Chatbot Fallback Copywriting

What the Bot Says When It Does Not Understand

Quick answer: The fallback message is the line a chatbot sends when it cannot place what the customer said. It is the most-read line the bot has, because it fires on every miss, and the least-written, because it ships with whatever the vendor typed into the default, or with nothing at all. The wordings the vendors put in front of you are variations on one sentence: Rasa Open Source's documentation example for utter_default (read 10 September 2026) is "Sorry I didn't get that. Can you rephrase?", Rasa Pro's shipped default utter_ask_rephrase is "I'm sorry I am unable to understand you, could you please rephrase?", Landbot's validation retry, as captured in our May 2026 session, is "I'm afraid I didn't understand, could you try again, please?", and Tidio's Lyro, in a handoff situation where it is set to keep the conversation, says "I'm sorry, the team is unavailable, so I can't connect you. 😔 Please try asking the question differently so I can help you better." Three of the four open with an apology and the other with a hedge, all four ask the customer to do the work again, and none of them says what the bot can do instead.

This guide is the rewrite. A fallback message has four jobs: own the miss without groveling, narrow the scope to what the bot handles, change the mechanism so the customer taps instead of retyping, and leave an exit to a person or a page. Six situations need six different messages, because "I don't know" after silence, after a wrong-format answer, after a question outside the bot's scope and after an integration timeout are different facts. And the message changes on each attempt: the first narrows, the second switches to buttons, the third stops asking. What follows is the copy for each rung and each situation, the vendor defaults read against it, the channel limits that decide how many characters you get, and a ten-message test to run before the bot goes live.

Why the fallback is the line to write first

Two numbers frame it. Our fallback-rate entry puts the healthy range for a tuned intent bot under 10 percent of messages and says to plan a training review between 10 and 15. Take a bot at 12 percent handling 2,000 conversations a month with four customer messages each: that is roughly 960 fallback messages sent a month, second only to the greeting in how often it is read, and the greeting fires once per conversation while the fallback fires on every miss. The welcome message gets a design session, the welcome-message guide and a review; the fallback gets the default.

The second number is what the customer does next. Intercom's escalation documentation for Fin lists "Loop: The customer repeats themselves across three turns" as an event on which Fin offers a human, and adds that "Fin will never offer escalation twice in a row. If an offer to escalate is triggered again immediately after a previous offer, Fin will escalate the conversation straight away to prevent looping." That is a vendor building the exit into the product because the loop is what customers do when the fallback gives them nothing to act on. The copy's job is to give them something before the third turn.

The four jobs of a fallback message

Every recovery message, whichever situation produced it, has to do four things. The order matters on a phone screen: the exit comes last, but it has to be there.

Own the miss. One clause, no theater. "I didn't catch that" or "That one's outside what I can help with" is enough. The vendor defaults over-apologize (Rasa, Rasa Pro and Tidio open with "Sorry" or "I'm sorry"; Landbot hedges with "I'm afraid") and the apology is dead weight in a message that has twenty characters per button. Do not blame the input ("Your message wasn't clear") and do not confess to being a bot as if that explained anything; the customer knows.

Narrow the scope. Say what the bot does handle, in the customer's words. This is the job every default skips, and it is the one Manychat's documentation gets closest to: its Default Reply tips are to "include contact information in your Default reply" and to "incorporate Buttons or Quick replies to guide users in the right direction if they get stuck." The scope is the list from your content strategy: the three to five things the bot answers, looks up or does.

Change the mechanism. If the customer typed and the bot missed, asking them to type again with different words is asking them to guess the bot's vocabulary. Offer the scope as tappable options instead. Rasa Pro's own clarification pattern does this with a list ("I can help, but I need more information. Which of these would you like to do: {{context.clarification_options}}?"), and every messaging channel we cover has quick replies or buttons for it; the buttons vs free text guide is the per-question rule for which questions should have offered them in the first place.

Leave an exit. A person, a phone number, a page. Not on every message, but by the second attempt at the latest and on the first for anything the bot cannot ever answer. Tidio's documentation makes the exit a setting: if Lyro "is allowed to transfer conversations to live agents during online hours, Lyro will offer the Talk to a person option," and if it is not, "it will ask the visitor to paraphrase their question, or simply explain that it lacks the necessary knowledge." The second configuration is a bot with no exit, and it is the one many small businesses end up with because the inbox is unstaffed.

