The neighbourhood already has everything. It just cannot be read.
A manifesto for LOCAL: a platform built on what this city's small businesses can actually offer, confirmed by the people who run them. What it is, what it refuses, what it does not yet know.
Walk down Noriega and you pass bakeries, restaurants and shops whose signs you cannot read. The food is probably exactly what you would want. You will never find out, because the sign is in one language, the menu is in another, and there is no version of that shop that exists in the language you search in. Multiply that by a city and you have the whole problem.
Now add the second half. The bakery that opened four months ago makes the best bread on the block and is invisible to search, because the only signal that scales on large platforms is accumulated reviews and clicks, and that signal runs against anyone new. And the third half: the restaurant is open to private dinners and says so nowhere, because the owner considers it obvious, the server has no authority to answer, and the customer does not know what to ask.
Three failures of the same kind. The capability exists. The demand exists. Nothing in between can read either one.
That gap is the entire product. Not another directory. A layer of confirmed knowledge about what a local business can genuinely offer, to whom, and on what terms, readable by a neighbour and by another business alike.
01 / Why this, why here
San Francisco is the hardest and the best place to try it
The hardest, because the incumbents are not idle. Google now does agentic search and calls businesses on your behalf. Yelp has licensed its corpus to a frontier model and ships an assistant across categories. The delivery platforms own the transaction, and the reservations platforms own the table. Every one of them is aiming at the same consumer surface.
The best, for a reason that is easy to miss: not one of them owns what a merchant can actually do. Maps reads menus, hours and structured feeds. It does not have a field that says works with small bakeries, takes orders from eight units, delivers within the city. Not because it could not, but because acquiring that requires a conversation with every owner, and a company optimised for global coverage will not have those conversations.
And this city has the raw material no other American city has in the same concentration: an extraordinary density of small, independent, multilingual merchants within a few square miles. Family bakeries, taquerias, salons, florists, machinists, roasters, kitchen contractors. A fabric of capabilities that is real, valuable, and almost entirely unreadable to anyone outside it.
$150/moYelp Ads, the entry price of attention
3-4%monthly churn in small-business software
150MToo Good To Go users, still not profitable
120MBeli ratings, built by a team of about five
$81MChowbus Series E, for point of sale and AI marketing
Read that band carefully. Attention is expensive and rented. Merchant tooling is already funded and crowded. And the two consumer products that worked locally did not win on features: Too Good To Go won by turning an update into a transaction, Beli won with a personal list people refuse to abandon. The lesson is not to build more software. It is to own data nobody else has bothered to collect.
02 / The mechanism
One knowledge base, two audiences, no new platform
The mechanism is a single one: the capabilities of a local business, confirmed by its owner, connected to the intent of whoever is looking. And the one looking may be a neighbour or another business. A bakery needs a particular flour, in small batches, delivered. A cafe needs a local roaster. A dinner organiser needs a restaurant open to events. Same data, same search, same assistant. Only the intent and the criteria differ.
01
The assistant talks with the owner and turns the conversation into structured fields: what they make, which orders they accept, who they work with, on what terms, from what volume, in what lead time, where they deliver, what constraints apply. One conversation of twenty to forty minutes at onboarding yields a dozen fields.
02
Every field is confirmed by the owner and published with a date and a subject: what exactly was confirmed, and as of when. A capability claimed but never exercised for a third party sits behind a label of stated by the owner, never presented as fact.
03
All of it enters one index: the permanent profile, the collaboration capabilities, the live layer. The live layer is stored as a field with a lifetime, not as a post. A fresh batch ready in a few minutes is an attribute of a result, not an item in a feed.
04
One intent router recognises who is asking and applies one of two rankings to the same data. A neighbour cares about distance, fit and freshness. A business cares about capability match, terms, volume, delivery area and lead time. The unknown is neither hidden nor invented: the system says not known yet, and offers to ask.
One mechanism, two audiences
This is a query-side layer, not a new product. The same fields serve both audiences; only the ranking differs. That means one intent classifier and one ranking profile instead of a second platform, a supplier marketplace or an agent framework. Architecture follows the product, not the other way round.
