, so this comes before conversion.

Retention rate. Customers active at the end of a period who were also active at the start, divided by customers active at the start, with new customers excluded from both. Compute it per cohort (customers who reached moment 1 in the same month) or it hides the programme's effect behind acquisition.

Block and opt-out rate. The cost, and on WhatsApp the number that feeds the quality rating. A retention programme whose block rate rises is buying churn.

Report the lapse campaign against a holdout: a random slice of lapsed customers who receive nothing, compared over the same window. Without it, the campaign takes credit for every customer who would have come back anyway, which for a seasonal business is most of them.

The stop rule

Set it before the first message goes out, because on a metered channel the customer who does not answer is a block risk, not a lead.

Two business-initiated attempts per moment, a fortnight or more apart for a lapse, a day or more for a survey chase, and then nothing until the customer moves. On WhatsApp a 131049 error is a stop for that recipient, not a retry. On Messenger an opt-out from a Messenger List removes the permission entirely, and Meta's own caps (one marketing message per user per day; sending blocked when marketing threads pass 20 percent of organic threads over seven days or 40 percent over one day, per the changelog our Messenger tutorial cites) mean a retention programme that leans on rung 6 throttles itself before the block rate does. In Intercom, put the stop in as an exit rule before the campaign is live; the help article's own warning is to "add the exit rule to your Series before tagging customers with the exclusion tag," because customers tagged first "may receive messages in the window between when the tag is applied and when the exit rule is saved." In Tidio, the per-person trigger limit does the same job for the return trigger. A survey detractor who has replied is inside a window and is not counted against the two attempts; the reply reset the ladder.

What our fifteen reviews record

Searched 12 September 2026 across the fifteen files matched by sample-reviews/*-review.md, case-insensitively: win-back|winback|win back appears in 0; re-engag in 4, of which three (AiSensy, Landbot, Intercom) mean the bot resuming after a human agent and one (Blip) is the recurrence metric above; retention in 3, all meaning data retention; churn in 2 (Chatfuel, on its own user base; Intercom, the "churn risk" event). The reviews' six-scenario protocol times a FAQ bot, a template approval and a handoff; none sends a business-initiated message to a lapsed contact and measures the reply. That is a gap in the protocol, recorded here, not a finding about the platforms. Corrections to editorial@chatbotscape.com.

Where it breaks

The programme cannot fix three things. A product that customers leave because it did not work is not retained by a message; the resolution rate guide is the prior work, and moment 3 is where the signal shows first. A US customer base on WhatsApp has no moment-5 channel at all under the current Meta rule, and the honest playbook for that case is email, SMS (the SMS chatbot setup guide covers consent and STOP handling) and, where the list opt-in exists, a Messenger Marketing Message. And a programme that measures itself without a holdout will report success every quarter until the block rate ends it.

FAQ

What is customer retention automation for a chatbot?

A set of rules that send, or offer, the right message at the five moments a customer can be kept (confirmation, first week, first support contact, survey, lapse), on the channel that is allowed to carry it. On WhatsApp, Messenger and web chat the automation is constrained by who spoke last: a customer's message opens a window in which the bot may say anything; outside it, the business needs a template, a paid marketing message, an opt-in, or an address.

Can a chatbot win back lapsed customers on WhatsApp?

Only with an approved marketing template, which Meta rations per recipient, withholds with error 131049 when a person has reached their limit, and does not deliver to US phone numbers (all read 12 September 2026). The cheaper route is a utility template the customer expects, ending with a question; a reply opens a 24-hour window in which the rest of the sequence is free of those rules.

What replaced Recurring Notifications on Messenger?

Marketing Messages, which Manychat's article of 27 August 2026 describes as paid through the business's ad account, in beta in a limited number of countries (20 markets including the US, excluding the EU, UK and Canada, per our Messenger tutorial), and sendable only to contacts on an opted-in Messenger List, capped at one per user per day. Recurring Notifications was retired in early 2026 (Manychat dates the start of the sunset to 12 January 2026) except for existing integrations in the EU, UK, Australia, Japan and Korea. Message tags were never allowed to carry promotion and three of them stopped working in 2026.

How many retention messages should a chatbot send?

Two business-initiated attempts per moment, then stop until the customer moves. A reply resets the count, because it opens a window. On WhatsApp a withheld send is a stop, not a retry; in Intercom the stop is an exit rule set before the Series goes live; in Tidio it is the per-person trigger limit.

How do I measure whether the programme works?

Permission rate at moment 1, reply rate at moments 4 and 5, retention rate per cohort with new customers excluded, and block or opt-out rate as the cost. Run the lapse campaign against a holdout group that receives nothing, or the campaign takes credit for customers who would have returned anyway.

Does this work for a US business?

Moments 1 to 4 do. Moment 5 on WhatsApp does not, because Meta does not deliver marketing templates to US numbers; use email, SMS, in-app or, for contacts on an opted-in Messenger List, a Messenger Marketing Message (the US is one of its 20 markets) for the lapse, and use WhatsApp for the utility messages customers expect.

Win-back campaign (glossary) — the lapsed-customer message and the channel rules that constrain it.
Drip campaign guide — the timed sequence as a mechanism and where it stops.
Lead nurturing automation guide — the same channel economics, before the purchase.
Abandoned cart recovery playbook — the short-window case.
Chatbot NPS guide — moment 4 in detail.
Chatbot survey design — moment 4 when the survey has more than one question.
Chatbot onboarding guide — moment 2 in detail.
WhatsApp broadcast campaign guide — what a marketing send does to a quality rating.
Chatbot metrics guide — the definitions behind the four metrics.

