Start with ten real support questions, not a chatbot. Pick questions your team answers repeatedly, identify the article that should answer each one, then test whether an agent can find that article and use it correctly.
An AI knowledge base has two jobs: give customers reliable articles to read and give agents a way to find the same content. A chat widget is optional. If you only want external agents to read your help center, focus on discovery and retrieval first. The AI-native knowledge base hub shows the three surfaces DocsAlot uses for that job: llms.txt, Markdown, and hosted MCP.
1. Pick one support workflow
Start with a workflow such as connecting a store, inviting a teammate, or exporting a report. Pull ten questions from recent support conversations and rewrite them without account IDs, emails, or other customer data. Keep the phrasing customers actually use.
For each question, record the expected answer, the canonical article URL, and whether the agent should answer at all. For example, a public article can explain how exports work. It cannot tell a customer why their specific export failed without the right account context.
Do not publish raw tickets as source material. Resolve conflicting advice, check what is still true in the product, and publish the reusable answer instead. Keep internal workarounds and account-specific instructions out of the public index.
2. Make each article answer one task
Use a title customers might type into search. Put the answer or outcome first, followed by prerequisites, steps, and a recovery path. Keep critical conditions in text, even when you have screenshots.
Here is a structure you can reuse for an export article. Fill the bracketed fields with verified product details:
1# Export a report as CSV2
3Who can export: [role or plan].4Before you start: [required permissions or settings].5
61. Open [exact location in the product].72. Select [exact button or menu label].83. Choose [date range or filters] and start the export.9
10Success: [where the download appears and what it contains].11If export is unavailable: [likely cause and what to check].12If your export failed: [support path for account-specific help].Link related articles instead of copying the same rule into several places. If two pages disagree about a plan limit, cleaner Markdown will not tell an agent which one is correct. Fix the source before tuning retrieval.
3. Publish the help center and verify the agent-facing files
Publish the canonical HTML articles for people. Then give agents a shorter path to the same source:
- Markdown: expose a plain-text version of each published article, including its conditions, steps, and useful links. Open a real Markdown page to see what the response looks like.
llms.txt: provide a compact index with a description of the help center and links to key articles. The llms.txt proposal describes the format; it does not require agents to read or index it.- Hosted MCP: if your intended clients support it, connect them to a server that exposes search or retrieval. Test the actual tools that the client sees. MCP provides access to content; it does not replace public discovery or automatically install a support widget. See the MCP architecture overview.
Test the index and at least one linked article from an unauthenticated session. Replace these example URLs with your own:
1curl -fsSL -i https://help.example.com/llms.txt2curl -fsSL -i https://help.example.com/guides/export-a-report.mdRead the response body. A successful status code alone is not enough: a site can return an HTML homepage or a sign-in screen in place of the file. Check that the index links to the intended host, the article returns readable content, and the Markdown still includes any important permission or plan limits. If you also use MCP, run a search and open the result in the connected client.
4. Test the answers before shipping
Run the same ten questions in the client your team or customers will use. Start fresh conversations, note which access method is enabled, and check the source yourself. Don't count a fluent answer as correct if it cites the wrong page or drops a restriction.
| Questions | What to test | What passing looks like |
|---|---|---|
| 1-4 | Common documented tasks | The agent finds the intended article and gives the correct steps. |
| 5-6 | Alternate wording | The same answer works with the words customers use, not just article titles. |
| 7-8 | Role, plan, or version limits | The answer retains the restriction and cites a source that supports it. |
| 9 | Missing or undocumented policy | The agent says it cannot verify the answer instead of inventing a rule. |
| 10 | Account-specific problem | The agent points to support without guessing about the account or exposing private data. |
For each miss, record the retrieved page, the actual answer, and what went wrong. If retrieval found an old article, fix the links or remove the conflicting source. If the right article left out a prerequisite, edit the article. If the agent still guesses when no source exists, do not expose that workflow to customers until it can handle the gap.
5. Give the knowledge base an owner
Assign someone to update the article when a product workflow or policy changes. Republish it, verify the Markdown and index, and rerun the affected questions. If you use a separate search index or answer widget, check that it is using the updated source rather than assuming a publish automatically refreshed every client.
Start with one workflow that passes these tests. Expand into billing, onboarding, or troubleshooting when you can keep the answers accurate. If you want to publish the help center and expose the same articles to agents, start with the DocsAlot AI-native knowledge base.