Written by the CustomerEagle team
A support assistant needs accurate content about your business to explain its policies and procedures. Missing conditions, outdated values and ambiguous instructions can undermine an answer before the model writes it. Content is one part of quality alongside retrieval, model behavior, configuration and human handoffs. This guide focuses on the part your support team can maintain: clear sources for the questions customers actually ask.
Why your content sets the ceiling
Grounded AI support retrieves relevant passages and uses them to prepare an answer. A source must contain the policy, condition or procedure the customer needs; a citation alone does not establish that the answer is correct. Check whether retrieval found the right passage and whether the reply represented it accurately. When the available evidence is insufficient, provide a useful route to a person.
Two teams using the same software can have different question mixes, content coverage, integrations and operating rules. Use your own conversation sample to find documentation gaps, and compare results using a consistent outcome definition. A more complete knowledge base is an input to better support, not a promised automation percentage.
1. One question per article
The most common structural mistake is the omnibus article: "Shipping, returns and refunds", three thousand words, eleven headings. It is convenient to maintain and close to useless for retrieval. When a customer asks how long a refund takes, the system has to pull the right fragment out of an article that is mostly about something else, and all the surrounding text about shipping actively competes with the part you wanted.
- Split by the question a customer would type, not by the department that owns the answer. "How long does a refund take?" is an article. "Returns policy" is a category.
- If an article needs more than one heading to answer its own title, it is probably two articles.
- Title each article with the question itself, in the customer's phrasing. Titles carry a lot of weight in retrieval, and they are also what a person scanning search results reads.
- Cross-link the split articles so a customer who lands on the wrong one gets to the right one in a click.
2. Put the answer in the first two sentences
Support articles have a habit of opening with context: why the policy exists, what changed last year, who to contact if this does not apply to you. Move all of it down. The opening sentences are what gets retrieved, what gets quoted back to the customer, and what a person in a hurry actually reads.
| Instead of opening with | Open with |
|---|---|
| We want you to be happy with your purchase, and our returns process is designed to be as straightforward as possible. | You can return an unused item within 30 days of delivery. The refund goes back to the original payment method within 5 working days of the parcel reaching our warehouse. |
| Delivery times depend on a number of factors, including the carrier and your location. | Orders placed before 22:00 on a working day ship the same day and normally arrive the next working day. Orders to Belgium take one extra day. |
| For questions about your invoice, please contact our billing team. | Your invoices are in Settings, then Billing, then Invoices. A copy is emailed to the billing address on the account on the first of each month. |
A useful test: if you deleted everything after the second sentence, would the article still answer its own title? If not, rewrite the opening rather than adding more to the end.
3. Write the values out — do not point at a screenshot
Text retrieval needs content that has been extracted and indexed. Do not assume a value shown only in a screenshot, image table or spoken video instruction will be available. Supported PDF and DOCX knowledge uploads can provide extractable text; check the processing result, especially for scans and complex tables. Write important values and click paths as ordinary article text as well.
- State numbers, deadlines and limits as text: 30 days, 22:00, €50, five working days, three delivery attempts.
- Write out the click path in words — Settings, then Billing, then Invoices — even when a screenshot shows the same route.
- Keep the screenshot. It is still the fastest way for a person to confirm they are in the right place. Just do not let it be the only place a fact exists.
- Transcribe the important part of any video. A caption file is content; a video player is not.
4. Cover the cases you would rather not put in writing
Most knowledge bases document the happy path thoroughly and stop where it gets uncomfortable. Those are exactly the conversations that reach a human, and they are the expensive ones. Every unwritten exception is a permanent tax on your queue.
- The parcel is marked delivered and the customer did not receive it. What should they check first, how long do you wait before opening an investigation with the carrier, and what do you offer afterwards?
- The item comes back used or damaged. What do you accept, what do you deduct, and how do you word it?
- The customer is a few days outside the return window. Write down what your team actually does — if agents routinely make an exception up to a week, say so, or the AI will be stricter than your own staff.
- The refund has not appeared. How long does the card issuer take on top of your own processing time, and what do you ask the customer to send you?
- The discount code was not applied. Do you refund the difference afterwards, and under what conditions?
Writing these down has a second benefit that has nothing to do with AI: your agents stop answering them five different ways.
5. Use the words your customers use
Retrieval matches meaning rather than exact keywords, so the occasional synonym is not fatal. Writing consistently in internal vocabulary is. If your customers say "invoice" and your articles say "billing document", or they say "cancel my subscription" and you say "terminate the agreement", you are asking the retrieval step to bridge a gap on every single query.
