How matching works

1. Your sentence becomes structure

A query like “me and 5 mates want somewhere busy around St Kilda tonight, decent food, beers, not too expensive” is parsed into the things that actually decide where you end up:

• Party size — including phrasing like “me and 5 mates” (six) or “table for 8”
• Location — a named suburb, or your device location if you share it
• Timing — “tonight”, “brunch”, “after drinks”, a day of the week
• Atmosphere — busy, cosy, quiet, fancy, and phrases that rule things out (“not too fancy”)
• Food and drink — dishes, cuisines and drink styles
• Budget — “$60 each”, “under $35”, or a general sense like “cheap”
• Occasion, dietary needs, seating and venue type

The results page shows you this parse under “Read as”, directly beneath your sentence and above the answer. If we’ve misunderstood something, you should be able to see it immediately rather than work it out from twenty bad results.

Today this is a rule-based parser: a large vocabulary of phrases mapped to structured meaning. It is fast, predictable and easy to correct. It sits behind an interface designed so a language model can take over the parsing without anything else in the system changing.

2. Every venue is scored across twelve dimensions

Each venue is scored on food, drinks, atmosphere, location, party-size suitability, occasion, budget, opening status, venue type, dietary coverage, seating and reputation.

A dimension only counts if you asked about it. Weights come from your query, not from a fixed formula. If you said nothing about budget, an expensive venue is not penalised for being expensive. This is why two different searches can rank the same two venues in opposite orders, and both be right.

Atmosphere is scored two ways

Venues carry both vibe tags and measured positions on three axes: noise and energy, formality, and intimacy — each from 1 to 5, shown on every venue page. Tags alone are too brittle, because “lively” and “busy” describe the same room. Axes alone lose the specifics that tags capture well, like “rooftop” or “sport on the TVs”. So both are used, and blended.

Party size is a real constraint

Every venue records the largest group it can seat together, whether it takes bigger bookings, and the size at which it starts requiring one. A table of eight will not be recommended a twelve-seat wine bar just because the atmosphere fits.

3. When we show a percentage, and when we don’t

A percentage implies a measurement. MenuMap only shows one when the query gave it enough to measure against — four or more scored dimensions. Below that you’ll see a qualitative band instead (“strong match”), and for a near-empty query, no match indicator at all.

A single-word search like “pizza” could easily be dressed up as “94% match”. It would be meaningless. The more you tell MenuMap, the more precise it is allowed to be about how well something fits — and it says so rather than inventing confidence.

4. Why you get one answer instead of thirty

Australia is not short of good restaurants, and nobody arrives here short of options. What people are short of is a reason to pick one — so a page of thirty ranked results hands the hard part of the job straight back to the person who asked. MenuMap leads with a single recommendation, at a size nothing else on the page competes with, and states the reasons it won in the same terms the query was scored on.

Underneath it sits a short row of alternatives, and these are not simply results two, three and four. Each one is chosen because it differs from the recommendation on a different, nameable axis — cheaper, closer, quieter, livelier, open now, more room, smarter — and is labelled with that difference. The reason for the shape is that the next question after “where should we go” is almost never “what was second”. It is “what if I wanted something a bit cheaper”, and that is answerable in one word.

A difference has to be large enough to notice before it earns a label. Somewhere four dollars a head cheaper is not “cheaper”, and somewhere two hundred metres nearer is not “closer”. An alternative also cannot win its axis by failing the request in some other way: it must still be a good answer, and it will not be offered at all if taking it means a substantially longer trip than the recommendation. The full ranked list is always one click away, sorted by match, distance or price.

5. What never affects ranking

• Payment. Placement is not for sale. No venue can pay to rank higher, appear in a collection, or be featured.
• Claiming. Claiming a venue does not move it up. Keeping an accurate menu can help you match more searches — because you match them, not because you claimed.
• Commercial relationships. Booking or ordering links do not influence position.

Reputation — where a rating exists — contributes a small amount as a tiebreak between venues that fit your request equally well. It never outweighs whether a place actually matches what you asked for, and ratings from a handful of people count for less than ratings from thousands.

5. When we widen a search

If a query is narrow enough that almost nothing qualifies, MenuMap widens it — including places that are currently closed, or venues in neighbouring suburbs — and tells you at the top of the results that it has. It will not quietly pad a page with weak matches and let you assume they were what you asked for.

6. Where this is still weak

Being straight about the limits:

• The parser is vocabulary-based, so genuinely unusual phrasing can be missed. Unmatched terms are logged so the vocabulary can be improved.
• Atmosphere ratings are editorial classifications, not sensor readings. A quiet Tuesday and a packed Saturday are the same venue.
• Menu prices change more often than any catalogue can track. See data and accuracy for how we handle that.