Who AI names when you ask for a local business, and who it never names

Project summary

I run a small engine that asks the same question a customer would ask, and writes down whether a business gets named. So far that's 1,175 saved answers across 30 real businesses in South Florida — auto shops, dealerships, dentists, med spas, a dermatology practice. Median business gets mentioned in 40% of answers. Ten of the thirty land under 30%. One auto shop came back at zero. Nobody is publishing this at the neighborhood level, so I'd like to: 100 areas, several trades each, everything raw and free.

What are this project's goals? How will you achieve them?

Here's the plan concretely. Take a zip code, ask best dentist, best fast food, best body shop, whatever the trade is there. Then the next zip. A hundred of them. Same method every time. Published raw and free, so a small business owner or a franchise owner can look and see where they stand and where they don't.

That's who this is for. Not agencies. I already tried agencies and it did not work, I'll tell you about that below. Owners.

The zero in my data is the auto shop where I volunteer on Thursdays. Nineteen answers on its own business name and not one mention. The assistants didn't even work out it was an auto shop, they read the name like a normal English phrase and answered about fast cars. Add the city and the word reviews, eleven of sixteen. All Grok and Perplexity. ChatGPT absent, Gemini nothing.

So it's findable. If you already know how to ask. Customers don't know how to ask. That's the whole thing.

Two things in here I haven't seen anyone measure. First, the models disagree with each other, constantly. Five assistants, same question, different businesses named. If that holds at a hundred areas then AI visibility isn't one thing, it's five things wearing the same coat. Second, drift — same questions, same subset, monthly for six months. A model gets retrained and the named set moves overnight and nothing outside the labs records that it moved.

There's a second instrument, same idea aimed at developers. I built it because I wanted to shift the plane a bit from what I already knew. An agent picks its own SDK and I wanted to understand how that works. 87 runs, a coding agent told to add a feature to an empty project. It doesn't recommend a library, it installs one, and it's in your lockfile before anyone reads the diff. Seven of nine task areas, every completed run wrote the same package. next-auth, stripe, resend, drizzle, inngest. One area picked its own vendor's SDK, which is a model preferring its maker's product. Armature published a much bigger study in September 2026 and where our tasks overlap their findings match mine, so I'm not claiming first. What's different is they sell getting picked by agents, so their corpus and their money sit in the same place. This one is supposed to sit apart from that.

Now the story of why I'm asking for research money instead of trying to sell this again.

Okay so. Ten weeks ago I started this thing. I'll tell you the whole story because I think it explains why I'm asking for research money instead of trying to sell something again.

It started easy, honestly. It started with just building. I taught myself how to structure the architecture properly, learned the technical parts I needed. The build itself took a week. Another week for fixes. This was over the summer.

The original purpose was, well — make my own project, make a name for myself. You know the thing, 17 years old, make a name, make money, start something serious. I've never really wanted to work at some company. It always felt to me like, not cool. I never wanted to picture myself that way.

So two, three weeks went into development and figuring out the plan. First I was selling to small businesses, thought I'd make a case study with my sister. Realized the problem: getting data constantly is very hard, it's much easier when you take it yourself. When you run it yourself. So I thought, I need whoever closes that problem, whoever does it themselves. Marketing agencies — they already have clients. I could just supply it white label. Great, I thought, white label, everything's chocolate.

Then the question came up: is that even a good offer for them? So I started improving the output, the audit itself, so it would show what to do and where, for agencies who deal with this stuff daily. Source mapping and all of that. Got it to a decent state. Then wanted to improve it more, so I made it smarter, more precise, more integrations, more algorithms. Got it to 200 answers.

And then I started knocking on what I already knew, which is cold email.

It did not work. Bought a domain, Google Business Profile. Sent maybe 20 emails. Later sent about 50 more. Zero responses. Zero, on hundreds of sends. I think it becomes effective at thousands, but definitely not at hundreds.

In parallel those weeks I was trying to survive through other social channels. Loosened my grip a bit. Some amount of time later, I stopped tracking time honestly, I discovered LinkedIn and started trying to contact people, writing comments. Built myself four machines so they could post for me constantly.

