Guide · AI support bots & help-center docs
Help center AI readiness audit: a step-by-step checklist
You're about to point an AI support bot at your help center. Maybe it's Intercom Fin, Zendesk's AI agent, Chatbase, or something you built on your own docs. Before you switch it on, it's worth asking a plain question: if a new support hire could only answer from these pages, word for word, would they get it right?
That's roughly what the bot will do. It retrieves pieces of your docs and answers from them. Whatever is missing, out of date, or contradictory in the docs becomes missing, out of date, or contradictory in the answers.
This guide is a pre-launch audit you can run yourself. It's built for a small team with a public help center, not a documentation department. If a bot is already live and answering wrong, start with why your AI chatbot gives wrong answers instead. That guide traces one failure back to its source.
What "AI-ready" means in practice
A help center is ready for a bot when, for the questions customers actually ask:
- There's one current answer. Not two pages with different values, not an old page next to a new one.
- The conditions sit next to the claim. Plan, role, region, and dates are in the same sentence or section as the answer.
- Each section makes sense alone. A retrieved chunk shouldn't depend on "as mentioned above."
- The answer exists in text. Not only in a screenshot, a video, or a support agent's head.
- The bot can reach it. The page is public, or deliberately shared with the bot, and it's in the bot's source list.
HelpCenter.io's checklist puts the goal in similar terms: content that is "accurate, understandable on its own, available to the intended assistant and audience, and tested against real customer questions."
Before you start: pick the questions
Don't audit the whole help center first. Audit the questions.
Write down a short list of the questions customers ask, weighted toward the ones where a wrong answer costs money or trust. Good sources are your support inbox, saved replies, and help center search terms. A starter list for a typical SaaS product:
- How much does it cost for my team size?
- Can I cancel, and what happens to my data?
- Do you support SSO?
- Which plan includes [the feature people ask about most]?
- How do I [the setup step people get stuck on]?
- Where is my data stored?
Ten is a reasonable first batch. HelpCenter.io suggests the same number and is careful to call it "a suggested starting point, not an industry benchmark." The point is to finish one loop, not to cover everything.
The audit, pass by pass
Run each pass on your question list. Record findings in a simple sheet (a template is below).
Pass 1: Find every page that answers each question
For each question, search your help center the way a customer would, and also for the underlying fact (the price, the limit, the feature name). Note every page that states an answer, not just the best one. Include pricing pages, blog posts, changelogs, and any files you plan to upload to the bot.
Pass 2: Check for contradictions
If two pages answer the same question, compare them line by line. Different values mean one is wrong.
Illustrative example: your pricing page says the Starter plan includes "up to 5 projects." A getting-started article, written before a plan change, says "Starter users can create unlimited projects." A bot asked "how many projects on Starter?" can quote either one. Pick the true value, fix the other page, and note which page now owns that fact.
Vendors treat this as a real category of problem. Intercom's Fin, for example, has a "Fix contradicting content" feature that surfaces articles "at odds with one another."
Pass 3: Check that conditions sit next to claims
Look for answers that are true only under a condition stated somewhere else.
Illustrative example: a feature page says "Export to CSV from any report." Three paragraphs down, a note says "Exports are available to workspace admins." A bot that retrieves the first sentence alone tells a regular member they can export. Rewrite it as one sentence: "Workspace admins can export any report to CSV."
Pass 4: Test whether sections stand alone
Copy one section of a key article into a blank document. Can someone tell what product, feature, and conditions it's about? Watch for:
- "As mentioned above," "see below," "this option"
- Headings like "Overview" or "Other notes" that say nothing about the content
- Steps that only make sense after reading an earlier section
HelpCenter.io suggests the same test: "copy one section into an empty document."
Pass 5: Find answers that exist only in images or people
Illustrative example: your SSO setup article is five screenshots with captions like "Step 1" and "Step 2." The actual field names and values are only in the images. Depending on the bot, it may get nothing useful from that page. Write the steps out in text next to the screenshots.
Then look for missing answers. If your team answers "Do you support SSO?" by hand every week and no public page says yes or no, the bot will still answer. It just won't have anything true to answer from. Write the page.