6. Make retrieval's job easy
Long articles are normally split into chunks before they are indexed, and each chunk is retrieved on its own. A paragraph that only makes sense after the three paragraphs above it will therefore be retrieved without them.
- Make each section self-contained. Repeat the subject rather than starting a section with "it" or "this plan".
- Use headings that state the question the section answers, not one-word labels.
- Prefer short paragraphs and lists to dense blocks. A list survives chunking; a wall of text can be cut mid-argument.
- Put conditions before conclusions. "If your order has not shipped yet, you can cancel it in…" is safer than a cancellation instruction whose exception appears two paragraphs later.
7. Freshness is a support metric, not a content metric
A wrong article is worse than a missing one. Missing content produces a handoff. Wrong content produces a confident, cited answer that is wrong — and the citation makes it more convincing, not less. Every price change, policy change and integration change needs a pass over the articles that mention it.
- Give every article a named owner and a review date. Unowned articles rot quietly.
- When something changes, search the knowledge base for the old value before you announce the new one.
- Retire aggressively. Two articles that disagree are worse than one that is slightly out of date, because you cannot predict which one gets retrieved.
- Control what is indexed. In CustomerEagle each help-center article carries a "Searchable by AI" setting, so an internal or seasonal article can exist for people without being used as an answer source.
A template you can copy
Almost every article in a knowledge base that performs well looks roughly like this.
- Title: the customer's question, in their words where possible.
- First paragraph: the answer, with the actual values in it.
- Conditions: when this does not apply, written as "if… then…".
- Steps: numbered, in words, with screenshots as support rather than as substance.
- Exceptions: the awkward cases, and what you actually do about them.
- Next step: where to go if this did not solve it, including how to reach a person.
- Footer: owner and last-reviewed date — for your team, not for the customer.
Test it with real tickets before you trust it
Test content quality with a repeatable question set in an environment connected to your own sources. Allow time to review answers, make corrections and run the questions again.
- Take fifty real tickets from last month, spread across topics — not the fifty easiest.
- Ask each question in the customer's wording with unnecessary personal details removed, in your configured workspace or an evaluation environment using your own sources. The public interactive demo shows the workflow with fictional sample content; it does not test your own knowledge base.
- Score every answer as correct, incomplete or wrong. Ignore tone entirely at this stage; tone is a settings problem, coverage is not.
- For each incomplete or wrong answer, write down which article should have answered it. That list is your content backlog, already in priority order.
- Re-run the same fifty once you have written them. The delta is what the content work bought you.
Then let the gaps find you
Manual testing gets you the first draft of the backlog. After that, the system knows more about what is missing than you do: every conversation where it had no confident answer is a documented gap. CustomerEagle logs those, groups them, and can draft an article from the conversation for a person to edit and publish. The drafting is automated; the publishing deliberately is not, because an unreviewed article immediately becomes a cited answer.
Keep a content backlog based on your own question mix. Publish reviewed improvements, inspect failed answers and rerun the same questions. Maintain the knowledge base alongside retrieval, model configuration and handoff rules, and measure whether outcomes improve. The financial effect also depends on your volume and plan; the per-agent allowances are on the pricing page.
Frequently asked questions
Write one article per customer question, put the answer in the first two sentences with the actual values as text rather than inside a screenshot, cover the exceptions your agents normally handle case by case, and use the words customers use instead of internal terminology. Then test the articles against real tickets and rewrite the ones that fail.
There is no fixed number. Start with a representative conversation sample, identify recurring questions and document the approved answer and conditions for each. Test coverage and answer accuracy before expanding. A larger article count alone does not establish that the knowledge base addresses your customers' needs.
Check whether an approved source contains the answer, whether processing succeeded and whether the source is searchable by the AI. Then inspect whether the question retrieves the right passage and whether required conditions are present. PDF and DOCX files with extractable text can supply sources through supported knowledge uploads; do not assume image-only or scanned details were extracted successfully.
Short and specific. Long articles are split into chunks before indexing and each chunk is retrieved on its own, so a passage that only makes sense in the context of the whole page will be retrieved without that context. One question per article, with self-contained sections, retrieves far more reliably than a single comprehensive guide.
Screenshots help readers recognize the right screen. Include important values, setting names and click paths as text as well, so the answer does not depend on whether your ingestion system recognizes image content. Check the extracted, searchable text before relying on a fact that appears only in an image.
Tie reviews to change rather than to the calendar: any price, policy or integration change should trigger a search of the knowledge base for the old value the same day. On top of that, give every article a named owner and a review date, and retire outdated articles instead of leaving two versions that contradict each other.
