Kanchen's Secret Sauce

Darjeeling and Kalimpong do not have a tidy list of homestays anywhere. Properties open, close, get renamed, get built above villages you cannot spell without help. The first job of the pipeline is to walk the map and find them all.

We do this by tiling the two districts into a grid of search cells and sweeping each cell with a set of local queries in multiple scripts and spellings. Dense pockets get re-tiled with tighter grids because a single search only returns so many results before it caps out - if we did not tile, entire ridgelines would silently disappear. Discovery runs in parallel across regions, which is why a full sweep of both districts takes hours rather than days.

The bar to enter the index is boring on purpose: an actual place on the map, in the hills, that anyone can visit. New candidates get added continuously and old ones get re-checked so a homestay that opened last month is not waiting a year for its turn.

This is where most ranking sites quietly give up. Big chains and mid-range hotels outspend homestays on listings, so they flood the results. If you do not filter them, your "homestay guide" turns into a hotel directory in six months.

Kanchen filters in two stages. First, a deterministic pass that recognises brand vocabulary and hotel-shaped patterns and blocks them cold - Ramada, Sterling, Sinclairs and their cousins do not make it past the door. Then, for the genuinely ambiguous cases (the boutique retreat that calls itself a "stay", the guesthouse-hotel hybrid), an AI adjudicator reads the listing the way a person would and decides. Every decision is cached, so we do not re-litigate the same property every week, and the cache is transparent to us for audit.

The result: the guide stays a homestay guide, and the "unscored" pool stays honest. When a hotel does slip in - and one always does - a single sweep re-classifies it and it is gone from the public surface within the hour.

For every property that clears the filter, we pull up to 500 guest reviews from the last 24 months. Any language people write in. Older reviews get archived, not deleted - they stop counting toward the score but stay in the record. When a new review lands, the oldest one drops out. The window is deliberate: ownership and management in these hills change faster than aggregate ratings can track, and a five-star review from 2019 can actively mislead you about the place today.

An orchestration of AI and APIs then reads every review and extracts structured signals from unstructured prose. "The climb from where the taxi dropped us was rough with bags" becomes a Path signal. "Woke up to Kanchenjunga glowing pink" becomes a View signal. Not keyword matching - meaning, in context, across scripts. This is the piece we would have had to hire a small army to do by hand, and the piece that lets us cover roughly a thousand properties every week.

Every property gets a score from one to five on five pillars. These were not invented in a workshop - they emerged from what guests kept spontaneously writing about, over and over, across thousands of reviews.

• Shelter - rooms, cleanliness, warmth, upkeep. Weight: 25%.
• Water - hot water reliability, plumbing, supply. Weight: 20%.
• Hearth - food, hospitality, the warmth of the hosts. Weight: 25%.
• Path - road access, last-mile, arriving with luggage. Weight: 15%.
• View - what you actually see at first light. Weight: 15%.

Shelter and Hearth carry the most weight because they are what a stay is: a bed you want to sleep in, and people you are glad to spend an evening with. Water is next because when it fails, nothing else matters. Path and View are real but conditional - a rough road is survivable if the rest is right, and even the best view cannot rescue a cold room.

Because the pillars are scored independently, no single strength papers over a weakness. A run of five-star View reviews will not save a property whose Water pillar has collapsed.

Here is the part almost nobody does, and the part we care about most. When the AI scores a pillar, it also reports how confident it is in that score - high, medium, or low. Twenty recent reviews all describing hot water is a high-confidence Water score. Three vague mentions is a low-confidence one. Pretending those are the same is how bad rankings get made.

So confidence enters the math directly, in two ways:

• Damping. A low-confidence 5.0 gets multiplied down before it hits the composite. High confidence keeps the full weight, medium loses a little, low loses more. A confidently-scored 4.8 will out-rank a shakily-scored 5.0 - which is how it should be.
• Tier demotion. If a property has too many low-confidence pillars, it is capped out of the top tiers entirely, no matter how flattering its individual numbers look. Two shaky pillars drop it from Deep to Standard. Four drop it from Deep all the way to Limited. The Deep band is genuinely high-confidence only, and Summary is reserved for the thinnest evidence.

We then label every property into one of four confidence tiers. Deep means 20 or more recent reviews read directly. Standard means 10 to 19. Limited means 5 to 9. Summary only means fewer than 5 direct reviews, so the score is inferred from Google's own property summary. The tier is shown right next to the score, because a number without context is a number without meaning.

The raw pillar scores are still shown honestly on the property page - a five is a five, with its confidence chip next to it. But the composite you see at the top, and the rank the property holds in the guide, are earned from real evidence.

The ranking follows a simple, unbreakable sequence:

1. Confidence tier first. Deep properties sit above Standard, Standard above Limited, and Limited above Summary-only. A slightly higher composite never leapfrogs a stronger evidence base.
2. Composite score within the tier. The weighted, confidence-damped composite decides who leads.
3. Volume of recent reviews as tie-break. When two properties are neck and neck, the one with more recent evidence wins.

That is it. No boost for anyone who talks to us. No demotion for anyone who does not. Money never moves a rank on Kanchen, ever - the algorithm cannot see who has paid us, because nobody can pay us for placement.

• Take paid placement, in any form, ever.
• Manufacture a decisive-looking score from four reviews.
• Hide the confidence tier behind a friendlier star rating.
• Rewrite or republish guest reviews. The AI reads them; it does not put words in anyone's mouth.
• Pretend to have data we do not have. If evidence is thin, we say so on the card.