Pass 6: Check access
Open each page on your list in a private browser window, logged out. If you can't read it, a bot that crawls public pages can't either. If you use Zendesk, visibility is set per article: its API documents user_segment_id as the setting that "defines who can see this article," with null meaning everyone.
Then open the bot's source list and confirm each page is actually included. Also check the reverse: nothing internal or meant for one customer segment should be in a public bot's sources.
Pass 7: Ask the bot, before customers do
If your platform has a preview or test mode, ask every question on your list. Then ask each one a second way, the way a rushed customer would type it. For each answer, check:
- Is it supported by the page you chose as the owner?
- Does it include the condition (plan, role, region)?
- Does the cited source actually say what the answer says?
HelpCenter.io's warning is worth keeping in mind: "A source link beside an answer does not establish that the answer is correct." Read the source.
A simple audit sheet
One row per question is enough. Illustrative row:
| Question | Pages that answer it | Conflict? | Condition beside claim? | Stands alone? | In text? | Public + in bot sources? | Bot answer OK? | Fix |
|---|---|---|---|---|---|---|---|---|
| How many projects on Starter? | Pricing, Getting started | Yes (5 vs unlimited) | Yes | Yes | Yes | Yes | No, said unlimited | Fix Getting started; pricing page owns the fact |
Keep the sheet after launch. Re-run the bot-answer column after every release that touches a listed question.
Fix in this order
- Contradictions on money and trust facts. Price, limits, cancellation, data, security. These do the most damage when wrong.
- Missing answers to common questions. A short, direct page beats a bot guess.
- Conditions separated from claims. Rewrite into single sentences.
- Sections that don't stand alone. Fix headings and "see above" references.
- Image-only instructions. Add text.
- Everything else. Tone, formatting, long pages.
Don't start by rewriting the whole help center. Fix the rows that failed, re-test those questions, and expand from there.
After launch
Readiness doesn't stay fixed. Every release, policy change, or pricing update can create a new conflict or stale page. Two habits help:
- Put "search docs for the old value, update, re-sync the bot" on the release checklist. Chatbot giving outdated answers covers that in detail.
- Keep your question list and re-ask it after changes.
FAQ
How do I know if my help center is ready for an AI chatbot?
Pick the questions customers ask most, especially where mistakes are costly. For each, check that exactly one current page answers it, with its conditions in the same place, in text, publicly reachable, and included in the bot's sources. Then ask the bot and check its answer against that page.
Do I need to rewrite my whole knowledge base first?
No. Fix the pages behind your priority questions, starting with contradictions on pricing, limits, and policies. Expand as testing and support tickets show new gaps.
How many questions should I test?
Start with a small batch you can finish, such as ten, then add more as you find problems. There's no official number. What matters is testing the same questions again after changes.
Can the bot read screenshots and videos?
Some can process images and some only get extracted text. Check your platform. Either way, write essential steps in text so every reader, human or bot, gets them.
What if a page is behind a login?
A bot that crawls your public site can't read it. Either make the answer public, add it to the bot's sources deliberately, or accept that the bot won't know it and tell customers where to ask instead.
Related guides
Run it with a kit
If you want this audit as ready-made files instead of building the sheet yourself, the Docs Contradiction Toolkit has a scorecard, an expanded checklist, anonymized worked examples, bot-probe prompts, and Notion/CSV templates. It's a $39 one-time download for one help center, and a first pass takes about 30–60 minutes. It isn't a chatbot or a subscription, and it won't guarantee your bot's answers. It helps you find the conflicts, stale pages, and buried conditions before the bot does. Everything in this guide works without it.
Sources
- HelpCenter.io, "The practical checklist for AI-ready help center content": https://helpcenter.io/blog/the-practical-checklist-for-ai-ready-help-center-content/
- Intercom Help, "Use AI-powered content recommendations to improve Fin" (section "Fix contradicting content"): https://www.intercom.com/help/en/articles/11394959-use-ai-powered-content-recommendations-to-improve-fin
- Zendesk Developer Docs, Help Center API, "Articles" (field
user_segment_id): https://developer.zendesk.com/api-reference/help_center/help-center-api/articles/