A message that does all four in one line, for a shop bot on a website widget:

I didn't catch that. I can help with order status, returns, or store hours — tap one, or type "person" to reach the team.

Twenty-three words. On WhatsApp the three options become reply buttons and the text loses the list.

Six situations, six messages

The single biggest error in fallback copy is one message for six different facts. The bot knows which happened; the copy should say so. Two of the six (the invalid answer and the silence) are, strictly, reprompts, because a question is still open; they are in the table because the same default text usually serves all six.

Two of these deserve a note. The out-of-scope message is the one the vendors handle worst by default: Rasa Open Source's documentation example for utter_out_of_scope is "Sorry, I can't handle that request." and Rasa Pro's shipped utter_cannot_handle is "I'm sorry, I'm not trained to help with that." Both are honest and both are dead ends; the rewrite keeps the honesty and adds the exit. Intercom's Fin documentation names the underlying split: the questions Fin cannot answer are content gaps, customer-data gaps ("such as order status or account details") and action gaps ("like updating a workflow or canceling an order"), and the FAQ chatbot entry covers why the last two cannot be fixed with more pairs. A fallback that asks the customer to rephrase "where is my order" is asking them to rephrase a question the bot has no way to answer in any phrasing.

The low-confidence message is a different animal from the other five and is the one most flow builders cannot send, because it needs a confidence score and a second-best intent. Rasa's two-stage fallback generates it ("Did you mean '{intent}'?" with Yes and No buttons, a wording read from the project's source code rather than its docs); Dialogflow CX and the confidence-policy guide cover setting the band it fires in. If your builder has no confidence surface, skip the row; a guessed "did you mean" with no score behind it is worse than the scope list.

The three-attempt ladder

The message also changes with the count, and the counters exist for exactly this reason. Google's Dialogflow CX exposes sys.no-match-1 through sys.no-match-6 and the same for no-input, "invoked instead of sys.no-match-default and in order." Voiceflow's global no-reply setting counts follow-ups per turn ("default 5") before ending the session. Our chatbot scripts guide sets the shape (narrow, switch mechanism, stop) and notes that none of our fifteen reviews records a numbered retry primitive (Voiceflow's documentation exposes a no-reply counter our review did not capture), so in a flow builder such as Manychat, SendPulse, Wati or Landbot, assume you count with a variable. This is the copy for each rung.

Attempt one: narrow. Shorter than whatever the customer was answering, with the scope or the format in it. No apology beyond one clause.

Not sure what you need — is it an order, a return, or store hours?

Attempt two: switch. Stop asking for text. Buttons for the scope, and the exit appears as one of them. The wording acknowledges this is the second miss without counting it out loud.

Let me make this easier. Tap what you're after: [Track an order] [Start a return] [Talk to a person]

Attempt three: stop asking. No third question. The bot states what it is doing and does it: handoff with the transcript, a ticket, or the contact line if nobody is available. Rasa's documentation calls this the "ultimate fallback action" and gives the handoff as the example; Intercom's Fin, when it "is unable to find an answer to a new user query," routes the conversation "automatically" to the team.

I'll get a person on this. I've passed along everything you've written — someone will reply here, usually within [your real number].

The fallback-rate guide describes the same ladder in two steps, options on the first miss and a person on the second; it collapses rungs one and two, and the shape is the same. Three rules across the rungs. Each message is new: a customer who reads the same line twice has learned that nothing they do changes the outcome. The count resets when a message matches: two misses, one hit and another miss is attempt one again, not attempt three, or the bot hands off customers who were making progress. And the ladder is per situation: an out-of-scope question goes straight to the rung-three message on attempt one, because narrowing cannot help.

The vendor defaults, read against the four jobs

The defaults are worth reading because they are what most bots say. Rasa, Tidio, Intercom, Manychat and Voiceflow were read on 9 and 10 September 2026; the Landbot and Chatfuel rows are the fields as our reviews captured them on 29 and 30 May 2026, and Rasa Open Source's wordings are the examples its documentation tells you to define, not text the engine ships.

