My chatbot gives outdated answers: what to hand a freelancer

The short answer

A chatbot that quotes last year’s price is almost never a problem with the AI model. It is a problem with what the bot is given to read, and how that material follows your changes. Somewhere there is an old copy of a fact the bot can still find, or a new fact that never reached it.

So do not brief “make it smarter”. Brief the sync: which source is the truth, how a change reaches the bot (including deletions), how you can see when it last updated, and who hears about it when an update fails. Then accept the fix with four tests: change a fact, delete a document, ask about both, and break the update on purpose.

First, find where the old answer comes from

Before you message anyone, spend ten minutes finding the source of one wrong answer. It changes what you ask for, and it changes the quote.

Where the old fact lives	How you can tell	What the fix is
An old document still in the bot’s store	The bot names an old file or page as its source	Remove it, and make the sync remove deleted documents from now on
A new version never reached the bot	You edited the document, the bot still shows the old wording	Fix or add the update step
Two documents disagree	Sometimes old, sometimes new, for the same question	Fix the documents; one fact, one place
Written into the bot’s instructions	The fact appears even with documents removed; no source shown	Move the fact out of the instructions into a document
A cached or saved answer	Same wording every time, even after a change	Clear it, and set how long saved answers live
The model’s general knowledge	No source at all; the fact is common knowledge about your industry	Tell it to answer only from your documents, and say so when they don’t cover it

The instructions case surprises people. Developers sometimes paste a price or an opening time into the bot’s system prompt to fix one answer quickly. It works until the price changes, and then nothing you edit in your documents makes a difference.

If the bot can show its sources, this whole step takes minutes. If it cannot, that is the first thing to ask for. The basics page on retrieval explains why the source behind each answer is the part you can actually check.

Why outdated knowledge goes unnoticed

A wrong answer from old information looks exactly like a right one. It is fluent, specific and confident, and it was correct once. Nobody gets an error. The customer either believes it or leaves.

My own experience with this is not with a chatbot but with notes and research that other jobs depend on. Two things I now do as rules:

• Secondhand facts get rechecked at the source. My content pipeline takes topic ideas from other people’s short videos. The idea can be reused, but any date, amount or eligibility rule is checked again at the primary source before it is used, because secondhand text mixes in rules that have since changed. A chatbot’s documents are secondhand too: someone wrote them at some point, and the world moved on.
• A note that says “done” is not the live state. I once wrote in my own run-book that I had changed a setting on my scheduled jobs. A week later I checked the jobs directly, and none of them had the change; it had needed administrator rights and never applied. The note was sincere and wrong. The same goes for “the knowledge base is up to date”: ask to see the date of the last successful update, not a statement that it is current.

What a sync should do

When you hire someone to fix outdated answers, you are really hiring them to build or repair a small scheduled job: it reads your documents, finds what changed and updates the bot’s store. Like any scheduled job, it needs five things.

The fifth line is the one I would not skip. Any scheduled job that logs into another service will one day find its login expired. One of mine kept running for two days while doing nothing, and every layer reported success (the incident). A chatbot sync can fail the same way, and the bot will keep answering confidently from its last copy.

A brief block you can copy

Paste this into the Knowledge and Maintenance fields of the chatbot brief, with your own details:

Problem. The bot gives outdated answers. Example: asked “[question]”, it says “[old answer]”; the current answer is “[new answer]”, in [document]. Source of truth. Documents live in [place]. I edit them there. Sync. Changes, new documents and deleted documents reach the bot within [time], by [on change / nightly / a button]. No facts in the instructions. Prices, dates, opening times and policies come from documents only, not from the bot’s instructions. Visibility. I can see the date of the last successful update and the number of documents in the bot. Failure. If an update fails or finds zero documents, [me, by email or chat] gets an alert the same day. Acceptance. The four tests below pass.

Including one real wrong answer helps more than any description. A seller who reads it can often tell you which of the six cases it is before quoting.

Four tests before you accept the fix

1. Change a fact. Edit one price or date in the source. Wait the agreed time. Ask the bot. It gives the new value and names the right source.
2. Delete a document. Remove one document entirely. Ask a question only it answered. The bot no longer gives that answer; it says it does not know or hands over.
3. Ask an old question. Ask the question that produced the outdated answer in the first place. It is gone, from every channel the bot is on.
4. Break the sync once. Ask the developer to make one update fail on purpose, for example with a revoked login. You receive the alert.

Tests 2 and 4 are the ones that catch most incomplete fixes. They follow the same idea as the test sheet for chatbots before you accept them: test what happens when something is missing, not only when everything is there.

What moves the price

The chatbot job page lists starting prices of $100 to $1,000 for the vetted gigs, checked in September 2026. A sync fix moves within and beyond that range with:

• Where your documents live. A folder of files is simpler than a live connection to a help center or shared drive that needs its own login.
• Removals. Detecting deleted documents takes more work than adding new ones.
• How fresh it must be. “Within a minute of every change” costs more than “every night”.
• How many places the facts live. If prices sit in documents, in the instructions and in a separate FAQ widget, all three need cleaning up.

Where it goes wrong anyway

• The documents themselves are stale. A perfect sync faithfully copies an outdated page. Someone has to own the documents, usually you, as chatbots on your own documents explains.
• The quick fix comes back. Under time pressure, someone pastes a fact into the instructions again. Repeat test 3 after every change to the bot.
• The sync is treated as a change, not maintenance. Agree up front that a broken sync is a fix, as in putting maintenance in the contract.

Wrap-up

Outdated answers are a plumbing problem, not an intelligence problem. Find where one wrong answer comes from, brief a sync that adds, changes and removes, make its last update date visible and its failures loud, and accept it by changing and deleting real documents. Then the bot is only ever as old as your documents, which is the one thing you can control.