Started going to meetups. Talked with an owner of a marketing agency, in person. He said "yeah, sounds fine, do it." I didn't even ask him for anything, I wasn't selling.

Then a few weeks later I discovered what Sales Navigator actually is, subscriptions and all that. Maybe 20 days ago I started writing. Realized I was limited because I'd bought the free trial. Sent roughly 17 to 20 messages. And it worked, sort of — people answered. I'd also learned how to write properly by then: first I ask about some detail, not from personal interest in selling, and I wait for the actual reply before going warm. Then I try to sell it sensibly, not like an idiot. Got something like 30 to 45 percent answer rate. Which for cold contact I know is good. And even when people complained about their problem, I offered my solution for free. Let me show you free, let me do it free.

One guy agreed. Alexander. He said the topic is interesting but for his market, his clients, it wasn't relevant, because he works with Russian speakers and for them regular Google is more understandable.

Later I worked out something for myself: this is the same as SEO, only more concentrated, more specific. And I understood that even if I want to do a fix-for-you for someone, it ends with — I need social media. If it's tied to a zip code. And for digital businesses social matters even more, you need followers. Digital is actually harder than local, because the weight of the platforms themselves matters, plus popularity, plus how often you're cited.

So I realized that, and then the main puzzle started. For about two weeks now, and still today, I'm breaking my head over what to do. It came to me out of millions of conversations, millions of burned tokens: maybe I'm selling wrong, maybe I'm selling to the wrong person.

Here's the thing I got wrong and I'd rather write it here than have you find it. Agencies don't buy audits. It's in their playbooks — audits and research are free, or at a loss. Execution is what's paid. I made the audit paid and ignored execution completely.

Competitors: Semrush is about $200 a month, I checked that myself. Other chats told me Profound gives away nearly everything for free, but I have not verified that, so I'm not claiming it.

I also burned twenty dollars plus a subscription on a freelance marketplace. Without reviews it's pointless. There are people there for whom ten dollars a day is good money, and you show up with no reviews, and you go nowhere.

The one thing that did work: my name. From constantly registering on things, if you google me, even Gemini says specialist, entrepreneur. It's a small thing.

Ten weeks. Nobody has paid me anything. Two grant rejections so far, about fifteen grants and twenty scholarships sent, some still open.

What actually exists: 21,000 lines of code. 1,175 saved live answers across 30 real businesses in South Florida — auto shops, dealerships, dentists, med spas, dermatology. 87 coding-agent runs on the developer side. And an integrity layer, thirteen deterministic checks that re-verify every claim against the raw answer on disk and block the client report if any check fails. The name matcher gets written twice, second time from scratch, so a bug can't agree with itself.

So the instrument works. The business around it doesn't. That's basically the whole reason I'm asking for this to be public research rather than going out and trying to sell it a third time.

I don't want to leave this as a loss.

How will this funding be used?

Okay so, the money. I'll be straight about the parts I'm not sure of.

Right now everything runs off my laptop. That's fine for thirty businesses. It is not fine for a hundred areas measured monthly for six months, because the laptop gets carried around, closed, put to sleep, reconfigured. I already have runs in my data with absurd durations for exactly that reason, the machine went to sleep in the middle. If the monthly waves don't go out from something that just sits still, the months aren't comparable and the drift series is worth nothing.

So a dedicated always-on machine, $800. That's the line I'd defend hardest and it keeps running after the grant money is gone.

The runs themselves. One job is eight questions, five repeats, five assistants, about 200 saved answers. My one full receipt came to $2.07 and I've been working with about $2 a run. Budgeting $2.50 to leave room for longer question sets and for prices moving.