The pattern is plain: most of the wordings meant for a miss (Rasa's documentation examples, Rasa Pro's rephrase line, Tidio's keep-conversation answer, the Landbot field as we found it) apologize and ask for a rephrase, and the builders leave the wording to you (Manychat and Voiceflow ship no text; Chatfuel gives you a field whose pre-filled value, if any, our review did not record), which in practice means the first thing you typed at 11 p.m. on launch night. Tidio's default is the most instructive because the emoji and the sentence are doing opposite things: the sad face signals that the bot cannot help, and the second sentence asks the customer to try harder anyway. Tidio's own help center notes that the predefined messages "are automatically translated for any users who message you and require a language that's different from your default," which is one more reason to write them plainly; the localization guide covers what survives machine translation.

Two of the defaults are worth keeping. Rasa Pro's clarification pattern is a rung-two message as shipped. And Tidio's online-transfer line ("Absolutely! I'm transferring you to a human right away") does the one job a transfer message has, which is to say the transfer is happening; the handoff design guide covers what to add about wait time.

Generative bots: the fallback is a prompt rule, not a text field

On an AI agent built on a language model, there is often no fallback field to edit. Voiceflow's global no-match offers "Generative - provide a prompt that tells the agent how to handle the situation" next to "Scripted - write a fixed response that's sent every time." Intercom's Fin has no editable cannot-answer line; escalation guidance, written in natural language, "influences Fin's replies and send-off messages," and its documentation's own examples are guidance strings such as "Offer to route the customer to a human agent if they ask the same question twice." Tidio's Lyro combines the two: fixed predefined answers for the handoff cases and free-text Guidance for when to hand off ("Escalate to a human agent immediately if the customer is angry or annoyed").

The four jobs translate into prompt rules, and the prompt-engineering guide covers where they go in the system prompt. Written out:

When you cannot answer from the knowledge base, say so in one sentence without apologizing more than once. Name the three things you can help with: order status, returns, store hours. Offer them as buttons. On the second consecutive miss, offer a person. Never ask the customer to rephrase a question about their specific order, account or payment; those go to a person immediately.

Two cautions. A generative fallback can invent an answer instead of admitting the gap, which is the failure the hallucination guide is about, and the "say so in one sentence" rule is the guard. And a generative rung-three message still needs a deterministic action behind it: the model can say "I'll get a person" only if the workflow actually routes, which on Intercom is a workflow step ("Step 6: If Fin can't resolve the conversation") and on Voiceflow a path, not a sentence.

Channel limits decide how many words you get

The mechanism-switch rung depends on buttons, and buttons have caps. The figures below are the ones our chatbot scripts guide read from Meta's live WhatsApp documentation on 12 August 2026; they are repeated here because they decide the copy, and that guide carries the WhatsApp character table.

WhatsApp reply buttons: three per message, 20 characters each, so "Talk to a person" (16) fits and "Speak with a support agent" (26) does not. The body of an interactive message is capped at 1,024 characters, which is not the constraint; the 20-character button label is. Messenger and Instagram quick replies are capped at 13 per message and disappear once tapped, per our quick-reply entry. A website widget usually has no hard cap but has a phone screen, and a fallback that runs past four lines is not read. Voice has a different budget: Alexa caps the audio files in a reprompt at 90 seconds against 240 for the main speech, and its documentation's tip is a "specific reprompt that lets the user know what to say next," which in practice is one sentence naming two things to say.

The consequence for copy: write the options first, in under 20 characters each, and the sentence second. If the option does not fit the button, it is the wrong option name, and the customer will see the truncated version or the API will reject the message. The conversational UX patterns guide covers the constraint layer channel by channel.

The ten-message test

Before the bot goes live, and again after every scope change, send it these ten and read what comes back. The test is ours; it takes about fifteen minutes.

1. A message in the scope, misspelled ("retrun policy"). Expect a match, not a fallback; if it falls back, the fix is training, not copy.
2. A message in the scope, phrased as a complaint ("why is my order late"). Expect the order-status path or a rung-one message naming it.
3. A one-word message ("hours"). Expect a match.
4. A greeting only ("hi"). Expect the welcome, not the fallback; small-talk handling covers the rest.
5. Something out of scope and answerable by a person ("can I get a wholesale price list"). Expect the out-of-scope message with the exit on the first attempt.
6. Something per-customer ("where is my order"). Expect the lookup if integrated, the handoff if not; never "please rephrase."
7. The same nonsense three times ("asdf"). Expect three different messages and a handoff or contact line by the third.
8. A wrong-format answer to a question ("tomorrow" when asked for an order number). Expect the format restated and the same ask, not a fallback.
9. Silence for the channel's idle period. Expect at most one check-in, then nothing.
10. "Talk to a person" as the first message. Expect the transfer, not the scope list; Intercom's documentation notes Fin "will follow the handover path without trying to generate a response" for that case, and yours should too.

