FAQ Chatbot — What It Is, the Two Ways to Build One, and How Many Q&A Pairs It Takes

Quick answer: An FAQ chatbot answers the questions you are asked over and over ("what are your hours", "do you ship to Canada", "how do I reset my password") from a list of question-and-answer pairs you wrote. It is the oldest chatbot use case and, in our protocol, the first bot: in every one of our fifteen platform reviews, the first scenario is an FAQ bot (a ten-question one in fourteen of them). The word "FAQ" describes the content, not the technology. The same list of pairs can power a keyword-matching flow that fires a scripted reply, or an AI agent that treats the pairs as a knowledge base and writes the answer in its own words.

The vendors have made the pair the unit of work and put limits on it. Read on 8 September 2026: Tidio's Lyro stores scanned pages, PDFs and manual entries as question-and-answer pairs, and its help center estimates that "up to 10 Q&A pairs → up to 40% automation" and "30+ Q&A pairs → up to 70% automation." BotPenguin caps a question at 200 characters and, on its best-practices page, an answer at 500 (its FAQs page says otherwise; see the table), and takes "a maximum of 200 FAQs per category per file." Intercom does not require pairs; it tells you to put small answers in "a list of FAQs" inside one article or in snippets, and "Fin will still be able to find them." SendPulse's keyword trigger matches "individual words and phrases as commands," with a keyword limit of 32 characters, and sends everything that matches nothing to a Standard reply flow. None of them decides the list for you, though Tidio and Intercom will draft entries from what they scrape or observe, and four of the five now tell you what is missing from it.

Two builds, one list

The phrase covers two different machines, and the difference decides what a pair has to look like.

The keyword or menu FAQ bot matches the customer's message against words you listed and plays a reply you scripted. SendPulse's keyword trigger is the plain version: you enter keywords ("Hello, price, what is the price, how much does it cost"), matching is case-insensitive, and you choose between Matches with and Contains. With Contains, "the trigger launches as soon as a subscriber's message contains a keyword"; with Matches, a keyword buried in a longer sentence makes the bot send a clarification request in which "the price keyword appears as a quick reply." Anything that matches no trigger lands in the Standard reply flow, which SendPulse generates for every bot and which "cannot be deleted." Manychat's Default Reply and the button menus in most builders are the same idea. A keyword bot never invents an answer, which is its virtue, and never answers a question you did not anticipate, which is its limit; the phrasing gap is the whole fallback problem.

The AI FAQ bot treats the pairs as sources. The customer's message is matched to the closest pair by meaning rather than by word (retrieval over embeddings, not string matching), and a language model phrases the reply from the pair's answer. Tidio's Lyro, BotPenguin's Bot Training, Chatbase's Q&A source and Intercom's Fin are all AI builds of this kind; of the pages read, only Intercom's describes the retrieval step, and the others imply it. The paraphrase problem largely disappears; the new problems are that the answer can drift from what you wrote, and that two pairs saying different things about the same fact make the bot pick one. BotPenguin's own FAQ for the feature puts it flatly: if the bot "is saying contradictory information," it is because "the bot has been trained on sources that give the bot contradictory information regarding the same piece of data."

In our reading of the reviews, most SMB bots end up hybrid: a menu for the five questions that carry a button ("Track my order"), AI pairs for the long tail, and a handoff for the rest.

What each vendor accepts as a pair

The table below lists, for each of the five platforms, what a pair is called, how it gets in, and the limit the vendor states. All read 8 September 2026.

The Tidio re-sync line is the one that bites. If you scrape your site, then hand-edit the pairs Lyro made from it, then re-sync because a page changed, the edits are gone. Our fix, not Tidio's, is to keep edited pairs as manual entries and treat the website folder as read-only, which is a content-ownership decision, not a settings decision; the content strategy guide covers where each answer should live.

How many pairs

Tidio is the only vendor of the five that publishes a number, and it is an estimate, not a benchmark. Its help center says, "Based on user data, we've estimated the amount of possible automation as follows," then gives four rungs: up to 10 pairs, up to 40 percent automation; up to 20, up to 50; up to 30, up to 60; 30 or more, up to 70. Read it as a shape rather than a promise: the first ten pairs do most of the work because they are the ten questions everybody asks, and every pair after thirty buys a smaller slice. "Automation" here is Tidio's word; the metric we track is containment, and the two are not the same thing, since a contained conversation is one the bot handled to the end, not one it merely answered.

BotPenguin gives the shape from the other side, for documents rather than pairs: "It's better to have 30 well-written, focused paragraphs than 5 long paragraphs," with each paragraph "under 800 characters" and "One paragraph = One point." Intercom sets the floor for its gap report: content-gap recommendations are generated on volume (a "High number of conversations where a question and answer can be found"), or "regular queries (1+ a day) on the same topic for at least 7 days," and "low-volume customers (with fewer conversations) may receive fewer or no recommendations."

Our editorial rule of thumb for a business with 1 to 100 employees, untimed and not a vendor figure: write the first 20 pairs from the last month of tickets and chats before launch, then let the unanswered-question report write the rest, one review a week. A test set of one paraphrase per pair, run before each edit, catches the pair that stopped matching.

What a good pair looks like

The vendors agree on more than they differ, and Tidio's help center has the clearest side-by-side. Its good example is "Q: What payment methods do you accept? A: We accept Visa, Mastercard, American Express, Discover, PayPal, and Google Pay. You can update your payment information in your account settings." Its bad example of the same pair is "Q: Payment type options? A: V M A D P G." The bad pairs, it says, "use unnatural language, lack context, and have formatting issues."

Three rules recur across at least two of the three vendors that publish writing guidance, and a fourth is disputed between them:

1. Write the question the way a customer types it. Tidio: use "How do I reset my account password?" and not "Password reset procedure explanation request." BotPenguin: "Always write the full question clearly (not just a keyword or phrase)."
2. Make the answer stand alone. Intercom calls this query-answer symmetry and asks you to write "as if you're doing a radio interview and you don't want to be quoted out of context"; its poor-practice example is a question followed by the single word "No." Tidio's edge-case advice is the same rule: "Provide context before details." An AI bot retrieves the answer without the question, so the answer has to carry the subject.
3. Spell out the acronym once. BotPenguin's example is "Who is the Chief Executive Officer (CEO) of SML Isuzu?"; Intercom's is "CSV (comma-separated values)." A retrieval step that has only ever seen "CEO" will not match "chief executive."
4. Keep fluid facts out of the pair. Tidio: avoid "FAQs about fluid things like pricing, availability, etc. Link to pages you can update dynamically instead." Intercom disagrees on the mechanism, asking for exact numbers in the content ("Returns can take up to 12 business days") because it does not want Fin to do arithmetic. The reconciliation is ownership: a number that changes monthly belongs in one place with one owner, and the pair links to it or is regenerated from it.

One rule the vendors state and most operators skip: BotPenguin bans "emojis, incomplete sentences, or slang" in pairs, and Tidio bans "marketing fluff." The pair is a fact record the bot reads, not a message the customer sees verbatim.

Where the next pair comes from

An FAQ bot's list is never finished, and the useful vendors know it. Each of the four AI platforms now reports the questions the bot could not answer and lets you write the pair from the report. The table below is what each report is called, what it shows, and how it turns into a pair, all read 8 September 2026.

Two things the table hides. First, the reports only see what the bot saw: Tidio's excludes email tickets, Intercom's excludes "conversations without teammate responses" and "abandoned conversations," so a customer who left after Fin's first non-answer is in no content-gap recommendation. Pull the fallback rate alongside the report. Second, the reports are noisy on purpose. Tidio lets you ignore a suggestion; BotPenguin's FAQ tells you to rephrase a question that is "too broad or vague" before training on it. The report is a list of candidates, and the pair is still written by a person.

