Mifanr 中文

A product that did nothing wrong, and how it died

September 7, 2026

A few days ago I read a due diligence report on a micro-SaaS. The product turned Instagram comments into direct messages automatically, and it was listed for sale at $7,000.

The numbers: live since March 1, and by early September it had $310 in monthly subscription revenue across 18 active subscriptions, with $695 in lifetime revenue. It shipped 39 product updates, passed the platform’s official review, and connected two payment providers. On its six core keywords it ranked on the first page zero times out of six, at a domain rating of 4. Its free tier of 300 messages a month was the lowest in its category. It went up for sale on August 24 at $12,000, and the price was cut five times in eight days down to $7,000. Four offers came in. It had not sold.

Every individual step was fine. That is what made me stop.

It died from something that never sets off an alarm

A product doing the wrong things gets negative signals. Users arrive and leave, features get complaints, conversion looks bad. All of that hurts, and all of it tells you where the problem is.

This product got silence. The demand was real, and about twenty competing products crowded onto the same keyword, which is the proof that demand existed. The product worked; it passed platform review. The price was reasonable. Nobody could find him.

The whole category was manufactured. In March, the upstream platform cut its free tier from a thousand contacts to twenty-five. A wave of users had to find a replacement at the same moment, and about twenty products poured into the gap. Spotting that window is not the hard part. Getting there before everyone else, carrying distribution nobody else has, is the hard part.

The two kinds of mistake differ by an order of magnitude

Pick by your own advantage and get it wrong, and you fail fast and loud. You thought you understood the problem and nobody bought. You know within a month, the cost is tiny, and the answer needs no interpretation.

Pick by market data and get it wrong, and you fail slowly and quietly. Market data is usually correct. The demand is real, twenty competitors prove it, and you did build something that works. You receive no negative signal at all. You receive silence. Six months, 39 updates, then a 42% price cut in eight days.

The two mistakes differ by an order of magnitude in time and money. So the question is not how to find the right demand. It is how to discover you are wrong while it is still cheap.

Market data clears mines; it doesn’t pick the field

“Which industry is good, which category is good” is a question that cannot work, and it is self-defeating: the answer is the same for everyone who asks it. Any promising category you can look up is already crowded the moment you look it up.

The same data works fine with the question reversed. That founder did not need anyone to tell him what to build. He needed to know that twenty products already sat on his keyword, that his free tier was the worst in the category, and that a search-led acquisition path does not work arithmetically in a category that crowded.

Research after you choose is mine-clearing. Research before you choose is suicide. The difference is only in the ordering.

The handbook I read had already corrected itself once on this. Someone argued publicly that for a business under a hundred million, market data means nothing, because in the end it comes down to you understanding something others don’t; a business that looks to everyone else like tending a grave is one you work with ease. The argument is right, and it has a trap in it, which comes later.

Five veto questions that cost nothing

This is the part I cared most about when turning the handbook into a skill. A handbook is a document, people skim documents, and what gets skimmed is usually the cheapest step.

The five vetoes are questions. No web access, a few minutes:

  1. Can you find it with one search? Yes means the ground is taken. You would be fighting a domain rating of 60 for the first page.
  2. Is the core mechanism one API endpoint belonging to someone else? Yes means you don’t set the ceiling. That product was capped by a platform rule allowing one reply per comment, so multi-step follow-up was not “not built yet”. It was structurally impossible.
  3. How long does switching take? Under ten minutes means no moat. That product’s entire state was five keyword-to-message pairs.
  4. Does the free tier cover 80% of what users need? Yes means a more generous free tier will kill you. A competitor in the same category offered unlimited.
  5. Do you have distribution nobody else has?

The first four can be dodged by changing what you build. The fifth cannot. A product where every piece of information about it was published by you does not exist.

The expensive part goes last

The handbook’s conclusions came from an eight-track parallel investigation that gathered 229 evidence-backed facts, covering the site, pricing and checkout, the SEO footprint with six keywords tested by hand, platform data, the founder’s public trail, about twenty competitors, and the underlying API’s official docs and policies.

That stage costs three orders of magnitude more than the five questions above. In a document they sit as two sections of equal weight. As a skill the order can be enforced: the first three stages cost nothing, a veto stops the run, and the investigation never starts.

The investigation stage has its own ordering, by signal-to-noise:

One class of data deserves caution. Search Console, keyword tools, and search suggestions have a structural blind spot, and it fails in reverse for newcomers: it only shows terms where you already have impressions, and terms you can’t rank for never appear. That failed product ranked zero of six on its core keywords, and those six terms were nearly empty in its own dashboard. The deeper problem is that a need has to be compressed into a search term before it can enter this data at all, and an unmet need has no name yet.

Other people’s pessimism is not a signal

Back to the trap. “A business that looks to everyone else like tending a grave” describes two kinds of people at once: the person with an asymmetric advantage who sees value others can’t, and the person actually buried in the graveyard. From the outside they look identical, down to the story they tell themselves.

The variable is whether you have the thing others don’t, and whether it can be verified. The handbook gives two tests, and you need both:

If you can’t name them and can’t reach them, that isn’t understanding others lack. That is using the phrase to reassure yourself.

I ran it on myself and stopped at the fifth question

The day I finished the skill I used it on a few directions I had been considering. All of them stopped at question five.

This site went live yesterday. The X account posted its first thread today. My distribution is close to zero. By this method, I should not be looking for a product right now.

That conclusion was more useful than I expected. It says the work at this stage is not thinking up products, it is growing distribution, because distribution cannot be bought and cannot be dodged by picking a different problem. Writing, posting, doing the arithmetic in public: that is the work.

This isn’t bad news. It is the priority order, correctly assigned.

If you only remember a few things

  1. Market data clears mines; it doesn’t pick the field. Where you go is decided by your asymmetric advantage. Research after you choose is mine-clearing; research before you choose is suicide.
  2. Run the five free veto questions before you spend anything on research. One search finds it, the core mechanism is someone else’s API, switching takes ten minutes, the free tier covers 80%, no distribution of your own.
  3. The fifth is the only one you cannot dodge by picking a different problem. If distribution is zero, build distribution, don’t hunt for demand.
  4. Search the grammar of complaint, not the noun for the need. “We wrote an internal script for it” means the pain is big enough that someone spent time, not yet big enough that someone shipped a product. That gap is the opening.
  5. Search data is for validation and mine-clearing only. By definition it cannot see a need that has no name, and the newer you are, the more of it is noise.
  6. Two- and three-star competitor reviews beat keyword data, because they prove the person paid.
  7. “Others think it’s a bad business” is not a signal on its own. Name three things others don’t know, reach twenty real users in a week. