New prices from January 1: preparing cutover changes in your assistant
New prices from January 1: how the new list can wait as a draft without the assistant quoting it in December, and how to switch over on the cutover date.
In plain language
If you change your prices for the new year, your assistant should only name them from that day.
Set up the new page as a draft first. The assistant does not use a page that says noindex in its head.
On the set day you make the page public and let the source be read again. Only then does it name the new prices.
Then check its answers in a test chat. If a new price is wrong, fix it in the source. Only bring back the old state if reading the source went wrong.
Around the turn of the year, many businesses change their prices, rates and terms, and often their opening hours between Christmas and New Year as well. For 2027 some have a fixed occasion on top: the statutory minimum wage in Germany rises on January 1. For a chat assistant, a cutover date like this raises a practical question. How can the new price list be ready in time without the assistant quoting it in December – and how does it reach the answers on the cutover date without the assistant still quoting the old amounts in January? This article walks through the path from taking stock, through preparing the list as a draft, to switching over and checking. It also shows where XICBOT helps and why a change deliberately does not take effect in the same moment.
Labels in the illustration
- Website
- Old price page: public, valid until Dec 31
- old page removed or noindex
- New price page: noindex, no sections
- New price page: public
- On the cutover: publish the page, re-read the source
- Re-read
- Knowledge
- Version n of the source: answers with old prices
- even after publishing, until the run ends
- Run
- Version n+1
- new prices
- Source rollback: newer versions are dropped
- Cutover January 1
- Dec 1
- Dec 15
- Dec 31
- Jan 1
- Jan 10
- schematic
- Draft or noindex: no sections
- Old prices: carry answers
- Run: switches at the end
- New prices: carry answers
January 1 as the cutover date
For many businesses the turn of the year is when new price lists, rates and terms take effect. For 2027 there is an occasion that many wage calculations share: from January 1, 2027 the statutory minimum wage in Germany is 14.60 euros gross per hour (Section 1 of the Fifth Minimum Wage Adjustment Ordinance). Anyone who sells working time, for example in trades, building cleaning, hospitality or care, therefore often recalculates hourly rates at the turn of the year. How that calculation works is not the subject of this article, and it does not say whether or by how much your prices should rise. This is about the step after that: the new list is settled, and the assistant on your website should quote it from the right day onwards.
Der Mindestlohn beträgt 1. ab 1. Januar 2026 13,90 Euro brutto je Zeitstunde, 2. ab 1. Januar 2027 14,60 Euro brutto je Zeitstunde. For an assistant, a cutover date is an unusual task. It answers from what is in its knowledge base, and a price list without a date looks just as valid in December as in January. Whether an amount already applies or still applies is something it can only pass on if the source says so. The changeover is therefore not a question of the language model but a question of the sources: which of them are in the collection at which point, and which of them may carry answers.
Too early or too late: two failure patterns
A price changeover can go wrong in the chat in two ways, and both are unpleasant. In the first case the new list is in the collection too early. A visitor asks about the hourly rate on December 12 and gets the amount that only applies from January. They plan with a price you would not charge them today, or they are puzzled by an increase that has not even been announced. In the second case the old list stays in place too long. On January 4 the assistant still quotes last year's rate, and the discrepancy only shows when the quote is written. Which price details belong in the chat in general is covered in the article on prices in the chat; here the question is only about timing.
| Failure pattern | Typical cause | What the visitor experiences |
|---|---|---|
| New prices in December | new price page already public without noindex, new file already added as a source, or new question-and-answer pair already active | an amount that does not apply yet |
| Old prices in January | no run of the source since the website changed; old file or old text still in the collection | an amount that no longer applies |
| Both amounts alternating | old and new list are both present as sources | one amount or the other, depending on the question |
| Old amount despite the new page | a question-and-answer pair or a text source with the old amount was overlooked | the old amount, although the website is correct |
Too early is as wrong as too late
Both failure patterns produce information that does not fit the day of the question. The goal is therefore not the fastest possible changeover but one that matches the cutover date: nothing new in the collection before it, nothing old after it.