What an FAQ bot cannot answer

The list has a natural edge. "Where is my order" is an FAQ in form and not in substance: the answer is different for every customer and lives in another system. Intercom's Recommendations sorts non-answers into three reasons, and only the first is a content gap; the other two are "customer data gaps," where "Fin needed information from an external system that wasn't available, such as order status or account details," and "action gaps," where Fin "needed to take an action in another system, like updating a workflow or canceling an order." Neither is fixed by a pair. They need an integration or a handoff rule, and a well-run FAQ bot says so instead of guessing. Tidio's list of things to leave out ends with "Questions the agent truly can't handle. Be honest about Lyro's limitations. Don't try to fake it."

What our fifteen reviews record about FAQ chatbots

All fifteen of our platform reviews mention an FAQ bot: grep -liE '\bfaq\b' returns 15 of 15, because the first scenario of our testing protocol is a ten-question FAQ bot. Times to a working FAQ bot, as the reviews report them: Chatbase 8 minutes, Typebot 11, Manychat 12 (the anchor), Tars 13, Tidio 14 (widget plus FAQ), Botpress 14, BotPenguin 15, Voiceflow 15, Wati 18 (including WhatsApp onboarding). Read these as protocol-anchored figures unless the review's hands-on section re-timed the scenario, which only Tidio's did (28 May 2026): the Manychat review labels its six-scenario figures a methodology-anchored simulation, Botpress's evidence box calls its 14 minutes a calibrated simulation, and Chatbase's changelog calls its assessment projected. The other six reviews frame their scenario-A figure as projected or anchored, or (SendPulse) as not yet run.

Two of fifteen use the term Q&A (grep -liE 'q&a pair|question-and-answer pair|Q&A' returns 2: chatbase, tidio); none uses the phrases "unanswered question" or "content gap" (grep -liE 'unanswered question|content gap' returns 0 of 15), though the BotPenguin review's screenshot caption describes the Un-answered Questions tab as one that "surfaces queries the bot couldn't answer, lets operators add answers back to the KB without leaving the trainer." Our Tidio review captured the Lyro AI Hub with "Suggestions 0" on 28 May 2026. The reports are a first-week blind spot in our reviews, which is a scope note for their next refresh, not a vendor gap.

Related terms

• Chatbot knowledge base — the whole content layer; the pair is its smallest unit.
• Fallback intent — what a keyword FAQ bot does with a message that matches nothing.
• Chatbot fallback rate — the metric that tells you how many pairs are missing.
• Intent recognition — how a bot decides which pair a message is asking for.
• Retrieval-augmented generation — the mechanism that lets an AI FAQ bot match by meaning instead of by keyword.
• Customer service chatbot — the use case most FAQ bots are the first version of.

FAQ

What is an FAQ chatbot?

A chatbot that answers the questions a business receives repeatedly from a fixed list of question-and-answer pairs the business wrote or approved. It can be a keyword flow that plays a scripted reply, or an AI agent that reads the pairs as a knowledge source and phrases the answer. Either way the list is the product, and the bot's report of unanswered questions is how the list grows.

How many questions should an FAQ chatbot have?

Tidio's published estimate, "based on user data," is that up to 10 pairs automate up to 40 percent of chats and 30 or more pairs up to 70 percent. Our editorial rule of thumb for a small business is 20 pairs from the last month of tickets before launch, then one pair-writing session a week from the unanswered-question report.

What is the difference between an FAQ chatbot and an AI chatbot?

The FAQ is the content; AI is one of two ways to serve it. A keyword FAQ bot matches words and plays a script (SendPulse's keyword trigger). An AI FAQ bot matches by meaning and phrases the answer from the pair (Tidio's Lyro, BotPenguin's Bot Training, Chatbase's Q&A source, Intercom's Fin). The AI build handles paraphrase; the keyword build never contradicts itself.

