A friend of mine in Singapore works in F&B, and one thing kept coming up whenever we talked: a surprising number of stores run their marketing with no strategy behind it. Run the ads, keep the social accounts going, bring in an influencer. The numbers move right after. A month later the lift has gone quiet, and nothing about what worked has stayed behind.
That loop is what we’re building against: MONOMARQ, a service that uses AI to diagnose a store’s marketing. Your health has a doctor — examine, prescribe, “come back next time,” and the last result decides the next step. For a store’s marketing, we couldn’t find anyone who runs that full cycle with you, at least not around us. My friend and I had also both read Kakuritsu Shikō no Senryakuron (“Probability Thinking”) by Morioka-san of Katana, and from the first conversation we kept landing on the same question: could that way of thinking be implemented as logic, in software? That was the other starting point.
What I want to write about isn’t the service. It’s one mistake we made almost immediately, and how I got the mistake itself wrong afterwards.
The first of five competitors had gone out of business
We diagnosed the first store in Singapore and showed my friend the output. Data that takes hours to pull together by hand, AI assembled in minutes. That part impressed him.
The problem came after. The first of the five competitors we listed was a restaurant that had gone bankrupt about a year earlier. He lives there, so he caught it instantly. “That place is gone.”
Was I scared? Not really. It was a pilot, not a paying client, and he knew going in that it was an experiment. But the very first of the five, of all of them. That one stung.
The data wasn’t wrong
For a while I told this story as “we had no step where a human checks what the AI produced.” Today I think that’s half of it.
What that report shows is how a store looks right now on Google Maps, Instagram, TikTok and review sites. Reputation, ratings, the reviews people wrote. It lists the view from outside.
By that definition, a closed restaurant belongs in it. It still comes up in search, the ratings are still there, and shoppers still compare against it. As the free first report we hand out, I don’t think the contents need to change.
What was broken sat somewhere else. We had never written down what that report was claiming to show.
The reader takes it as “five competitors currently trading.” We were handing over “five competitors as the web sees them.” The gap sat between those two readings, not in the accuracy of the data. What was missing was a single caveat line: this is how it looks on the web.
On top of that, Google Maps labels closed places “Temporarily closed,” and our AI check never read that field. That one is going in as a required check. But it’s one field added, and the caveat is the real fix. And the caveat isn’t in yet. Where it goes in the first report, and how it should be worded, isn’t decided. That’s the part I’m least sure of right now.
If you’re wondering which of your data is safe to hand to AI in the first place, that’s a separate piece.
The same gap is in the report on your desk
I don’t think this failure mode is ours alone.
A monthly report arrives from the agency. Impressions up, click-through up. The numbers are correct. What those numbers were claiming to show is usually nowhere on the page. The client reads “this moved sales,” the agency handed over “response on the placement,” and neither side ever says so out loud.
Because the numbers themselves aren’t wrong, there’s nothing to object to. The next spend gets decided on top of the gap.
So the thing we’re trying to add to our own report is the same thing worth asking for as a client. “What did this number measure?” Something tied to revenue, or response on the web? A vendor who answers that gives you a report you can make the next decision with.
We use AI all over our own company — I’ve written about how we run it inside our development team — and I still missed this. In the piece on chasing AI I argued for picking one concrete problem and killing it; even inside that one problem, this is how you trip.
What we sell isn’t “campaigns that work”
What MONOMARQ sells isn’t campaigns that work. It’s the state of knowing what worked.
You run a campaign and find out it had no effect. I count that as a result. Once you know it didn’t work, you can stop paying for it, and the money and hours you free up get pointed somewhere by the last result.
“Knowing it failed doesn’t add a single dollar of revenue”
You’re right. Revenue comes from the campaigns themselves.
But go back to that store for a second. You pay an influencer and sales tick up that month. If you don’t know what worked, next month’s decision is another “feels right.” If you knew the weekday lunch crowd never moved, your next move changes.
One round, small difference. Now stack it for a year. You get a store that repeated “feels right” twelve times, and a store that repeated “last time’s result” twelve times. Twelve is nothing exotic: pay once a month for one year and you’re there.
This is also why MONOMARQ doesn’t promise “sales up X percent.” What we can promise is that you’ll know whether it worked or it didn’t. That’s the line.
Our own company was the same
And this wasn’t only a store problem. We’re a services company. We hold a weekly meeting, we look at numbers. Then I tried counting the campaigns I could say verifiably worked, and there were almost none. We had spent the hours looking at numbers, and we had never written down what our own numbers were claiming to show either.
So when we design a marketing diagnosis, and when we help a client work out where AI actually belongs in their operation, we now start by writing down what each number measures.
If MONOMARQ fails, I think it fails on distribution. At our size we can’t reach stores one by one, so the path runs through partnering with companies that already have stores under them and connecting one working example to the next. If that doesn’t connect, this fails. Either way, I’ll keep posting what we learn here.
FAQ
How do we start measuring marketing effectiveness?
Before any big framework or dedicated tool, take just your next single spend and decide one thing before paying: what you will look at once it ends. The same month last year, or weeks with the campaign against weeks without it — rough is fine. If you picked what to look at before the money moved, you can’t grade it after the fact.
Can AI diagnose marketing?
The collecting and comparing gets much faster. In our case, data that takes hours by hand takes minutes with AI. But we once put a restaurant that had already gone bankrupt into a competitor report. What was missing there wasn’t accuracy. It was one line saying what that data was claiming to show.
What should we ask a marketing vendor before paying?
Look for one thing in the proposal and in the monthly report: what each number actually measured. If it isn’t written down, ask before you sign. A vendor who has an answer to that turns even a failed campaign into material for your next decision.
Testing a new venture small, with AI? Let’s compare notes.
MONOMARQ is one example of how we test new ventures in small steps with AI. Give me 30 minutes and I’ll walk through the rest of this story from the angle of what it would look like at your company. At minimum, you’ll leave with a hint about what to test small next.
Book 30 minutes → Message us instead →
Shogo Harada原田 祥吾
CEO · Linnoedge Inc. · LinkedIn↗
Operating IT offshore development and overseas expansion support businesses across two bases: Tokyo and Vietnam. A leader who believes in “Systems over Spirit,” structuring cross-border businesses that often tend to be opaque. Committed to providing “reproducible quality” to organizations and clients rather than relying solely on individual skills.