Start with an inventory: where prices appear
Before anything is prepared, it pays to list every place where the affected amounts and terms appear. Prices rarely live only on the price page. They appear in service descriptions, in PDF price lists, in frequently asked questions, in conditions for travel costs or minimum orders, and sometimes in a single sentence on the home page. What matters for the assistant is the route by which a detail enters its knowledge base. With XICBOT these are essentially four source types – website, file, text and question-and-answer pairs – plus the maintained details that belong to the building blocks of the assistant itself. Each of these places changes in a different way, and that determines what happens to it on the cutover date.
| Where the detail appears | How it changes | What to do on the cutover date |
|---|---|---|
| Website page | with the next run of the website source; due according to its fetch interval, seven days by default | create the new page in advance with noindex, publish it on the cutover date and choose Re-read |
| File, for example a PDF price list | when you change it; a file has no scheduled refresh | add the new file on the cutover date, remove the old source |
| Text source | when you change it; like a file, without refresh | add the new text on the cutover date, remove the old source |
| Question-and-answer pair | when you change it; without a run | create the new pair in advance and set it to inactive, activate it on the cutover date, switch off the old one |
| Maintained details in the assistant | with a new version of the assistant | publish the changed details as a new version on the cutover date |
Two aids help you find these places. In the Knowledge area you can view sources and their documents, and with “Check answer” you can see which source passages an answer to a price question would be built from today; every role in the account may do both. Which price questions visitors actually ask is shown in conversation analytics. Whatever comes up often there belongs in the check after the switch later on. A simple table with the location, source type, old and new value and the name of the person responsible on the cutover date is enough as a result.
Preparing the new list in advance
The new price page does not have to stay in a drawer until December 31. It can already be finished on the website as long as it is marked as a draft. For the XICBOT knowledge crawler, that marker is noindex in the robots meta tag of the page: such a page is read, but it is not split into sections and does not feed into any answer. It is only read, however, if the knowledge crawler reaches it: the crawler works through the start address, the sitemap, the pages from the last run and links in the page text, at most 25 pages per run. If a run reaches the page, it appears in the document list of a source with the note that it does not feed into answers. That lets you check in December whether the page is found and whether the marker takes effect. How pages can be kept out of the knowledge base in general is covered in the article on exclusions in the knowledge base.
A noindex page and a draft source receive no section at all — the barrier stands before writing, not before reading.
For the other source types this does not apply in the same form. A file or a text changes when you change it and is split into sections as soon as it is added. A draft source, as mentioned in the quotation above, can only be set up as such when it is created and cannot be published later, so it does not help with the cutover date. The new PDF price list is therefore best kept ready on your side, not in the collection, until the cutover date. Question-and-answer pairs, on the other hand, have a switch of their own: only active pairs with an answer are used. A new pair with the new amount can be created in December and set to inactive right away; on the cutover date a single tick is then enough. Good preparation therefore has six parts.
- New page with noindex: fully designed, with the marker in the head of the page and not linked from navigation or menus. So that the knowledge crawler still reads it, it is listed in the sitemap or linked from a page the crawler reads. Some content management systems leave noindex pages out of the sitemap; then only a link helps.
- Start of validity in the text: the new page states itself when it applies, for example “valid from January 1, 2027”.
- End of validity on the old page: in December the current list already says “valid until December 31, 2026”. The assistant states figures the way they stand in the section; if the date is in the same section, it can mention it too.
- Files and texts ready but not added: the new PDF price list is prepared and is uploaded only on the cutover date.
- Pairs created but inactive: the new pair carries the new amount and is switched off; the old one stays active until the cutover date.
- One responsible person: who switches over on the cutover date is settled, and that person has the required role in the account.
How the knowledge base switches over
To make the cutover date plannable, it helps to look at how a change reaches the answers. A run does not write into the body that is currently answering but into the next version. The switch happens only at the end of the run, and only if the new version differs from the previous state. If the run fails, the old state keeps answering, and a run that has written nothing deletes nothing either. A document that has not changed is skipped; a page that has become a draft loses its sections. For the cutover date this means: the old price page disappears from the answers as soon as a run no longer finds it or reads it as noindex. The new one is added as soon as a run reads it without the marker.
A changed price list does not take effect in the same moment, and we do not write it that way either.
On its own, a run comes according to schedule. A website source carries a fetch interval, seven days by default, adjustable from one day up to a year; the value zero means it is not refreshed automatically at all. A daily run picks out which sources are due and asks with ETag and modification date: if your server reports a page as unchanged, it is not split again. Anyone who only changes the website on January 1 therefore waits until the source is due again. With the default interval that can be several days in which the assistant quotes the old amounts.
