Chatbot Content Strategy

Deciding What the Bot Answers, Where Each Answer Lives, and Who Keeps It True

Quick answer: Three of the commonest first-month chatbot failures are content failures wearing a technology costume. The bot answered a shipping question with last year's policy because two pages disagreed; it refused a question that was in the help center because the article buried the answer under a heading it never retrieved; it made up an order status because nobody decided that order questions are not the bot's to answer. A content strategy is the set of decisions that prevents those three: what the bot covers and what it routes elsewhere, where each answer lives and who owns it, how the content is written so retrieval finds it, and how the list is fed and pruned every week. The vendors now do more of this than they used to. Read on 8 September 2026, Intercom publishes a 14-factor "content readiness" checklist and a weekly report that sorts the non-answers it keeps into a content gap, a data gap or an action gap; Tidio lists the questions Lyro missed in live chat with an occurrence count; BotPenguin exports the unanswered questions by date range; Chatbase flags missing sources in the data-sources rail. What none of them does is decide scope for you, say which owner is right when two sources disagree, or give you a rule for when an answer is retired. That is this guide.

Step 1. Draw the scope map: answer, look up, act, or hand off

The first decision is not what to write but what not to. Every question a customer might ask falls into one of four columns, and only the first is content.

Intercom's Recommendations report is built on the same split. Read 8 September 2026, it groups the unresolved conversations it keeps into "content gaps," "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." The point of the split for a small team is that only the first column is fixed by writing, and a content strategy that tries to fix the other three with more articles produces a bot that describes the refund policy fluently and cannot refund anything. Tidio says the same thing from the content side: leave out "Questions the agent truly can't handle. Be honest about Lyro's limitations. Don't try to fake it."

Build the map from demand, not from your site map. Pull the last 30 days of tickets, chats and the site's search-box log, cluster them, and assign each cluster to a column. The knowledge base guide covers the clustering; the output here is the map, one row per cluster with its column, plus a note you will fill in later: which report told you the cluster exists (step 5). Our editorial rule of thumb, untimed: 20 to 30 clusters cover the first launch for a five-to-ten-topic business; a cluster with fewer than three questions in a month is not worth a pair yet.

Step 2. Keep a source-of-truth register, one owner per topic

The second failure in the opening paragraph, two pages that disagree, is the one Intercom now detects, BotPenguin explains, and neither can fix. Intercom's Sunday scan "prepares up to 20 new recommendations for you to review on Monday," "around 15" potential contradictions and "around 5" duplicates, "depending on what's found." BotPenguin's FAQ page explains a bot "saying contradictory information" as one "trained on sources that give the bot contradictory information regarding the same piece of data" and tells you to "clear this contradiction." Neither can tell you which version is true, because that is not in the data.

So the register is the strategy's core artifact: one row per topic on the scope map's Answer column, with the topic, the single place its truth lives, the owner who can change it, the vendor format it is delivered in, and the date it was last confirmed. Intercom's content guidance asks for the same thing in the audit step: "divide your content by owner and ask each team or SME to review their knowledge area."

The format column matters more than it looks, because the five vendors store content differently and each storage choice has a maintenance cost.

The rule that falls out of the table: each fact lives in exactly one format on one platform, and the register says which. A shipping policy that exists as a help-center article, a Tidio manual pair and a line in a SendPulse prompt is three places to forget on the day the policy changes, and the bot will serve whichever version retrieval ranks first.

Step 3. Write for retrieval, and know where the vendors disagree

The knowledge base guide covers drafting. What belongs here is the short list of rules that decide whether a written answer is ever found, because the third failure in the opening paragraph, the buried answer, is a formatting failure. Intercom's checklist of 30 June 2026 has 14 factors; on six of them at least one of the other two vendors lands in the same place, and those six are the cross-vendor minimum.

1. The answer restates the question. Intercom calls it query-answer symmetry: write "as if you're doing a radio interview and you don't want to be quoted out of context." Its poor example is a question followed by "No." Tidio's version is "Provide context before details," and its bad examples "lack context." A retrieved chunk can arrive without its heading, so a bare "No" answers nothing.
2. One topic per section, and the heading's words appear in the paragraph below it. Intercom asks you to "include some of the header text in the paragraph below it, in case Fin or Copilot fails to capture all headings." BotPenguin's fallback rule is "One paragraph = One point = Under 800 characters."
3. Numbers are exact. Intercom: "'A few minutes,' 'some requests,' 'shortly' — none of these pass." BotPenguin: "Supports up to 100 users." not "Supports many users."
4. Acronyms are spelled out on first use. Both vendors give the same example shape: "CSV (comma-separated values)" at Intercom, "Chief Executive Officer (CEO)" at BotPenguin.
5. Limitations are written down. Intercom's most direct line: "If a limitation isn't documented, Fin doesn't know to flag it." Tidio's product-question guidance says the same: "Be clear about which product questions the agent can and cannot address in its current state."
6. Images have text beside them. Intercom sends "a maximum of 3 images in any single AI answer" and wants step-by-step text either way; BotPenguin says not to include images at all because "AI doesn't process them."