Score it honestly: any run that produces the same sentence twice in a row fails, and any run where "where is my order" gets "please rephrase" fails. The QA testing protocol has the wider regression set; this ten is the fallback slice.

What our fifteen reviews record

Searched 10 September 2026 across the fifteen files matched by sample-reviews/*-review.md, case-insensitively: fallback appears in 9; default reply|default answer|standard reply in 2 (Manychat, SendPulse); didn't understand|don't understand|rephrase in 1 (Landbot); reprompt|re-prompt in 0. The Landbot review captured the Validation error message field on a Question block with its default copy and called the per-block placement "meaningfully less friction than platforms that route validation through a separate rule editor." The Chatfuel review captured a Fallback message field on the AI behavior form alongside the greeting and message-length pickers. The SendPulse review captured a Standard reply trigger on the bot-structure tab with its own sent and engagement counts (37 sent, 0 engagement on the test bot), and the Manychat review the Default Reply hook on a channel page. None of the fifteen reviews scored the wording of a fallback message, and none ran the same-nonsense-three-times probe; the protocol tests handoff, not the sentences before it. That is a gap recorded here, not a finding about the platforms.

Where it breaks

The copy cannot fix three things. A fallback that fires on a third of messages is a training or scope problem; rewriting the message makes each miss cheaper and does nothing to the count, which is the fallback-rate guide's work. A rung-three message that promises a person when the inbox is unstaffed is a lie the customer discovers in an hour; Tidio's offline configuration and Intercom's "bot only" mode both exist because the honest message when nobody is there is "nobody is there, here is what happens to your message," and the escalation playbook covers the routing. And a fallback written once and never reread drifts out of scope as the bot grows: the scope list in the rung-one message is a copy of the content strategy's scope map, and it goes stale the day a new topic ships. Put the fallback messages in the same review as the knowledge base, on the same cadence.

FAQ

What should a chatbot say when it doesn't understand?

Four things, in one or two lines: that it missed; what it can help with; those options as buttons or quick replies where the channel allows, so the customer taps instead of retyping; and how to reach a person. For example: "I didn't catch that. I can help with order status, returns, or store hours — tap one, or type "person" to reach the team." Avoid the shipped defaults, which apologize and ask the customer to rephrase without saying what would work.

How many times should a chatbot ask before handing off?

Twice, with a different message each time: the first narrows the ask, the second switches to buttons with a person among them, and the third does not ask at all; it hands off or states what happens to the message. Dialogflow CX allows up to six numbered attempts, and Voiceflow's no-reply setting, once switched on, defaults to five follow-ups on silence, but those are ceilings. Intercom's Fin treats three repeated turns as a loop and offers a human.

Should the fallback message apologize?

Once, briefly, if at all. "I didn't catch that" carries the acknowledgment. The wordings from Rasa's documentation, Rasa Pro and Tidio open with "Sorry" or "I'm sorry," Landbot's field, as captured, with "I'm afraid," and on a channel with 20-character buttons the apology costs space the scope list needs.

What is the difference between a fallback message and a reprompt?

A reprompt repeats a question the bot asked and the customer did not answer usably (silence in voice, a wrong-format reply in chat); it restates the ask and the expected format. A fallback handles a message the bot could not place when no question was pending; it states the scope and an exit. The reprompt entry covers the mechanism and the per-platform limits.

Can I customize the fallback message on Intercom Fin?

Not as a text field. Intercom's documentation says escalation guidance "influences Fin's replies and send-off messages," and Fin's own FAQ, as quoted in a January 2026 community thread, says the default handoff line "cannot be directly edited or removed" but can be shaped by guidance. The workflow step "If Fin can't resolve the conversation" sets what happens; guidance shapes what Fin says on the way.

What should the fallback say on WhatsApp?

The same four things, with the options as reply buttons: three per message, 20 characters each. Write the button labels first ("Track an order," "Start a return," "Talk to a person") and the sentence second. The chatbot scripts guide carries the WhatsApp character table.