The cutover date needs a trigger
Neither the noindex marker nor a date triggers a run on January 1. The changeover in the assistant begins with a run of the source, and you either trigger it yourself or have it triggered. Planning for that usually shortens the period with old amounts to the duration of one run.
The sequence on cutover day
On the cutover date itself, the order matters. First the website changes, then the assistant reads it again, then it is checked. The other way round, a run still reads the old state, and the new page is missing until the next run. January 1 is a public holiday; many businesses therefore switch over on the last working day of the old year or the first working day of the new one. Which point in time fits depends on when the new prices are meant to apply to your customers. That is a decision for your business and, if in doubt, a question for your professional advisers, not for the assistant. Once the point in time is settled, the steps follow in this order.
- Change the website: publish the new price page or remove the noindex marker, and add it to the navigation or the sitemap so that the run reaches it; remove the old page, redirect it to the new one or set it to noindex in turn.
- Re-read: in the Knowledge area, choose “Re-read” on the website source. This queues a run that works in the background; until it ends, the previous state keeps answering. If a run is already open for the source, the request joins it. If that run was already working before the page was published, choose “Re-read” again once it has ended. Afterwards, check in the document list that the new page appears there without the draft note.
- Swap files and texts: add the new PDF price list or the new text as a source and delete the old source, so that two lists are not in the collection at the same time.
- Switch pairs: activate the new question-and-answer pair and switch off or delete the old one. This takes effect without a run.
- Adjust maintained details: if an amount sits in the building blocks of the assistant, publish a new version. New conversations start with it; running ones stay with their previous version.
- Check: ask the most frequent price questions in a test conversation and look at the source references, as described in the section on checking below.
The hours between versions
Even with good preparation, a short window remains. Between publishing the page and the end of the run, the assistant answers from the previous version. Conversations that began before a new version of the assistant also keep the building blocks they started with. Someone who asks about prices on New Year's Eve or on the morning of January 1 may therefore still receive the old state. This window cannot be closed completely, but it can be shaped so that an answer within it does not mislead.
- Validity dates in both lists: if the old list says “valid until December 31, 2026” and the new one “valid from January 1, 2027”, each answer can mention its period.
- A note about the change: an active question-and-answer pair “Are your prices changing at the turn of the year?” can already say in December that a new list applies from January 1 – without amounts, as long as you do not want to state them yet.
- Only one valid list in the collection: if two versions of the same detail are present as sources, either one can end up in an answer. Old files and texts go out on the cutover date, not at the next tidy-up.
- Borderline cases to your team: a question that touches both periods, such as a job that starts in December and ends in January, usually belongs with a person. A sentence in the assistant's handover rule makes that explicit.
Checking after the switch
Whether the changeover has arrived shows not in the status of the source but in an answer. In a test conversation the assistant answers questions against the current configuration without any visitor being affected; source references and tool calls appear in the transcript. Ask the three or four price questions that come up most often and check which source the answer comes from. If the reference still points to the old page or an old file, the run has not finished yet, or the old source is still in the collection.
| Test question | Expected from the cutover date | Warning sign |
|---|---|---|
| How much is an hour of work with you? | new amount, reference on the new price page | old amount or reference on the old page |
| Is there a flat travel charge? | detail from the new list with its start of validity | two different amounts in two answers |
| Does the old price still apply to my December job? | handover to your team or the rule from your source | a rule that is in no source |
| Are you open between Christmas and New Year? | hours from the current source | hours from the previous year |
This can be repeated with the check questions of the acceptance run. They ask questions with a model answer along the same path as a real conversation, and the verdict is reached without a language model on mandatory terms, forbidden terms and literal overlap. That fits the cutover date exactly: the new amount goes under “Must contain”, the old one under “Must not contain”. You start the run yourself, before and after the cutover date; one beforehand will then show deviations as expected, one afterwards should pass without them. The probe set in the Knowledge area is separate: if you maintain probe lines with a question and the expected document there, they are measured daily after the nightly refresh to see whether the search finds that document. How a probe set is built is described in the article on test cases before go-live. In the first days of January it is also worth looking at the analysis of conversations: price questions without a source hit show where something is still missing.