Baseline corpus, 500 runs at $2.50 = $1,250

Drift, 200 runs times 6 monthly waves at $2.50 = $3,000

Subscriptions, $140 a month for 7 months = $980

Dedicated machine = $800

Storage and bandwidth for the public corpus = $150

My time, 7 months at $571 = $4,000

Contingency = $320

Total = $10,500

Subscriptions are $140 a month because two of the five assistants can't be read through an API in a way that reproduces what a real user sees, so those go through paid consumer tiers. And the coding agent runs on a flat monthly subscription rather than metered billing, which is why there's no per-run cost line for the developer side. What it eats is subscription months and machine hours, both already in the table.

The line for my own time is the one I expect you to push on, so here it is plainly. Ten weeks in I have 21,000 lines of code and zero revenue. I've been funding this out of whatever was left after the bills, roughly $400 of my own money over the past year, subscriptions and a domain and about $27 of metered spend. That is the entire budget this has ever had. $571 a month is what makes this the work I do instead of the work I get to on weekends.

At the $1,500 minimum I run a reduced corpus, maybe two hundred businesses instead of the full hundred areas, publish the code and the replication guide, and drop the drift series. The instrument survives. The six-month question doesn't get answered.

Every run logs its own cost and that log gets published with the data, so you can check this budget against what actually happened.

Who is on your team? What's your track record on similar projects?

Okay so. Ten weeks ago I started this thing. I'll tell you the whole story because I think it explains why I'm asking for research money instead of trying to sell something again.

And then I started knocking on what I already knew, which is cold email.

Started going to meetups. Talked with an owner of a marketing agency, in person. He said "yeah, sounds fine, do it." I didn't even ask him for anything, I wasn't selling.

Competitors: Semrush is about $200 a month, I checked that myself. Other chats told me Profound gives away nearly everything for free, but I have not verified that, so I'm not claiming it.

The one thing that did work: my name. From constantly registering on things, if you google me, even Gemini says specialist, entrepreneur. It's a small thing.

Ten weeks. Nobody has paid me anything. Two grant rejections so far, about fifteen grants and twenty scholarships sent, some still open.

So the instrument works. The business around it doesn't. That's basically the whole reason I'm asking for this to be public research rather than going out and trying to sell it a third time.

I don't want to leave this as a loss.

What are the most likely causes and outcomes if this project fails?

My thirty businesses are all South Florida and I picked most of them myself. Auto shops, med spas, dentists. That's not a random sample, it's a sample of who I could reach. So the concentration I'm seeing might be an artifact of who I measured. Thirty is enough to notice something. Not enough to claim it. That's the honest state of it, and it's the main risk.

Access is the second one. Two of the five assistants can't be read through an API the way a real user sees them, so I go through paid consumer tiers, and those get rate-limited or closed. One of the five has a defect I haven't solved: it answers from memory, no web access, zero citations. The flag saying a search happened only means one was requested. I label it memory-only in my reports and I'd label it the same publicly.

Third, the schedule. Waves have to go out monthly or the months aren't comparable. School gets heavy, machine dies, something slips, and there's a hole in the series.

Fourth, nobody uses it. I write the guide, release the code, no one runs it on their own city. This is the one I'd bet on and the one I can do least about. The data still exists either way.

I turn eighteen in October, before this page closes. That's why the deadline sits where it does.

How much money have you raised in the last 12 months, and from where?

Nothing that counts as raising. No grants, no investors, no institutional money.

Revenue: zero. Ten weeks in and nobody has paid me anything. One local business has had the work for free, which I'd rather say now than have it read later as a hidden customer.

It's gone the other way. Roughly $400 of my own money over the past year, mostly assistant subscriptions and a domain, about $27 of that metered model spend. I also burned twenty dollars plus a subscription on a freelance marketplace trying to find clients, and learned that without reviews you go nowhere there. That is the entire budget this work has ever had.

I've applied to around fifteen grants and prize programs this year for my work generally, plus about twenty scholarships. Two rejections so far, Emergent Ventures in September being one. Several still open. None of them were for this specific project.

One application for this specific work is coming. Hacker Initiative opens in October and I intend to ask them for $6,500 toward the same measurement. If that lands, or anything else does before this page closes, I'll post it here, say what it covers, and reduce what I'm asking for.