Where they disagree is the fluid fact. Tidio's FAQ guidance says to "avoid FAQs about fluid things like pricing, availability, etc. Link to pages you can update dynamically instead." Intercom wants the number in the content, with worked examples, because it does not want Fin to calculate ("Don't ask Fin to do math"). Both are right about their own retrieval, and the register resolves it: a fluid number has one owner and one home, and every other mention links to it or is regenerated from it on a schedule. We would not put a price in a Tidio manual pair; we would put it in the page Tidio scans, and re-scan on price-change day.

One editorial addition the vendors do not make: write the refusal. A pair whose answer is "we do not currently offer X; the closest thing we offer is Y" is content, and it is the pair that stops the bot inventing X. Our hallucination guide covers the weekly check for invented answers; the refusal pair is the cheapest prevention.

Step 4. Split content by audience before it contradicts itself

A contradiction is often two true statements for two customers. EU and US customers have different data policies; wholesale and retail have different return windows; a logged-in user and a visitor see different menus. Intercom's own example for segmenting recommendations by audience is "different data policies for EU vs. US customers," and it recommends segmenting content-gap recommendations by audience so "Fin only analyzes conversations and content relevant to a specific group of customers." Tidio attaches an Audience to each data source, "the default target audience being 'everyone'," and the audience is built from contact properties and channels with and/or conditions.

The strategy decision is whether you have audiences at all. If every customer gets the same answer, do not split; every split doubles the register. If two customer groups get different answers to the same question, the split is not optional, because, on Intercom's own example, its contradiction scanner can otherwise flag the two correct answers as a conflict, and a bot without audience targeting will serve the wrong one to whoever asks first. Add an Audience column to the register on the day the second group appears, not before.

Step 5. Run the gap loop weekly from the vendor's report

The loop is the part of the strategy that keeps working after launch, and it now has a vendor-supplied input on four of the five platforms. The table below is where each report lives, what it cannot see, and the cadence the vendor itself suggests, all read 8 September 2026. What each report shows row by row, and the mechanics of turning a row into a pair, are in the FAQ chatbot entry and are not repeated here.

Two rules make the loop a strategy rather than a chore. First, the report is a list of candidates, not a to-do list: Tidio lets you ignore a suggestion, BotPenguin tells you to rephrase a question that is "too broad or vague," and Intercom asks you to accept, edit or reject. A row goes into the register only if it belongs in the Answer column of step 1; a "where is my order" row is a data gap and goes to the integration owner instead. Second, the report is blind to the customer who left. Tidio cannot see email; Intercom drops abandoned conversations; a keyword bot on SendPulse sees nothing at all. Read the fallback rate and a sample of transcripts alongside the report, on the analytics guide's method, or the loop only ever fixes the gaps the reporting tool happened to notice.

Our editorial cadence for a one-person owner, untimed: 30 minutes a week on the report, one pair or article per session, logged in the register with the date; a monthly hour to re-read the ten highest-traffic answers against their source of truth. Intercom's prioritization advice is the same shape: "Sort your content by 'Last updated' to find the content most likely to contain outdated information," and fix "high-traffic articles" first.

Step 6. Retire on purpose: duplicates, contradictions and stale pairs

A list that only grows accumulates the second failure. On the pages read, three of the five vendors give you a mechanism for duplicates or stale pairs and none gives you a rule, so the rule is editorial.

Intercom's Sunday scan surfaces duplicates ("two articles that both contain very similar instructions on how to reset a password") and contradictions, and lets you edit, delete or mark done; it also warns that "Recommendations are static as of the time they were generated," so a preview may be stale by the time you read it. Tidio's CSV import asks, for a duplicated question, whether to "replace the old version" or "skip the duplicate," and lets you disable a pair for Lyro or Copilot without deleting it. BotPenguin refuses the same FAQ in two categories outright.

The rule: every row in the register has a confirmed-on date, and a row not confirmed in 90 days is reviewed, not assumed. When a policy changes, the owner updates the one home, re-syncs or re-imports whatever is derived from it, and disables rather than deletes the old pair for a week, so the regression run can show the change took. Tidio's Used-by toggle is why this costs nothing: a deleted pair is gone, a disabled one comes back with a click.

