Every AI answer comes with a confidence between 0% and 100%. The threshold decides what happens with it:
- At or above the threshold, a match gets the status Matched, and a no-match answer No match.
- Below the threshold, both get the status Review and, by default, a light yellow fill, so they stand out. To turn the fill off, clear Highlight rows to review in Settings (see Where the results go); the Status column and Show review rows still work.
The default threshold is 80%. Change it in Settings; the new value applies to the next run. The add-in remembers the value you set, so a threshold you changed earlier stays as it is.
Choosing a threshold #
Raise it, for example to 90%, when a wrong match is expensive: a vendor master clean-up, or lists with several legal entities of one group, where a parent and its subsidiary have near-identical names. More rows go to Review, so there is more to check, and more of the look-alikes land there instead of passing as Matched. It does not catch all of them, so still spot-check Matched rows for legal entities. The extra rule “A subsidiary is not the same as its parent company” helps with the same look-alikes. Lower it for a first rough pass: fewer rows go to Review, and more uncertain answers are marked Matched.
A quick review routine #
- Click Show review rows and go through them. These are the matches the AI was least sure about.
- Look at the No match rows. If you know a partner exists, see Why did a row get No match? On a List B with more than 200 entries, Search thoroughly checks them against every entry.
- Look at the Duplicate in List B rows. The name is in List B more than once, so the copied “(matched)” columns stay empty. Pick the right row yourself.
- Spot-check a handful of Matched rows, especially where your lists are full of similar names or hold several legal entities of one group.
Confidence is a guide, not a guarantee #
A high confidence means the AI is sure, not that it is right. Most mistakes land in Review, but not all of them: for example, when the right entry was not among the candidates the AI saw, it can answer No match with high confidence. Treat the results as a very good first pass and check them before you merge, delete or pay anything based on a match. How accurate is it? explains where it gets it wrong.