03 / Nine principles
What we will not do, stated as rules
These are not aspirations. They are the operating constraints, and everything else in this document exists to protect them.
P1
A better match is not for sale. Not in consumer search, not in business results, not later. Relevance is set by a person's need, not a business's budget.
P2
No paid placement in business results at all. Not even labelled. There a person is deciding a purchase for their company, and any payment for placement zeroes trust in the entire capability graph.
P3
AI invents nothing. Not prices, not stock, not services, not terms. An unknown field stays empty or explicitly marked, never filled with a guess.
P4
Only what the owner explicitly confirmed is published. No capability appears from the assistant's suggestion unless the owner repeated it in their own words.
P5
Confirmation is the origin of the data, not a guarantee of quality. We say so in the interface, beside the badge. Otherwise confirmed is read as good, and trust is built on a misreading.
P6
An empty profile is not published. The catalogue grows more slowly than the number of merchants walked, and that is the correct order. One empty profile devalues the badge for everyone else.
P7
The free profile is enough to take part in search. You pay for freshness, detail, tools and hiring. A free tier that cannot reach results is not a tier, it is a demo.
P8
Live content is voluntary and its absence carries no penalty. No counter reading twelve days since your last post. Convenience turned into a duty stops being convenience.
P9
The reputation window shows the recent; it does not hide the old. Freshness means prioritising the last three months, with earlier history one click away. The moment a window hides history it is the same rating with a different number.
Two of these cannot be broken once
P1 and P2. A single sold position, or one paid slot in a procurement decision, ends the only asset this product has: trust that what is shown is shown on merit.
04 / Refusals
What LOCAL is not
NO
Not Maps or Yelp. It does not compete on venues, reviews or traffic. The difference is that a buyer no longer assembles information from six sources and still ends up uncertain.
NO
Not a directory. Empty profiles are not published, and participation is not a condition of being shown. An unconnected business appears when it best answers the query; the connected merchant's advantage is detail, not position.
NO
Not a marketplace at the start. Order, payment and booking stay in the merchant's own system, and LOCAL points there. Integrations come later, if their value justifies the difficulty.
NO
Not a social network. Live content is not a feed, views are not the goal, and nothing is mandatory. Video of the work lives in the profile for months; video of fresh baking lives in the live layer for hours.
NO
Not an ad platform. Organic position is not sold in any form, and business results carry no placements at all, labelled or otherwise.
NO
Not a translator. Language is how meaning is carried, not how words are swapped. The original name stays beside the explanation, and general knowledge about a dish is separated from confirmed facts about one shop.
NO
Not a closed club. Search needs no registration, personal data stays under the user's control, and an account alone never proves a visit: influencing reputation requires a trace of an enquiry.
05 / What is actually an asset
Everything copyable is not an advantage
Intent search, explanation, translation, the map, saves, personalisation, video: a large platform can reproduce all of it in a release or two. Only what cannot be obtained by indexing and cannot be bought with money is protected.
What copies in a release, what accumulates by hand
Asset
Copyable in a release
Accumulates only by hand
Intent search, explanation, translation
Yes
No
Map, saves, personalisation
Yes, except the person's own list
The list accumulates with the user
Confirmed availability, corridor by corridor
No, requires walking the street
Yes
Capability graph: terms, volumes, partnerships
No, requires a conversation with every owner
Yes, and it serves two sides at once
Neighbourhood relationships and trust
No
Yes
Density beats coverage, and this is the hard part
The capability graph is valuable only where it is dense. Incomplete capability data is worse than none: a professional calls three suppliers, finds none suitable and does not come back. So build one category wholly, not evenly across a city. Every bakery in the Sunset, plus every flour supplier that serves them. A business query must find an answer, or it finds nothing and never returns.
06 / Economics
How it pays for itself, in what order
The shape of the cost is the good news. The onboarding conversation happens once and yields a dozen fields, so cost per merchant is front-loaded rather than weekly: more expensive to bring in, cheaper to keep. For a business that runs on relationships, that is the right shape, because most of the work lands where a person is present anyway.
01
The founder's own labour. Setting up and running a profile for a one-off fee, reference 150 to 500. Available immediately, before any product exists, and the only income depending on neither audience nor technology.
