The AI overview page opens with a progress bar, a score out of four and a sentence telling you what to do next. It is a useful screen, and like most progress meters it is more useful once you know what it is counting — because the four steps are not four independent checks, and the score can be read precisely if you know how it is built.
Here is what each step measures, in the order the panel puts them, and the two things the number genuinely cannot tell you.
Step one: a billing path exists
The first step asks a single question: has anybody decided who supplies the model key? It is satisfied either by a managed plan being active on the site, or by a provider key having been saved in bot setup. Those are the two routes, and choosing between them is the first real decision in an AI rollout.
Note what this step does not do. It checks that a key is present, not that the key works. A key with a character missing, a key from an account that has since been closed, a key that was revoked last month — all of them read as ready here. The panel is tracking whether a decision has been made and recorded, which is a reasonable thing for a setup checklist to track, but it is not a connection test.
Step two: replies are switched on
The second step reads one setting: whether AI replies are enabled for the site. That is all. It is the difference between having configured a bot and having a bot that is allowed to answer anybody.
This step is worth separating from the first because the two go wrong in opposite directions. Teams who complete billing and stop here have a bot that is paid for and silent. Teams who switch replies on before adding any knowledge have a bot that is loud and empty, which is the more expensive mistake and the reason the next step exists.
Step three: at least one knowledge source is enabled
The third step counts the training sources on the site that are currently enabled. If the count is more than zero the step is satisfied.
Two consequences follow from that being a count. The first is that it is a count of sources, not a measure of content: one thin page counts exactly as much as a full site crawl. The step is a floor, not a quality bar, and the panel is honest about this in its own wording — it asks for real company detail rather than claiming to have found any.
The second is that the count only includes enabled sources. If somebody disables everything while investigating a bad answer, this step drops back to needing content and the score falls, even though nothing was deleted. That is the correct behaviour, and it is worth knowing before you go looking for what broke.
Step four is not a fourth check
This is the part that repays understanding. The fourth step is not an independent test. It is satisfied when the first three are all satisfied and there is no reply-capacity block. It is a summary of the other three with one extra condition attached.
That has a precise consequence for the number on screen. Because step four is the conjunction of the rest, the score cannot land on three out of four by having three separate things done and one outstanding — if any of the first three were missing, step four would fail too and you would be at two or lower. There is exactly one way to score three:
- billing decided, replies enabled, knowledge present — and the managed reply allowance exhausted.
So a score of three out of four is not a general nearly-there. It is a specific message: everything is configured and you are out of replies. The badge says so too, and the fourth card turns red rather than amber. If you see three, go straight to usage and capacity; nothing else is wrong.
The capacity block only applies to one mode
The capacity condition is checked for managed sites only. On a managed plan it compares what you have left — included replies plus any you have bought, minus what has been used — or, during a trial, the trial allowance against the trial replies used. Either way, when the remaining figure reaches zero, launch readiness is blocked and the recommendation changes to reviewing limits.
If you supply your own key, this check is skipped entirely, and it should be: your spend lives in your provider account and we are not metering it. But that means a site on its own key can show a full four out of four while the provider account behind it is out of credit or rate limited. The panel is not hiding anything from you; it genuinely does not have that information. On your own key, the balance you need to watch is the one on the provider's dashboard, not ours.
Ready to test means ready to test
When all four are satisfied the badge reads ready to test, and the wording is deliberate. Nothing in the four steps involves sending a message. A site can reach four out of four without a single question having been answered by the bot, correctly or otherwise.
So treat four out of four as permission to start, not as a result. The next action after a full score is to open the widget as a visitor and ask the three questions your customers actually ask, then read the AI monitor to see what the bot said. That is the check the panel is telling you to run, and it is the only one that tests the thing you care about.
Reading the panel when it looks wrong
Every figure on this screen is read defensively: if a lookup cannot complete, the page still renders and the value falls back to its empty state rather than showing an error. That makes for a robust screen, and it has one implication worth carrying. A step that reads as not done immediately after you were certain you finished it is worth re-checking on a refresh before you go and redo the work.
The recommended-next-step card underneath follows the same order — billing, then bot setup, then knowledge, then capacity, then day-to-day monitoring — so it will always point at the earliest unfinished step rather than the most interesting one. Once everything is done it settles into pointing at usage, which is the right place to be looking once the bot is live.