Sources

Meta for Developers, WhatsApp Business Platform, Per-user marketing template message limits (developers.facebook.com/documentation/business-messaging/whatsapp/templates/marketing-templates/per-user-limits, undated on the page), read 12 September 2026: "WhatsApp may limit the number of marketing template messages a WhatsApp user receives from any business in a given period of time"; "adapts automatically over time based on a person's recent engagement levels"; "WhatsApp does not currently deliver marketing template messages to WhatsApp users with United States phone numbers"; the EEA, UK, Japan and South Korea exclusion; "If a WhatsApp user responds to a marketing message, it starts a 24-hour customer service window. Marketing messages sent within this window do not count towards the limit"; "wait at least 24 hours before attempting to resend"; error code 131049.
Manychat Help, Marketing Messages on Messenger (article 24351480518684, updated 27 August 2026), read 12 September 2026: "built on top of Meta's paid advertising capabilities"; "Unlike Recurring Notifications, which Meta will begin sunsetting globally on January 12, 2026, Marketing Messages on Messenger is a paid solution. Meta charges for its usage through the business's credit card connected to the Ad account(s)"; "currently in beta and has been rolled out by Meta in a limited number of countries"; "Both options apply to campaigns that run for 30 days"; "reviewed and approved by Meta. After approval, there is currently a 10-minute waiting period"; "an estimated reach based on the campaign budget."
Manychat Help, How to send messages outside the 24-hour and 7-day windows in Messenger and Instagram (article 14281199732892, updated 27 August 2026), read 12 September 2026: "Meta is deprecating the Message Tags feature on Messenger on Monday, February 9, 2026"; "Message tags cannot be used for promotional content. This includes product announcements, event invitations for users who haven't registered, cart abandonment reminders, or notifications about new blog posts"; the 7-day Human Agent window ("Automated messages are not allowed during the 7-day window"); "starting August 10, 2023, notification tokens no longer expire"; "in February 2024, Meta paused allowlisting for the Instagram DM List Beta."
Tidio Help Center, Flow editor: Triggers (article 5453654525212, dated 19 November 2025), read 12 September 2026: the trigger list, including "Visitor returns to the site: The flow starts when the visitor returns to your site (is not a new visitor)"; "The visitor hasn't contacted you for some time: The flow starts when the visitor hasn't contacted you back for a specified time (counted in days). The flow will work only if the visitor has left you an email address"; "Agent marks conversation as solved"; "decide how many times they should be fired per person."
Chatbotscape, Messenger chatbot tutorial (/academy/messenger-chatbot-tutorial, sources read 25 August 2026): the Marketing Messages 20-market list and its EU, UK and Canada exclusion; Meta's per-user cap, 12-hour token cooldown and the 10 April 2026 frequency caps (from Meta's Messenger Platform changelog); the 27 April 2026 error-code-100 date for the three retired tags; and the Recurring Notifications carve-out for the EU, UK, Australia, Japan and Korea. Not re-read from Meta for this guide; cited to the tutorial and its dates.
Intercom Help, Series explained (article 4425207; the page carries a relative stamp, "Updated today," so its revision date is 12 September 2026), read 12 September 2026: "Series lets you create automated, multi-step messaging campaigns using Intercom's Outbound channels"; entry rules as "the set of filters a customer must match to enter the series"; "Try to match for 3d"; "If the customer doesn't trigger the event within 30 days of reaching the block, they'll go down the 'did not match' pathway or disengage"; "Send an email → Wait for 3 days → Send a follow up chat message"; "Customers can re-enter the same Series multiple times if you enable 'Re-enter Series'"; "they will exit the series completely, and cannot re-enter"; "Add the exit rule to your Series before tagging customers with the exclusion tag ... they may receive messages in the window between when the tag is applied and when the exit rule is saved"; "Customers can also exit a Series if they reach the end of a path, hit the series goal or don't engage with messages for a set time."
Chatbotscape review corpus (the fifteen platform reviews at /reviews), searched 12 September 2026 from the repository root. Denominator: ls sample-reviews/*-review.md | wc -l returns 15. grep -liE 'win-back|winback|win back' sample-reviews/*-review.md returns 0; grep -liE 're-?engag' sample-reviews/*-review.md returns 4 (aisensy, blip, intercom, landbot); grep -liE '\bretention\b' sample-reviews/*-review.md returns 3 (aisensy, botpress, tars); grep -liE '\bchurn' sample-reviews/*-review.md returns 2 (chatfuel, intercom). Passages cited: intercom-review.md lines 195 and 633; blip-review.md line 646; manychat-review.md line 251; sendpulse-review.md line 127.
Ahrefs Keywords Explorer, US overview, queried 12 September 2026: the figures in this guide's keyword note.
Chatbotscape evaluation methodology. /methodology (continuously updated).

About this guide

Chatbotscape launched in 2026 as an independent review site for chatbot platforms. This guide is part of our SMB chatbot Academy and is written for the owner or marketing lead of a small business who is answerable for whether customers come back, and who is choosing which channel a chatbot should use to help. It reads the retention primitives of four vendors (Intercom, Tidio, Manychat and Meta's WhatsApp and Messenger platforms) as worked examples and carries two observations from our own reviews (SendPulse, Blip); it does not rank them. Where a platform has a Chatbotscape review, the link goes to it and the review carries the affiliate disclosure.

Methodology

The Meta, Manychat, Tidio and Intercom pages named in Sources were read in full on 12 September 2026, and every rule and setting name carries its source and date; the SendPulse, Blip, Manychat and Intercom review observations come from our own dated review sessions and are cited by line. The corpus counts are ours and the commands are printed. The editorial judgments on this page, and they are judgments rather than measurements: the five-moment split, the permission ladder and its order, the "one full cycle before a lapse campaign" rule, the two-attempt stop rule, the cohort definition of retention rate and the insistence on a holdout. See /methodology.

Last updated

13 September 2026.