Support KB Article Structure: How to Organize Articles for AI Retrieval
Support KB Article Structure: How to Organize Articles for AI Retrieval
Support KB articles should be structured so customers can find the answer quickly and AI systems can retrieve useful pieces of information independently.
Good KB content is not just well-written. It's built so AI can retrieve it accurately and customers can use it without reading everything around it.
That requires structure at every level. Not just the article itself, but the sections inside it and even the language used within it. AI doesn't always work with an article as one complete unit. It often works with smaller pieces of information. Those pieces need to stand on their own.
Get any one of these levels wrong and retrieval becomes unpredictable, even if the writing itself is good.
Atomic vs hub — the article level
Every article in your KB does one of two things. It either answers a question or it organizes answers. Those are different jobs and they require different structures.
An atomic article answers one specific intent completely. A customer arrives with a question, gets the answer, and can act on it without going somewhere else. Think of a task that needs to be solved by following a five step process.
A hub article organizes multiple related intents. Instead of providing a solution to each intent, it groups related topics and points customers toward individual atomic articles.
Guides are hubs.
Hubs primarily exist for navigation and orientation. AI systems may retrieve them, but they generally make weaker answer sources because they organize information rather than fully answer a specific question.
When creating hubs, make sure you provide context or brief orienting text along with the atomic articles you link out to.
What a bad list of links looks like:
Step 1: Set up your payments
Set up Stripe
Configure payouts
Manage subscriptions
What a better list of links looks like:
Step 1: Set up your payments:
Set up Stripe — Connect Stripe before accepting payments.
Configure payouts — Control how and when revenue is deposited.
Manage subscriptions — Update billing plans and recurring payment settings.
Short descriptions in addition to hyperlinked titles of supporting atomic articles help customers navigate and give AI enough context to understand how the topics connect.
Answer first — the opening of every article
The opening of a KB article should answer the question the title asks.
Customers scan for answers. AI systems also rely on clear, self-contained information. If the opening only introduces the topic instead of answering it, customers keep searching and AI retrieves context instead of value.
What a bad opening looks like:
In this article, we'll explain how subscription cancellation works and walk through a few things to consider before getting started.
Avoid openings that only explain what's coming without providing a high level answer.
Common culprits:
"In this article, we'll..."
"This guide will walk you through..."
"Before we get started..."
What a good opening looks like:
You can cancel your subscription at any time. If you're on an annual plan, cancellation takes effect at the end of your current billing cycle.
The opening should provide an extremely brief answer to the title of the article.
The rule
Your opening 1–2 sentences should:
Answer the question the title asks
Stand alone if retrieved without the rest of the article
Skip the lead-up and preamble
If someone only reads those first sentences, they should still walk away with the answer. They will not have all the detailed steps, but the answer is clear.
Text should stand on its own
Images and videos can support an article, but don’t rely on visuals alone. Instructions, requirements, limitations, and important decisions should always exist in text.
Screenshots and videos should supplement and help explain information. They should not be the only place information lives.
Customers may skim images. AI systems primarily rely on text. If important information only exists in a screenshot or video, it becomes harder for both people and AI to use.
Sections as standalone answers
Most customer-facing AI systems use some form of Retrieval-Augmented Generation (RAG). Rather than reading your entire knowledge base, they retrieve smaller pieces of relevant content and use that information to generate a response.
The exact retrieval process varies across platforms, but the principle stays the same: the quality of the response depends heavily on the quality and structure of what gets retrieved.
Every section needs to stand alone because AI systems often retrieve smaller pieces of an article rather than the article as a whole. Even when an article covers one intent, individual sections should still make sense independently.
The easiest way to enforce this is to treat every section as a question-and-answer pair. The heading doesn't need to be written as a literal question, but the content underneath it should directly answer one.
A section that opens with "as mentioned above" or "following the previous step" is a section that will fail when retrieved independently.
What a bad “chunk” looks like:
Browser compatibility
Use one of our supported browsers: Chrome or Firefox.
What a better “chunk” looks like:
Browser compatibility
The file upload feature is not supported in Safari. Use Chrome or Firefox for the best experience.
Each article section should still make sense if retrieved out of context from the rest of the article.
Keep ideas small
A paragraph should usually communicate one idea.
Long paragraphs often combine instructions, explanations, exceptions, and troubleshooting into one block of text. That creates mixed signals for customers and AI.
What combining too many topics looks like:
To enable two-factor authentication, go to Settings and enable the option. This improves account security and is available on Business and Enterprise plans only.
Avoid combining instructions and availability into one block. A customer who already knows what the feature is and just wants to enable it has to read through irrelevant information.
What splitting topics looks like:
Enable two-factor authentication
Go to Settings
Select Security
Toggle two-factor authentication on
Plan availabilityTwo-factor authentication is available on Business and Enterprise plans only.
One idea per section creates cleaner boundaries for customers and AI.
Use specific language
Content should make sense even when read in isolation.
What generic language looks like:
It updates automatically.
This feature is available on some plans.
What specific language looks like:
Billing settings update automatically.
SSO is available on Enterprise plans.
The reader and the AI should never have to guess what words like "it," "this," or "that" refer to.
Use consistent terminology
AI matches on language.
If you call the same thing by different names across your KB, retrieval becomes less reliable.
Pick a term for every concept in your product and use it consistently. If it's called a workspace in the product, it's a workspace everywhere. Not a project. Not an account. Not a space. If your plans are Starter, Business, and Enterprise, those are the only words that appear in your content.
This breaks down in practice because different writers use different language, articles get written at different times, and nobody is auditing for consistency. The result is a KB where the same thing has three names depending on who wrote the article and when. An AI agent retrieving content for a question about workspaces will miss articles that called it something else.
Consistent terminology is a maintenance decision as much as a writing decision. Define your terms once, document them somewhere your team can reference, and treat inconsistency as a structural error, not a style preference.
Include the language customers use naturally
Customers rarely describe problems using your product terminology alone.
Someone looking for help with a feature labeled “Deactivate” may search for “turn off”, “disable”, or “remove access” instead.
The goal is not to stuff articles with alternate keywords. It is to make sure customer language and product language overlap enough for people and AI to connect them.
Acknowledge common customer phrasing naturally where it helps clarify meaning.
What including alternative customer language looks like:
To deactivate or turn off a team member’s access, open the team member’s profile and select Deactivate.
Structure determines how content behaves. Conditions tell AI whether that content applies to the customer asking.