Related guides

• Reduce chatbot fallback rate — the diagnosis: why fallback fires and the five buckets it falls into.
• Chatbot scripts — the seven blocks, the character budgets, and the failure ladder as a mechanism.
• Chatbot escalation playbook — the triggers that send a conversation to a person.
• Chatbot handoff design guide — what the transfer message says and what the agent receives.
• Chatbot welcome message design — the first message, which sets the scope the fallback repeats.
• Chatbot confidence policy — the band in which a "did you mean" message fires.
• Reprompt (glossary) — the second ask as a mechanism, with Alexa, Dialogflow CX, Voiceflow and Rasa limits.
• Fallback intent (glossary) — the intent that catches the message nothing else did.

Sources

• Rasa Open Source documentation, Fallback and Human Handoff (legacy-docs-oss.rasa.com, last updated 25 April 2025), read 10 September 2026: utter_out_of_scope "Sorry, I can't handle that request."; utter_please_rephrase and utter_ask_rephrase "I'm sorry, I didn't quite understand that. Could you rephrase?"; utter_default "Sorry I didn't get that. Can you rephrase?"; the Two-Stage Fallback sequence and "an ultimate fallback action is triggered (e.g. a handoff to a human)"; the handoff example "Would you like to be transferred to a human assistant?" The affirmation wording "Did you mean '{intent}'?" is read from rasa/core/actions/action.py on the project's main branch the same day, not from the docs page. Rasa Pro documentation, Patterns (rasa.com/docs/reference/primitives/patterns, last updated 9 September 2026), read 10 September 2026: utter_cannot_handle "I'm sorry, I'm not trained to help with that."; utter_ask_rephrase "I'm sorry I am unable to understand you, could you please rephrase?"; utter_clarification_options_rasa "I can help, but I need more information. Which of these would you like to do: {{context.clarification_options}}?"; utter_human_handoff_not_available.
• Tidio Help Center, Lyro - the conversational AI agent (article 9003475527196, updated 17 July 2026), read 10 September 2026 in a browser: "You can decide what Lyro is going to do if a visitor asks to speak with a live agent, or if Lyro is unable to answer the visitor's question"; "If Lyro is allowed to transfer conversations to live agents during online hours, Lyro will offer the Talk to a person option"; "If Lyro is supposed to keep all conversations while you're online, then it will ask the visitor to paraphrase their question, or simply explain that it lacks the necessary knowledge"; "the predefined handoff messages are automatically translated for any users who message you and require a language that's different from your default"; the editable predefined messages (four named in the article's text; the screenshot shows five fields, with Create a ticket split by online and offline). The default wordings in the table are transcribed from the article's screenshot of the Predefined answers panel (attachment 22969276105884); the emoji is rendered as it appears there. Tidio Help Center, Lyro Guidance (article 20810956911004, updated 4 September 2026), read 10 September 2026: "Escalate to a human agent immediately if the customer is angry or annoyed."
• Intercom Help, Manage Fin AI Agent's escalation guidance and rules (article 12396892), read 10 September 2026: "Loop: The customer repeats themselves across three turns"; "Fin will never offer escalation twice in a row. If an offer to escalate is triggered again immediately after a previous offer, Fin will escalate the conversation straight away to prevent looping"; "By default Fin posts a brief handoff message (e.g. "Let me connect you with a teammate") before routing"; "Applied guidance now influences Fin's replies and send-off messages"; the example guidance "Offer to route the customer to a human agent if they ask the same question twice." Intercom Help, Deploy Fin AI Agent over chat (article 8286630), read 10 September 2026: "Step 6: If Fin can't resolve the conversation"; "if Fin is unable to find an answer to a new user query, the conversation will be automatically routed to the team." Intercom Help, Use Fin AI Agent in Workflows (article 10032299), read 10 September 2026: "If a customer sends a message similar to "talk to a person" as their first message when Fin is triggered, Fin will follow the handover path without trying to generate a response." Intercom Help, Configure Fin's behavior when no human agents are available (article 15234261, 22 May 2026), read 10 September 2026: the "bot only" mode and the fixed 4-minute follow-up. Intercom Help, Optimize Fin instantly with the help of AI (article 11390088, 14 May 2026), read 8 September 2026 for /glossary/faq-chatbot: the content, customer-data ("such as order status or account details") and action ("like updating a workflow or canceling an order") gap definitions. Intercom Community, Can I customize the customer-facing message Fin sends when an escalation rule happens? (8 January 2026), read 10 September 2026, quoting Fin's FAQ (fin.ai help article 10763810, not fetched directly): "While this default message itself cannot be directly edited or removed, you can influence this process by creating Guidance in Fin."
