The decision to leave a hosted chat widget is usually made twice: once when the invoice grows faster than the business, and again a month later when someone asks where the conversations are stored. This is the migration itself — what transfers, what genuinely does not, and the order that avoids a gap in cover.
What moves, and what does not
Being honest about this up front saves a bad week later.
- Your knowledge moves for free. This is the part people expect to be hard, and it is the easy part: a retrieval chatbot reads your existing pages, posts, products and PDFs. There is no export to prepare — the content is already on the site.
- Your past conversations usually do not move. Most platforms export a transcript file, but there is nowhere meaningful to import it: it is not knowledge, and the visitors are gone. Export it for your records before you cancel, because access ends with the subscription.
- Your canned replies move by hand. The answers your team types over and over are the best possible seed for a Q&A set. Copy them across — it is an afternoon, and it is the highest-value hour of the migration.
- Contacts and leads should be exported first. If the widget captured email addresses, get the CSV out while the account is still open.
- Your operators’ logins do not move. Human handoff runs on WordPress accounts instead, which is usually a simplification.
The order that avoids a gap
Do not cancel first. Run both for a few days — the old widget on some pages, the new one on others, or the new one only for logged-out visitors — and compare what they answer.
- Export everything you are entitled to while the subscription is live: transcripts, contacts, canned replies, and any FAQ content you wrote inside their tool.
- Install the plugin and connect a provider key. Your own key with your own provider, which is what removes the per-message bill.
- Index your content and check what it learned. Sync your pages, then read the entries list. If a page indexed badly, fix the page — that content was probably hard for people to read too.
- Paste your canned replies in as Q&A pairs. These are your sharpest answers and they beat anything generated.
- Set the business facts. Opening hours, phone number, address and price positioning entered once, so a stale line on an old page cannot contradict them.
- Test with the questions from your old transcripts. You have a list of real customer questions — use it instead of guessing. This is the only migration test that means anything.
- Point handoff at a real inbox and confirm someone is watching it before the old platform stops routing chats.
- Run both for a week, then cancel. Check the unanswered-questions log daily; those gaps are your remaining content work.
What will feel different
A retrieval chatbot on your own content behaves differently from a hosted widget, and knowing how avoids a false alarm in week one.
- It says “I do not know” more often, on purpose. A hosted widget often improvises around a gap. Grounded answers stop where your content stops — those refusals are a content to-do list, not a fault.
- Answer quality tracks your site. A vague pricing page produces vague pricing answers. The upside is that fixing an answer means editing a page, not retraining anything.
- There is no shared inbox app. Handoff lives in WordPress admin. For a one or two person team that is fewer places to look; for a shift-based support team it is a real downgrade, and a platform may genuinely suit you better.
- The bill splits in two. A licence, plus the AI provider at cost. The provider line is usually the surprise — it is normally a few euros a month, not the platform figure it replaces.
Before you cancel: the checklist
- Transcripts and contacts exported and stored somewhere you control.
- The new assistant answering your top ten real questions correctly.
- Handoff tested end to end, by a person who is not you.
- The widget visible on the pages you expect, and hidden where you do not want it.
- Your privacy policy updated: the data now lives on your hosting, and the disclosure line names the AI provider you chose.
That last one is easy to forget and is the one a regulator would ask about. The plugin generates the policy text for you, but it is your policy — read it.
Questions
Can I import my old chat transcripts as knowledge?
You can, but it is usually a bad idea. Transcripts contain wrong answers, half-answers and one-off promises, and a retrieval system will treat all of it as true. Mine them for the QUESTIONS people asked, then answer those questions properly as Q&A pairs.
How long does the move take?
For a site whose content is already decent: an afternoon to install, index and test, then a week of running both while you watch the gaps. The work is in the content, not the software.
What if the answers are worse after switching?
Nearly always a content problem rather than a model problem, and the entries list shows which passage produced the answer. Fix the page or add a Q&A pair, and the answer changes immediately — no retraining.