The Listicle Is the Label

There is a specific problem with adjective tags on points of interest, which is that nobody can agree on what they mean and everyone is wrong.

Ask a classifier whether a restaurant is "romantic" and you get a confidence score backed by vibes and whatever the training set thought romance was in 2021. Ask a user to rate it as romantic and you get a sparse, biased, gameable signal that mostly reflects the demographics of people who leave reviews. Ask me and I'll just make something up.

The thing I've been thinking about instead: let the internet's decade of listicle production do the labeling, and just harvest the consensus.

The mechanic is almost embarrassingly simple. Take a random adjective — "romantic," "hidden," "cozy," "underrated" — combine it with a city, and search for "Top X [adjective] places in [city]." The web has produced an absolutely staggering volume of this content. Travel bloggers, local guides, food publications, tourism boards — they've been manufacturing adjective-tagged place lists since before SEO was a word people said at dinner. The corpus is enormous and it is already labeled.

What you end up with is not a classifier's opinion. It's a provenance chain. The Candle Room in Prague is romantic because Condé Nast Traveler and three separate travel bloggers and a local Czech food site all put it on a "most romantic restaurants in Prague" list between 2018 and 2023. That's a different kind of claim than a model inferring romance from candlelight mentions in reviews.

You hold the proof. You hold the source. The confidence isn't a float between 0 and 1 — it's a count of independent claims over time, which is actually how humans decide things.

The loop-via-API version of this is what makes it interesting at scale. Spin it across a matrix of adjectives and cities and you're not just building a database, you're building the reference corpus for this category of subjective POI attribute — one that didn't really exist in a structured form before.

And then, almost as an afterthought, this becomes something else: an LLM discoverability metric. How many times does a model's training data describe a place as romantic? That's now a number you can have, because you've been systematically collecting the claims that would've gone into that training data. You end up with a proxy for how a model will answer questions about a place before you even ask.

The whole thing runs on the assumption that the listicle industrial complex, despite being mostly terrible, is not systematically lying about adjectives. Which, honestly, seems fine. They may be wrong about which brunch spot is most "hidden gem" in Austin, but they're not conspiring about it.

That's probably good enough.