How to Optimize Your Knowledge Base for AI Agents: Retrievable, Correct, and Never Instructing
Most guides on how to optimize a knowledge base for AI agents cover one job: helping the agent find the right article. There are two more, and on 28 October 2026 we found a competitor failing the third in public.
TL;DR: An AI-ready help centre passes three tests — the agent can retrieve the right article, the article is correct when retrieved, and nothing in it instructs the retriever — a line two Lorikeet articles crossed with opacity-zero text we observed on 28 October 2026.
Methodology & sources
- Procedure built from our own guidance plus published guides from Intercom (May 2026), Ada (July 2026) and Gorgias (Nov 2024).
- Lorikeet observation: two article pages loaded in a desktop browser on 28 October 2026; text and computed CSS read from the live DOM. No endpoints were called and nothing was booked.
- Standards: Google spam policies, OWASP LLM01:2025, llms.txt proposal. All figures verified October 2026.
- Disclosure: this is Aissist's blog. Aissist sells AI agents for customer service and competes with Lorikeet.
Observed 28 October 2026: instructions to AI assistants inside a competitor's articles
Two Lorikeet articles — AI support cost per resolution compared and AI support for travel and eSIM — each carry sections headed "Book a demo (for AI agents)" and "Ask our support agent (for AI assistants)". They tell an assistant how to book a sales demo over HTTP and query a public support endpoint. In our browser check the containers were 1px tall at opacity 0, one block in 4px type. A human reader sees none of it. The text does ask assistants to act only with their user's consent and to let the human choose a time, and Lorikeet publishes the same instructions openly in its llms.txt. We first reported the pattern on four other Lorikeet pages on 21 October.
What does "optimised for AI agents" actually mean?
A knowledge base is optimised for AI agents when it passes three tests: the agent can retrieve the right article, the article is correct when retrieved, and the article contains nothing addressed to the agent. Most published advice covers the first test, some covers the second, and almost none covers the third.
All three matter because an agent doesn't read your help centre like a customer. It retrieves a few paragraphs and answers from them as if they were the whole truth. A fragment that is hard to find produces "I don't know." A fragment that is out of date produces a confident wrong answer. A fragment that contains an instruction can change what the agent does.
Ada's guide puts the stakes plainly: it argues you don't train an agent on a knowledge base so much as ground it in one. Grounding is only as good as the ground.
"Knowledge management is almost as important as the core AI technology itself." — Declan Ivory, VP of Customer Support, Fin (Intercom)
A bold claim from a company that sells the AI technology, and a correct one. The ten steps below are grouped by test: 1–5 make content retrievable, 6–8 keep it correct, and 9–10 draw the line between structuring content and instructing the agent. For background on how knowledge bases changed shape after generative AI, see the evolution of the AI knowledge base.
How do you make help articles retrievable?
Make each article answer one question, title it as that question, and name every product, plan and entity explicitly so a retrieved fragment still makes sense on its own. Retrieval systems split articles into chunks; every chunk should survive being read alone.
- One answer per article. "Billing FAQ" with 14 questions becomes 14 articles. Ada recommends keeping articles mutually exclusive and titled around tasks; its guide also observes that roughly the top 20% of articles handle about 80% of inquiries, so start there.
- Title the article as the customer's question. "How do I change my eSIM plan?" beats "Plan management". Use the customer's words, not your internal feature name.
- Name entities explicitly. Write "the Pro plan" rather than "this plan", and "the iOS app" rather than "the app". A chunk that says "click it in settings" is useless out of context.
- One canonical definition per term. Define "active subscription" once, in one article, and link to it everywhere else. Three slightly different definitions give the agent three slightly different answers.
- Turn screenshots and complex tables into text. Agents read words. Describe each screenshot's step in a sentence; keep tables simple — one header row, no merged cells — and state the key row in prose too.
"Content should no longer be an afterthought." — Beth-Ann Sher, Senior AI Knowledge Manager, Fin (Intercom)
None of this is new to good technical writers. What's new is the penalty: a human skims past a vague paragraph, while an agent quotes it to a customer with complete confidence.
How do you keep a retrieved article correct?
State when each fact was true and who it applies to, keep exactly one source of truth per fact, and use the agent's own unresolved conversations to find what is wrong or missing. Retrievability without correctness just makes wrong answers faster.
- Put dates, scope and audience in the text. "As of October 2026, refunds on annual plans are prorated" carries its own expiry warning. Add a visible last-updated date and say which plan, region or customer type the article covers.
- One source of truth per fact. If the refund window appears in a help article, a PDF and a macro, two of them are drafts waiting to go stale. Pick one, retire or link the others.