What our fifteen reviews record

The counts and line citations are printed in the FAQ chatbot entry's sources and are not repeated here: all fifteen reviews build an FAQ bot first, none uses the phrase "content gap" or "unanswered question," and the only records of a gap report are the BotPenguin review's caption of the Un-answered Questions tab (29 May 2026) and the Tidio review's Lyro AI Hub capture showing "Suggestions 0" (28 May 2026). The Intercom review was last updated 28 May 2026, two weeks after the Recommendations article's date, and does not cover the report; its next refresh should. Our reviews measure the first week, and the gap loop is a second-month habit: a scope note for the reviews, not a finding about the vendors.

Where it breaks

The register lives in someone's head. The most common version of this strategy is one person who knows where everything is. It works until they are on holiday on price-change day. The register is a spreadsheet with five columns, six once you have audiences; it takes an hour.

The vendor's report becomes the strategy. Intercom's Recommendations is the most detailed of the five, and it is gated behind a paid add-on, filters out abandoned conversations, and needs volume to generate anything. A small business that waits for the report to tell it what is missing will wait a long time. Read the transcripts.

Scanned content is edited in place. Tidio's re-sync wipes manual edits to scanned pairs; the same shape exists wherever a synced source is corrected inside the bot rather than at the source. Fix the page, not the pair.

Audience splits multiply before they are needed. Two audiences means two registers, two report views (Intercom regenerates recommendations when you change segmentation, "which may take a few hours") and two test sets. Split on the day the second answer appears, not on the day you imagine it.

The strategy stops at launch. The 14 factors, the register and the scope map are launch work. The loop in step 5 is the strategy; without a weekly slot on a calendar it does not exist.

FAQ

What is a chatbot content strategy?

The set of decisions above the writing: which questions the bot answers from content versus looks up, acts on or hands off; where each answer lives and who owns it; how content is written so retrieval finds it; whether it is split by audience; how the list is fed each week from the bot's unanswered-question report; and how stale, duplicate and contradictory content is retired.

How is this different from building a knowledge base?

Building the knowledge base is writing the articles and pairs. The content strategy decides what should be written at all, who is allowed to change it, which vendor format it lives in, and how it is maintained. Our knowledge base guide is the writing; this page is the governance around it.

Which chatbot platforms report the questions the bot could not answer?

Read 8 September 2026: Intercom (Recommendations, content gaps, Pro add-on), Tidio (Suggestions, live chat only), BotPenguin (Unanswered Questions under Conversation History) and Chatbase (Suggestions in Data sources; Pro-tier per our review, though a Suggestions entry shows on the free-tier rail and the changelog names no plan). SendPulse has no named report; unmatched messages run its Standard reply flow.

How often should chatbot content be reviewed?

BotPenguin's own advice is "Weekly reviews are ideal," and Intercom generates content-gap recommendations weekly. Our editorial cadence for a small team is 30 minutes a week on the report and one change per session, plus a monthly hour re-reading the ten highest-traffic answers against their source of truth. Intercom's tip for choosing which: sort by "Last updated."

Should I put prices in the chatbot's content?

Tidio says no: "Avoid FAQs about fluid things like pricing, availability, etc. Link to pages you can update dynamically instead." Intercom says numbers must be exact in the content. The reconciliation is one home with one owner: put the price on the page the bot scans or the article the owner controls, never in a hand-written pair, and re-sync on the day it changes.

Why does my chatbot contradict itself?

Because two sources disagree about the same fact and the bot picked one. BotPenguin's FAQ says exactly this. Intercom's weekly scan lists the contradictions; fixing them means deciding which version is true and deleting or disabling the other, which is what the source-of-truth register in step 2 exists for.

Related guides

• Build a chatbot knowledge base — the writing this strategy governs.
• FAQ chatbot (glossary) — the question-and-answer pair as a unit, and each vendor's limits on it.
• Conversational analytics guide — reading transcripts by hand for the gaps the reports cannot see.
• Chatbot regression testing guide — proving a content change took without breaking a neighbor.
• Reduce chatbot hallucinations — the weekly check for invented answers the refusal pair prevents.
• Chatbot integration guide — where the look-up and act columns of the scope map go.
• Customer self-service guide — which answers to publish openly and which to keep behind a login.
• Chatbot knowledge base (glossary) — the content layer as a concept.

