We set out to verify 100 AI income claims. We published one.

The plan said 100 verified cases. After six rounds of research in one day we had five entries in the research sheet and exactly one on the site.

This is the record of how that happened. We are publishing it because the gap between "100" and "1" is the most useful thing we learned, and because a site that asks you to trust its verification should show you the verification failing.

What actually happened

Of the five logged: three reached our A grade for revenue evidence, two reached B. None of the three A-grade cases made it onto the site, and one of the two B-grade cases was disqualified for a different reason entirely. What survived was a single B-grade case, published without a revenue figure.

The method changed four times, and each change was forced

Round 1 — interviews cannot produce an A

Every candidate stalled at B. The reason was always identical: the only source for the revenue figure was the founder, in an interview. An editor selecting and polishing a story does not verify the number in it. Someone else retyping a figure is not confirmation.

Rounds 2–3 — changing the source fixed it

Moving to services that verify revenue through a payment-processor integration produced A grades immediately. The lesson was uncomfortable: rounds of effort against a bad source produce nothing, and the fix was not working harder.

Round 4 — an A grade says less than we assumed

Before adding the three A-grade cases to the site, we read their websites properly for the first time — earlier rounds had only checked that the URL resolved.

We had been reading "not disclosed" as "probably solo." It is not. It is undisclosed. We were looking for side-hustle cases, so we leaned that way.

An A grade says the revenue is real. It does not say a person did it alone, and it does not say you could repeat it.

All three came off the case list. They still appear in our pricing analysis, where team size is irrelevant, with a fixed notice saying they are not solo businesses.

Round 5 — our own grading rule contradicted itself

Looking for solo makers pointed at open-startup dashboards: founders who publish their own revenue. Then the rule stopped us.

Our A tier required evidence independent of the founder. Our list of qualifying evidence included "public MRR dashboards (open startup pages)" — which the founder publishes, on their own server.

Left alone, we would have awarded an A on a founder's own dashboard while printing "independent evidence exists" beside it. We redrew the line around a single question: who put that number on the screen? A processor or a platform means A. The founder means supporting evidence, however public and however long-standing.

Round 6 — "AI" in the name is not "AI" in the method

One B-grade case, AI Directories, described itself plainly on its homepage: "We manually submit your startup to 100+ directories."

The AI is the customer segment — AI startups — not the method. We had inferred automation from the name and the category. This is exactly the failure mode we had built a field for three rounds earlier, and we had not applied it to this record.

Our own mistakes, since we are the ones asking you to trust this

We guessed URLs. Twice. In round 1 we recorded a product URL that we had constructed rather than read from the source, and it resolved to a different company entirely — a marketing agency with no AI involved. We removed the case and wrote a rule. In round 3 we did it again with two more products. Writing the rule did not stop it; changing the procedure did — take the URL only from the string the source page actually returns, never from the shape of a name.

We let a hoped-for answer set the reading. "Solo status not disclosed" became "probably solo" because we wanted solo cases. Nothing in the data changed; our reading did.

The five checks you can run yourself

None of these need special access. All five come from a failure above.

1. 1

   Step 1. Open the URL, and notice where you got it

   A link you assembled from a product name is not a source. If you cannot copy it from the page that made the claim, you do not have it yet. This is the mistake we made twice.

2. 2

   Step 2. Look for a careers page

   30 secondsto rule out a solo operation

   Careers, a team page, investor logos, a compliance badge like SOC 2 — any one of these means it is not one person. This single check removed three of our cases.

3. 3

   Step 3. Ask who put the number on screen

   A payment processor or an app store is independent. A dashboard on the founder's own site is public and durable, but it is still the founder. Both are useful; only one is confirmation.

4. 4

   Step 4. Divide revenue by customers, then find the published price

   If a business reports both, the division should land near its real pricing tiers. Two of our three cases checked out this way. One does not publish pricing, so it stays uncheckable — and we label it that way.

5. 5

   Step 5. Separate what AI does from who buys

   “AI” in a company name often describes the customers. Read how the work actually gets done. One of our cases says “manually” on its own homepage.

What this cost, and what we still do not know

Six rounds in a single day, roughly twenty web fetches. The expensive part was never finding candidates — it was checking them, and every check that mattered was one we added after being wrong.

We do not know how to find solo, third-party-verified cases at scale, and we now think the scarcity is structural rather than a research problem. Getting an A grade means the revenue passes through a service that verifies it, and the people who register with those services are disproportionately preparing to sell. Real solo makers usually have only their own dashboard, which is a B by our rules.

So the honest shape of this site may be B-grade solo cases published without revenue figures — what was sold, to whom, how, and no number. Our first published case is exactly that.

We also have zero domestic Korean cases. We have not started.

The one thing worth taking from this

Every rule we now use was written after we got something wrong. That is not a sign the method is weak; it is the only way a method gets built. But it does mean one thing about anyone showing you verified income claims, us included:

Ask what their check missed last time. If they cannot tell you, they have not been checking long enough to know.