Where do I find the questions my FAQ chatbot could not answer?

Read 8 September 2026: Tidio under Lyro AI Agent > Suggestions (live chats only, not email tickets); BotPenguin under Bot Training > Conversation History > Unanswered Questions; Chatbase under Data sources > Suggestions; Intercom under Fin AI Agent > Analyze > Recommendations with the reason filtered to Content gaps (Pro add-on). SendPulse has no named report; unmatched messages run the Standard reply flow. Our Chatbase review records the sources-suggestions feature as Pro-tier.

Can an FAQ chatbot answer "where is my order"?

Not from a pair. The answer is per-customer and lives in another system, which Intercom classifies as a customer data gap rather than a content gap. It needs an integration that looks the order up, or a handoff to a person, and the bot should say which.

Sources

• Tidio Help Center, How to create a good FAQ for Lyro? (article 10624642054556, updated 9 April 2026), read 8 September 2026: the automation-rate estimate ("Based on user data, we've estimated the amount of possible automation as follows," 10/40, 20/50, 30/60, 30+/70); "How do I reset my account password?" versus "Password reset procedure explanation request"; the good and bad payment-methods pairs; "use unnatural language, lack context, and have formatting issues"; "Avoid FAQs about fluid things like pricing, availability, etc. Link to pages you can update dynamically instead"; "Marketing fluff"; "Provide context before details"; "Questions the agent truly can't handle. Be honest about Lyro's limitations. Don't try to fake it."
• Tidio Help Center, Data sources - Lyro's knowledge base (article 14543666652316, updated 22 May 2026), read 8 September 2026: the add-knowledge methods; "Scan priority pages option is limited to scanning up to 60 pages"; "You can add up to 500 URLs overall"; a CSV "can contain a maximum of 500 entries" and "the total number of entries cannot exceed 10.000"; the re-sync note ("removes the associated Q&A pairs and replaces them with new ones," "your manual edits will be gone"); the Create answer button in a conversation; the Manual label ("directly from this tab, or from a chat where Lyro failed to answer"). Tidio Help Center, Lyro Suggestions (article 24182486548892, updated 19 March 2026), read 8 September 2026: "only available for live conversations, and not for tickets (emails)"; "any questions that Lyro was unable to answer successfully when chatting with visitors live"; the occurrences, intent and date columns; pre-filled answers; "Save as a Q&A" and the Manually added label; the Playground Add answer button.
• BotPenguin Documentation, Best Practices for Training Your AI (help.botpenguin.com, website and mobile-app bot section), read 8 September 2026: "One paragraph = One point = Under 800 characters"; "30 well-written, focused paragraphs than 5 long paragraphs"; question "Maximum length: 200 characters," answer "Maximum length: 500 characters"; "Max 200 FAQs per category (split if needed)"; "Always write the full question clearly (not just a keyword or phrase)"; the SML Isuzu CEO example; "Including emojis, incomplete sentences, or slang"; "Embedding FAQs in the main file" listed as a mistake. BotPenguin Documentation, FAQs (same section), read 8 September 2026: "Maximum File Size: 20 MB"; "the answer can be effectively as long as you want"; the two-categories error; "the bot has been trained on sources that give the bot contradictory information regarding the same piece of data." BotPenguin Documentation, Unanswered Questions (same section), read 8 September 2026: Bot Training > Conversation History; date ranges and export; "Add FAQ"; "Weekly reviews are ideal"; rephrasing a question that is "too broad or vague."
• Chatbase, Suggestions for Missing Sources (changelog, 8 October 2025), read 8 September 2026: "Your AI Agent can now detect knowledge gaps - showing you which customer questions went unanswered because information was missing"; "One-click data upload to fix gaps." Chatbotscape, Chatbase review (/reviews/chatbase-review): the Data sources rail (Files / Text snippets / Q&A / Website / Notion / Tickets / Suggestions) counted 30 May 2026; the 8-minute scenario A.