02
Hiring and job listings. Local businesses hire constantly, the pain is acute, willingness to pay is real: job boards charge $75 to $400 per posting. Entrepreneur-to-entrepreneur value, independent of any consumer audience.
03
Business introductions. The strongest line. A consumer result is worth cents; a business relationship repeats and is worth thousands a year, and the query arrives with money behind it. Two honest forms: a one-off fee for an introduction or a verification, or a commission on a first transaction. Selling position here is banned by P2.
Tools come second year: automation, integrations, analytics, at $20 to $70 a month with 3 to 4 percent monthly churn, sold by the hundred. Advertising exists only on the consumer side, labelled and contextual, and never in business results. Merchant tooling is already funded and crowded; the tool is the entry, not the product.
07 / What we do not know
The honest list
A strong document differs from a handsome one by naming the unknown, so here it is in full.
01
Whether LOCAL will find a buyer. The main untested risk. No correction to supply-side costs removes it: onboarding merchants more cheaply is not the same as finding people who need them.
02
Whether density can be reached. Without density the capability graph is worse than its absence, and this is the one risk that kills the business layer outright.
03
Data discipline. Confirmed stock decays in hours, permanent fields in months, and somebody must hold it. The live layer is optional, so it will be patchy, and a buyer cannot count on it.
04
The volume of business queries. At fifteen merchants it cannot be measured. In the pilot the business layer is a bonus signal, not a criterion.
05
Reputation at scale. Designed now, tested later. Complaints in the pilot will be zero, because users will number in the dozens.
06
The product's aesthetic. An app you open just to explore your own city cannot be checked by research, only by a prototype with real people.
08 / The test
Six weeks, fifteen merchants, five queries
01
Ten to fifteen independent merchants from different districts, prioritising recently opened businesses willing to experiment. Recruited through Instagram, personal contacts and neighbourhood communities, framed as a B2B partnership rather than a plea for help. Free to start, with no promises of customer flow.
02
A small working version, not a full app. The first two weeks matching queries to offers by hand, to collect real phrasings, then the small app on the same queries. Manual matching tests the conversation; the working version tests result quality.
03
Five queries as the test set: a minimalist vegan cake under $150, a really good pastrami sandwich nearby, a hard gel pedicure technician, an unusual bouquet from a local florist, which new cafes have opened. Each run through Maps, AI Mode or ChatGPT, Yelp and Instagram, then through LOCAL.
04
Relative thresholds, never absolute: not how many queries there were, but for how many of them LOCAL produced a concrete confirmed offer that none of the four large players did.
05
Spread the supply across the city, concentrate the demand: a person in Bernal Heights will not cross town for a bakery in the Sunset.
06
Measure: the share of queries with a concrete confirmed answer; the share of honest nothing suitable; repeat use (an anonymous device counter, since entry needs no registration); the share of confirmed results among those shown; enquiries and orders reaching merchants.
07
Measure the business layer separately: how many merchants asked for a supplier and how many received an answer they acted on. Never mix the two.
The kill criterion, named in advance
If, out of forty real queries, the combination produces a confirmed result that none of the four large players gave for fewer than a quarter of them, the combination does not work. That is an answer, not a reason to keep going.
And the boundary that keeps this honest: viability is tested by merchant money, realisation by result quality, scaling by a second neighbourhood launched by a different person. This pilot answers only the second. Expecting all three is a planning error that produces a verdict of did not work where the wrong question was simply asked.
09 / The ask
What this needs to be proven
01
Someone who walks. One person, six weeks, fifteen conversations in one or two neighbourhoods, on their own streets, in their own languages. Not a sales hire. A neighbour with a notebook and a phone.
02
One category first, wholly. Every bakery, cafe or salon in a defined area plus the suppliers that serve them, so a business query finds an answer instead of emptiness. Depth in one place beats a thin layer everywhere.
03
Forty real queries, run twice. Once against the four large players, once against LOCAL, with the difference written down. That number decides whether this is a company or a note.
LOCAL connects not only a neighbour with a business, but the capabilities of some businesses with the needs of others, on data the owner confirmed themselves. Everything above is a way of not losing that simple thing.