What a perfect Boolean query still misses
Every analyst has done this. They write the query, run it, read what comes back, and spot a word they hadn't thought of. So they add it and run it again. Then a phrase nobody would have guessed turns up, and in that goes too.
We call this building a query. It's worth being honest about what it really is. A semantic search, done by hand, one guess at a time.
The tool has a good case. Anyone can read a Boolean string, run it again next quarter, and get the same result.
But underneath it sits an assumption we rarely say out loud: to find a conversation, you have to already know the words people used in it. The query is a guess about language, written before you've seen the evidence. So a mention count tells you how much of the internet used your words, not how much was talking about you.
And there's no error message for the difference.
The 86% that never showed up
Earlier this year we searched for one luxury automotive brand twice over the same month, once with keywords and once starting from a question, then compared the two. This is our own data and I'd rather say that plainly than dress it up as neutral research.
The question-led search found the conversation. So we asked one thing of the keyword query: how much of that same conversation had it actually seen?
On YouTube, 25%. On Facebook, 13%. On Instagram, 4%.
That last figure means the keyword search walked past ninety-six posts in every hundred that were sitting right there.
Now put those next to how much writing each platform gives a query. YouTube has a title, a description and captions. Facebook has post text. Instagram has a caption, and then the meaning disappears into a video nobody wrote down.
The results follow the writing almost exactly.
That isn't a badly built query. It's a query working perfectly, right up to the edge of what somebody typed, then stopping. Across all three platforms it saw 14% of the conversation and reported that as the whole thing.
What stays with me is that every quality check would have passed. Everything it returned was relevant. The 86% it missed produced no error message. You can check whether what you found is right. You can't check what you never found, because that would mean already knowing it was there.
What natural language actually buys you
It isn't a friendlier search box.
A Boolean string is doing two jobs at once. It has to cast wide enough to catch the conversation and tight enough to keep the rubbish out, and those jobs fight each other. Every word you add to catch more lets more rubbish in. Every NOT you add to clean up cuts real mentions too. That fight is why the tuning never ends.
Split the jobs and the fight goes away. Describe the subject, and a system can generate far more ways of saying it than anyone would sit and type. Then check what comes back against the question that was asked. Two steps, each doing one job.
Which means you start from a question about a subject instead of a guess about wording. Finding out what's there stops being the thing you finish before you search. It becomes something the search hands you.
What it costs, and what you shouldn't give up
None of this comes free.
Repeatability is the hard one. A Boolean string is something you can hand to a colleague. A question isn't, and answers can shift when the model underneath shifts. If you can't run an analysis again six months later and defend it, you're holding a finding you can't stand behind.
Knowing how far you reached comes next. If a search casts wide but nobody says where it stopped, it's unclear what the results add up to. A result you can't see the reason for it isn't evidence, it's a suggestion.
None of that is a teething problem. Some things have to survive intact: every finding traceable to its source, honesty about what was searched and what was left out, and a record of each version so an analysis holds up months later. Above all, people do the interpreting, because a system finding a conversation isn't the same as understanding one.
The question nobody builds a query for
We've spent twenty years getting very good at describing conversations before we've heard them. That worked while conversations were still being written down.
Boolean isn't losing here. It's precise, it's open to inspection, and it answers its question well. Keep it for that. Just stop letting it decide what you're allowed to find.
The most valuable question in this profession was never "what are people saying about us." It's "what are people saying that we'd never think to ask about." One of these tools can hear that. The other can't.
Queries will keep getting easier to write. Conversations will keep getting harder to predict.