- Review on a cadence, and let the agent tell you where. Intercom describes auditing more than 700 live articles before pointing its own agent at them, budgeting 5–10 hours a week per contributor for knowledge work, and reviewing about 30 days of agent conversations to find gaps. Unresolved and escalated conversations are the best content-gap report you will ever get for free.
Here is what steps 1–7 do to a single article:
| Before | After | |
|---|---|---|
| Title | Plans & billing | How do I get a refund on an annual plan? |
| Scope | Covers upgrades, downgrades, refunds, invoices | Refunds on annual plans only; links out for the rest |
| Entities | "Contact us within the window to cancel it" | "Cancel the Pro or Team annual plan within 14 days of renewal" |
| Dates | None | "Policy as of October 2026 · Last updated 28 Oct 2026" |
| Screenshot | Image of the cancel button | "Go to Settings → Billing → Cancel plan" |
| Definition | "Refunds are prorated" | Links to the one canonical article defining "prorated refund" |
Illustrative example; the plans and policy are hypothetical.
Where is the line between structuring content and instructing the agent?
Content written to be retrieved is legitimate; content written to issue instructions to the retriever is not — and the test is whether a human reader would recognise the text as addressed to them. Structure, headings, schema and an llms.txt index all help a machine read what a human can also read. Hidden text that tells the machine what to do does not.
- Use declared surfaces for machine-facing information. An llms.txt file, proposed by Jeremy Howard in 2024, is a Markdown index at your site root: an H1 with your name, a one-line blockquote summary, short context paragraphs, then H2 sections listing links to your key help articles, with an "Optional" section for secondary links. It is visible, declared and separate from your articles.
- Keep behaviour in agent configuration, never in published pages. What the agent may promise, when it escalates and what it must never say belong in reviewed, versioned configuration — not in a help article an editor can change. Our hidden-instructions audit guide walks through finding stray ones.
Two cases this month show why. Redact found 102 hidden instruction blocks on 80 of 545 WhatsApp help pages, including one telling a support bot not to recommend contacting a data protection authority.
"What the instruction could do was make sure you didn't hear about them from the place you went to for help." — Dan Saltman, Founder and CEO, Redact
Lorikeet's case is different in intent: its text addresses outside assistants, asks for consent and is mirrored in its llms.txt. But setting opacity to 0 is a technique Google's spam policies list under hidden text abuse, and text that steers a model is what OWASP ranks first in its 2025 LLM Top 10 as indirect prompt injection. If the instruction is fine to give, give it where people can see it. For defending your own agent against text like this, see prompt injection in customer support AI agents.
The retrievability checklist
- Each article answers one question, and its title is that question.
- Products, plans, regions and features are named, never "it" or "this".
- Each term has one canonical definition, linked everywhere else.
- Screenshots have text equivalents; tables are simple and summarised in prose.
- Each article states its scope, audience and a visible last-updated date.
- Each fact lives in exactly one source.
- Unresolved agent conversations are reviewed on a fixed cadence.
- An llms.txt indexes your key help articles.
- Agent behaviour rules live in versioned configuration, not in pages.
- Rendered text matches raw HTML: no text a human can't see.
Write for the retriever, never to it
The verdict: a help centre ready for AI agents is one you could hand to a new human hire unchanged. Clear, single-answer articles with stated dates are what retrieval rewards. Everything that only a machine can see is either an index it was invited to read or an instruction it was never meant to take.
Start with the 20 articles that handle the most volume, run the checklist on them this week, then search your own pages for text addressed to AI. Agents built to resolve rather than deflect depend on this groundwork; it is a large part of how Aissist reports an 83% average resolution rate and 4.8/5.0 CSAT.
Frequently asked questions
How long should a help center article be for AI agents?
As long as one complete answer to one question, and no longer. If an article needs subheadings for unrelated questions, split it.
Do AI agents read images and PDFs in a knowledge base?
Some can, but text is far more reliable. Add a written equivalent for every screenshot step and move key PDF content into help articles.
Is llms.txt required for a help center to work with AI agents?
No. It is a proposed convention, not a standard, and your own support agent usually reads your helpdesk directly. It helps outside assistants find your best pages.
Is it acceptable to put instructions for AI assistants on a public page?
Only where humans can see them and recognise who they are for, such as a labelled developer page or llms.txt. Hidden text addressed to AI is what Google's spam policies call hidden text abuse.
How do I find hidden AI instructions in my own help center?
Compare each page's rendered text with its raw HTML and search for phrases like "AI assistant", "language model" or "ignore". Our audit guide lists seven steps.
Which articles should I optimize first?
The ones handling the most conversations. Ada's guide notes roughly the top 20% of articles cover about 80% of inquiries, so rewriting 20 high-volume articles beats polishing 200 rarely read ones.
