How a report is made

Five passes, not one

The difference between a useful analysis and text generated at random is entirely in the number of passes. There are five here, each with its own prompt and its own model, and each with the power to stop the report from being published at all.

1. The radar, and the desk that says no

The first pass reads the news looking for events, not trends: a rule taking effect, a subsidy running out, a platform rewriting its fees, an operator going under. "The sector is growing" is not an event and does not survive this stage.

What it finds goes to a commissioning pass whose job is to reject. A candidate has to clear six tests — among them: could one person actually start this, and are there real companies that have already tried it? If nobody has ever done it, we can only imagine it, and imagining is not what this site is for. Most candidates are turned down, and the reason is recorded each time.

2. Research

The first model queries the web with real searches — up to eighteen per report — and builds a raw dossier. Its instructions are explicit: find filed accounts, platform figures, institutional sources. And hunt for failures with the same energy it hunts for successes, because an industry told only by its winners is an industry told badly.

Every figure it collects is labelled at one of three levels: verified (filed accounts, official filings, data published by a platform), stated (the company claims it but nobody audited it), and estimated (derived, with the arithmetic shown).

3. Precedents

A separate pass, with its own searches, goes looking for the companies that have already run this business — by name — and establishes what became of each one. It is instructed to spend at least half its effort on the failures, because that is the half that does not surface on its own: dissolutions, insolvency notices, founders' accounts of why they stopped, brands that quietly went quiet.

For each case it has to state a mechanism, not an adjective. "Tough market" is not a cause. "One client was 60% of revenue and left, and the fixed costs had been sized on that client" is a cause. Where the public record does not support one, the report says the cause is unclear rather than guessing at it.

Anyone can find the winners: they have websites. The dead ones are the point.

Those cases are stored as data, not just prose, which is why you can see them listed in every report. A feasibility judgement without precedents is an opinion.

4. Writing

The second model receives the dossier and the sources actually consulted, and writes the report around the strongest thesis the data supports — not a neutral summary. It may use only the URLs gathered in the previous pass; it is not permitted to produce new ones.

5. Adversarial verification

The last model has one job: demolish the report. It assumes at least one error is in there and looks for figures without a source, numbers too round to be true, revenue confused with profit, regulation cited wrongly. It re-searches the riskiest claims.

It starts with the named companies, and it checks every one of them — does the company exist, is the outcome stated correctly, do the figures match the public record? A company that does not exist would be the worst thing this site could print, so any case that cannot be confirmed is struck from the report and from the data behind it.

When a number fails verification it is removed, not replaced with another number.

At the bottom of every report you'll find how many corrections were applied. We publish that on purpose: if it were always zero, the fact-checker wouldn't be working.

Two languages, not one translation

The English and Italian versions of a report share the same research but are not translations of each other. The regulatory and tax sections differ because the constraints differ: an analysis that ignores Italian social security contributions is useless to an Italian reader, and one that dwells on them is noise to everyone else.

What never gets published

• A report with fewer than four sources is never generated.
• A report with fewer than three documented precedents is never generated — and neither is one where every precedent is a success. A list of winners only means the research found the easy half.
• A report that violates the data schema breaks the site build.
• Every report must carry a section on regulatory and tax constraints, one on who tried this before, and one on what can go wrong. Without them it is incomplete.

The limits, stated plainly

No human reads a report before it goes live. Verification is automatic and, however strict, is not infallible: it can miss an error, especially on regulation that has just changed or on very small companies with little public data.

Figures are true as at the date shown. An incentive live in March may be exhausted by September; a tax threshold can change with a budget law. Before making a decision that involves your savings, re-check the numbers at source — which is why we always link them.

And above all: this is information, not advice. A report doesn't know your situation, your skills, your city, your capital or your risk tolerance. It exists to get you to your accountant well prepared, not to replace them.

Why we publish negative verdicts too

Most content about starting a business has an incentive to tell you yes: it's selling a course, a consultancy, a piece of software. There's nothing to sell you here, so an avoid verdict costs exactly as much to produce as a promising one. If a business doesn't stand up, the report says so and shows the arithmetic. Those are the reports that save you the most.

Reporting an error

If you find a wrong figure, say so: the report gets corrected and the correction stays visible. A site that publishes daily can afford to be wrong. It cannot afford not to fix it.