If something goes wrong: roll back with care
In case a run on the cutover date produces something unexpected, XICBOT keeps five versions per source. An older version can be restored in the Knowledge area; afterwards that version answers again, and the newer ones are discarded. A rollback is therefore lossy going forward. As long as an intake or refresh is waiting or running for the source, no version can be restored. More important than the mechanics is the question of what a rollback is good for. It is the remedy for a failed run, for example if a single page delivered only a maintenance notice during the run. It does not help against a wrong new amount – the restored version from December carries the old prices.
| Feature | Versions of a knowledge source | Versions of the assistant |
|---|---|---|
| What they contain | sections from pages, files and texts | building blocks such as role, greeting, maintained details and handover rule |
| How a new one is created | with a run that changed something | with “Publish new version” |
| What restoring does | the chosen version answers again, newer ones are discarded | the old state is published as a new version, none is lost |
| What helps beforehand | looking at the document list of a source | looking at the difference from the active version |
A wrong amount is corrected in the source
For a knowledge source, restoring means going back to the old state, with all the amounts from back then. If only a single new value is wrong after the cutover date, the shorter route is to correct the source and re-read it. How an assistant can be updated without interruption is described in updating an AI assistant without switching it off.
What XICBOT brings to the cutover
-
Draft page without sections
A page with noindex receives no section at all. The barrier stands before writing, not before reading.
-
Switching at the end of the run
A run writes into the next version. If it fails, the previous state keeps answering.
-
Re-read instead of waiting
After a change on the website, “Re-read” queues a run instead of waiting for the fetch interval. If one is already open, the request joins it.
-
Pairs with their own switch
Question-and-answer pairs can be created in advance and activated only on the cutover date. This takes effect without a run.
-
References in the test conversation
Source references and tool calls appear in the test conversation. That shows which source the new amount comes from.
-
Five versions per source
An older version can be restored. Newer ones are discarded in the process, so check first.
The cutover date is a special case of ongoing upkeep. Anyone who has already defined responsibilities, cycles and expiry dates for the knowledge base will quickly find the steps of this article in their plan; what such an upkeep plan looks like is described in the article keeping the knowledge base up to date. Which source types XICBOT reads, and how fetch interval, versions and drafts work together, is described on the knowledge base page. For the changeover on January 1, 2027, a list of the affected places, a prepared page with noindex, a fixed point in time with a responsible person and a short check in the test conversation are usually enough.
Frequent questions
May the assistant quote the new prices as early as December?
That is your decision. If it should only quote them from the cutover date, the new list does not belong in its answers until then: as a page with noindex, as a finished file outside the collection or as an inactive question-and-answer pair. If you want to announce the change beforehand, a separate pair can point it out without stating amounts.
Does the assistant switch over by itself at midnight on January 1?
Usually not. A change takes effect with the next run of the source, not in the same moment, and neither the noindex marker nor a future date triggers a run on the cutover date. The most reliable route is to choose “Re-read” after publishing the page, or, if we look after your assistant, to tell us the cutover date in advance.
Why does the assistant still quote the old amount after the changeover?
Usually the old amount is still somewhere in the collection: in a question-and-answer pair, in a text source, in an old PDF file or on a page that the last run still read. Maintained text and FAQ sources also carry more weight than website text. The test conversation shows the source references of the answer and therefore usually where the old value still sits.
What does the minimum wage have to do with the assistant?
Directly, nothing. For many businesses it is an occasion to recalculate hourly rates at the turn of the year: from January 1, 2027 it is 14.60 euros gross per hour (Section 1 of the Fifth Minimum Wage Adjustment Ordinance). How you derive your prices from that is a decision for your business; in individual cases professional advice can help.
Can I undo the changeover?
For a knowledge source, one of the kept versions can be restored; the newer ones are discarded in the process, and the restored version carries the amounts from back then. For the building blocks of the assistant, the old state is published as a new version without any being lost. A single wrong value is usually corrected faster in the source.
How do I check whether the new prices have arrived?
Ask the most frequent price questions in a test conversation and look at the source references. If you keep check questions for the acceptance run, you can enter the new amount as a mandatory term and the old one as a forbidden term; a run after the cutover date should then pass without deviations. In the first days afterwards, questions without a source hit in conversation analytics show where something may still be missing.
Ask it yourself, Assistant of this website
Automated assistant