Sources

• Intercom Help, Optimizing content for Fin (article 7860255, written by Beth-Ann Sher, dated 30 June 2026), read 8 September 2026: the 14-factor content readiness checklist; "as if you're doing a radio interview and you don't want to be quoted out of context" and the "No." example; "include some of the header text in the paragraph below it, in case Fin or Copilot fails to capture all headings"; "'A few minutes,' 'some requests,' 'shortly' — none of these pass"; "CSV (comma-separated values)"; "If a limitation isn't documented, Fin doesn't know to flag it"; "Fin will include a maximum of 3 images in any single AI answer"; "Don't ask Fin to do math"; "divide your content by owner and ask each team or SME to review their knowledge area"; "Sort your content by 'Last updated'"; the high-traffic-first tip.
• 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; "different data policies for EU vs. US customers"; "Fin only analyzes conversations and content relevant to a specific group of customers"; "which may take a few hours"; the weekly and trigger cadence; "checked every Sunday," "up to 20 new recommendations," "around 15" and "around 5"; "expire automatically after being 4 weeks old"; the filtered-out list; "Low-volume customers (with fewer conversations) may receive fewer or no recommendations"; the password-reset duplicate example; "Recommendations are static as of the time they were generated"; the edit, delete, reject and mark-done actions; "cluttered with redundant information" in the duplicate section. Intercom Help, Optimize Fin instantly with the help of AI (article 11390088), read 8 September 2026 in a browser: the three gap types and their definitions; "The Unresolved Questions report has been replaced by recommendations."
• Tidio Help Center, How to create a good FAQ for Lyro? (article 10624642054556, updated 9 April 2026), read 8 September 2026: "Avoid FAQs about fluid things like pricing, availability, etc. Link to pages you can update dynamically instead"; "Questions the agent truly can't handle. Be honest about Lyro's limitations. Don't try to fake it"; "Be clear about which product questions the agent can and cannot address in its current state"; "Provide context before details" and "lack context"; the five-step improvement loop and "Repeat steps 2-4 regularly." Tidio Help Center, Data sources - Lyro's knowledge base (article 14543666652316, updated 22 May 2026), read 8 September 2026: the re-sync note ("removes the associated Q&A pairs and replaces them," "your manual edits will be gone"); "up to 500" URLs and "cannot exceed 10.000" CSV entries; the duplicate-import choice ("replace the old version" or "skip the duplicate"); the Used by toggle; Audiences with "the default target audience being 'everyone'" and conditions combined with and or or; the recommendation to periodically check the Suggestions section (spelled "Suggstions" on the live page). 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)"; the ignore option.
• 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"; "Supports up to 100 users." not "Supports many users."; "Chief Executive Officer (CEO)"; "Embedding FAQs in the main file"; "AI doesn't process them." BotPenguin Documentation, FAQs (same section), read 8 September 2026: "the bot has been trained on sources that give the bot contradictory information regarding the same piece of data"; the two-categories error. BotPenguin Documentation, Unanswered Questions (same section), read 8 September 2026: Bot Training > Conversation History; date range and export; "Weekly reviews are ideal"; "too broad or vague."
• Chatbase, Suggestions for Missing Sources (changelog, 8 October 2025), read 8 September 2026: "which customer questions went unanswered because information was missing." Chatbotscape, Chatbase review (/reviews/chatbase-review): the Data sources rail counted 30 May 2026.
• SendPulse Help Center, Standard reply flow trigger (updated 28 May 2025), read 8 September 2026: the no-matching-trigger condition; the Filter that routes to the AI Step element or sends a reply-shortly message. SendPulse Help Center, The AI Agent element in chatbots (updated 17 January 2025), read 8 September 2026: "you need to add prompts to every AI Agent element" and the global-variable workaround.
• Chatbotscape review corpus (the fifteen platform reviews at /reviews), searched 8 September 2026; the search strings and counts are printed in the sources of /glossary/faq-chatbot. Passages cited: botpenguin-review.md line 482; tidio-review.md line 609; intercom-review.md frontmatter last_updated: "2026-05-28".
• Ahrefs Keywords Explorer, US overview, queried 8 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. It reads five vendors' content and training documentation as worked examples of the decisions an operator has to make; it does not rank them, and the platforms named carry affiliate links disclosed above.

Methodology

The three Intercom help articles, the three Tidio help-center articles, the three BotPenguin documentation pages, the Chatbase changelog entry and the two SendPulse help-center articles were read in full on 8 September 2026, and every quotation carries its source. The corpus counts are our own and the commands are printed in the glossary sibling. The editorial judgments on this page: the four-column scope map in step 1 and its 20-to-30-cluster and three-questions-a-month thresholds; the five-column register in step 2; the choice of six factors from Intercom's fourteen in step 3 and the reconciliation of the fluid-fact disagreement; the split-on-the-second-answer rule in step 4; the 30-minutes-a-week and monthly-hour cadence in step 5; the 90-day confirmation and disable-before-delete rule in step 6. None of these is a vendor figure and none was timed in a deployment.

Last updated

9 September 2026.