• Intercom Help, Optimizing content for Fin (article 7860255, written by Beth-Ann Sher, dated 30 June 2026), read 8 September 2026: "as if you're doing a radio interview and you don't want to be quoted out of context"; the "No." poor-practice example; "include them as a list of FAQs (frequently asked questions) in one public / internal article or create snippets. Fin will still be able to find them"; "use headers (H1, H2, H3) when structuring your FAQs"; "Fin will include a maximum of 3 images in any single AI answer"; "CSV (comma-separated values)"; "Returns can take up to 12 business days." Intercom Help, Use AI-powered content recommendations to improve Fin (article 11394959, written by Beth-Ann Sher, dated 14 May 2026), read 8 September 2026: "missing, unclear, duplicated, or contradictory"; the Pro add-on note; the daily-or-weekly triggers ("Regular queries (1+ a day) on the same topic for at least 7 days"); "checked every Sunday," "up to 20 new recommendations," "around 15" contradictions and "around 5" duplicates; "expire automatically after being 4 weeks old"; snippets "immediately added to Knowledge"; the filtered-out list ("Conversations without teammate responses," "Abandoned conversations"); "Low-volume customers (with fewer conversations) may receive fewer or no recommendations." Intercom Help, Optimize Fin instantly with the help of AI (article 11390088, dated 14 May 2026), read 8 September 2026 in a browser: the three gap types and their definitions ("such as order status or account details"; "like updating a workflow or canceling an order"); "The Unresolved Questions report has been replaced by recommendations"; the older article 8890980, opened in a browser the same day, returned Intercom's page-not-found message.
• SendPulse Help Center, Keyword flow trigger (updated 11 January 2026), read 8 September 2026: "The maximum keyword length is 32 characters"; the example keywords; case-insensitive matching; Matches with versus Contains and the clarification request with the keyword "as a quick reply"; "the bot selects the trigger linked to the first keyword listed"; the throttle interval. SendPulse Help Center, Standard reply flow trigger (updated 28 May 2025), read 8 September 2026: "triggered when a user sends a message without a keyword (command) or if no matching trigger exists"; "System flows cannot be deleted"; the Filter that checks for AI integration and routes to "your AI assistant in the AI Step element," or sends a message "to inform users that you will reply shortly"; SendPulse's element article, The AI Agent element in chatbots (updated 17 January 2025), read the same day, names the same element AI Agent.
• Chatbotscape review corpus (the fifteen platform reviews listed at /reviews), searched 8 September 2026 from the repository root. Denominator: ls sample-reviews/*-review.md | wc -l returns 15. grep -liE '\bfaq\b' sample-reviews/*-review.md returns 15; grep -liE 'unanswered question|content gap' sample-reviews/*-review.md returns 0; grep -liE 'q&a pair|question-and-answer pair|Q&A' sample-reviews/*-review.md returns 2 (chatbase, tidio). Passages cited: botpenguin-review.md line 482 (the Un-answered Questions caption); tidio-review.md line 609 (Suggestions 0) and line 544 (14 min); chatbase-review.md line 506 (the Data sources rail with Suggestions), line 675 (the Pro-tier sources-suggestions note) and the 8-minute measurement; botpress-review.md line 231 (14 min) and line 52 (the calibrated-simulation note); voiceflow-review.md line 198 (15 min); manychat-review.md line 757 (12 min) and line 532 (the methodology-anchored simulation label); tidio-review.md line 194 (the hands-on re-timing); chatbase-review.md line 826 (the projected-assessment label); typebot-review.md line 339 (11 min); tars-review.md line 190 (13 min); wati-review.md line 156 (18 min); botpenguin-review.md line 162 (15 min).
• Ahrefs Keywords Explorer, US overview, queried 8 September 2026 — the demand, difficulty, CPC, global-volume and parent-topic figures in this entry's keyword note.
• Chatbotscape evaluation methodology. /methodology (continuously updated).