• Manychat Help, Default Reply in Manychat (article 14281159586588, updated 3 December 2025), read 9 September 2026: "It acts as a safety net, ensuring users always receive a reply"; "It's a great idea to include contact information in your Default reply, allowing users to reach out to your business directly for questions the system cannot answer"; "You can also incorporate Buttons or Quick replies to guide users in the right direction if they get stuck or need additional help"; the every-time and once-every-24-hours firing options.
• Voiceflow documentation, Behaviour (voiceflow.com/docs/documentation/build/behaviour, dateModified 31 July 2026), read 9 September 2026: Global no-match "Generative - provide a prompt that tells the agent how to handle the situation" and "Scripted - write a fixed response that's sent every time"; Global no-reply "Max no-reply messages per turn - how many times the agent will follow up before ending the session (default 5)."
• Google Cloud, State handlers, Dialogflow CX documentation (stamped "Last updated 2026-08-26 UTC" in the copy fetched 9 September 2026 and "Last updated 2026-09-03 UTC" when re-read on 10 September; the quoted text was identical), read 9 and 10 September 2026: sys.no-match-[1-6] "invoked instead of sys.no-match-default and in order"; the parallel sys.no-input-[1-6] row.
• Amazon Alexa Skills Kit, Request and Response JSON Reference (undated) and Manage Skill Sessions and Session Attributes (last updated 20 March 2026), read 9 and 10 September 2026: the 90-second reprompt and 240-second outputSpeech audio limits; "Tip: Include a specific reprompt that lets the user know what to say next."
• Meta, WhatsApp Business Platform interactive-message documentation, as read on 12 August 2026 for /academy/chatbot-scripts: three reply buttons per message, 20-character button titles, 1,024-character body. Not re-read for this guide; the figures are cited to that page and its date.
• Chatbotscape review corpus (the fifteen platform reviews at /reviews), searched 10 September 2026 from the repository root. Denominator: ls sample-reviews/*-review.md | wc -l returns 15. grep -liE '\bfallback' sample-reviews/*-review.md returns 9; grep -liE 'default reply|default answer|standard reply' sample-reviews/*-review.md returns 2 (manychat, sendpulse); grep -liE "didn.t understand|don.t understand|rephrase" sample-reviews/*-review.md returns 1 (landbot); grep -liE 'reprompt|re-prompt' sample-reviews/*-review.md returns 0. Passages cited: landbot-review.md lines 708 and 714; chatfuel-review.md lines 577–578; sendpulse-review.md lines 545–546; manychat-review.md line 672.
• Ahrefs Keywords Explorer, US overview, queried 9 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 support lead of a small business who is answerable for what a chatbot says when it fails. It reads the fallback defaults, documentation examples and settings of seven vendors (Rasa, Tidio, Landbot, Intercom, Manychat, Chatfuel and Voiceflow) as worked examples, with Dialogflow CX and Alexa read for the attempt counters; it does not rank them, and the platforms named are those whose documentation, or our own dated review screenshots, state the default or the setting plainly enough to quote. Where a platform has a Chatbotscape review, the review is linked; those reviews carry affiliate links, disclosed on each.

Methodology

The Rasa, Tidio, Intercom, Manychat, Voiceflow, Dialogflow CX and Alexa pages named in Sources were read in full on 9 and 10 September 2026 (one Intercom article, on the three gap types, on 8 September for the FAQ chatbot entry), and every default and setting name carries its source; the Tidio default wordings are transcribed from a screenshot in the vendor's own article, which is noted, and the Landbot and Chatfuel fields come from the dated screenshots in our reviews, not from vendor documentation. The corpus counts are ours and the commands are printed. The editorial judgments on this page: the four jobs, the six-situation split, the three-rung ladder and its copy, the reset rule, the prompt-rule translation and the ten-message test. They are reasoning from the vendor guidance quoted and from the failure modes in our reviews, not a measured result; where a vendor's own guidance supports a rule (Manychat on contact information and buttons, Amazon on a specific reprompt, Intercom on the three-turn loop) it is quoted beside the rule. The 12 August 2026 WhatsApp figures are cited to the guide that read them.

Last updated

10 September 2026.