Evaluation criteria
Feature grids flatter every vendor, because everybody ticks the same boxes. The differences that show up six months later are structural: how your data leaves the system, how the widget is installed and updated, what the AI is grounded in, and what happens to your configuration when you change plans.
A short evaluation list you actually apply beats a long one you skim. These are the questions that separate products once the demo is over.
- Can you export transcripts yourself, in a usable format, without asking support?
- Is the install a snippet you control, and can you remove it as easily as you added it?
- Is the AI grounded in content you supply, or in general knowledge that can invent specifics?
- What is the real unit of billing, and what happens to it in a quiet month?
- Where does chat data live, and who on your side can read it?
Choosing live chat software
Evaluating vendor security
What to log in live chat
Ask us too
The point of an evaluation is not to find a product with no limits; it is to find one whose limits you can live with and can see in advance. Ask the awkward questions early, and prefer a specific answer over an enthusiastic one, whoever is giving it.
If a vendor cannot tell you where AI conversations are processed, how long transcripts are kept, or how you would leave, those answers do not improve after you sign.
- How long are transcripts retained, and can we set that ourselves?
- Which third parties process our chat content, particularly for AI answers?
- What exactly happens to our data and configuration if we cancel?
- What is the migration path in, and the migration path out?
What AI chat sends to the model provider
Transcript retention policy
Answering security questions
How a switch goes
Nobody should cut over on a Friday afternoon. The low-risk version of a migration runs both products briefly: keep the incumbent live on the pages that matter, put the new widget on a smaller section, and compare the conversations rather than the marketing.
Most of the migration work is not technical. It is deciding which canned responses survive the move, which trained content is still true, and who is going to own the channel afterwards. Export your history before you cancel anything, because access usually ends with the subscription.
- Export transcripts from the incumbent while you still have an account
- Port canned responses by hand and delete the ones nobody used
- Run both widgets on different sections for a fortnight, not the same page
- Remove the old snippet from every template, including cached ones
Migrating from another provider
Common setup mistakes
Launch checklist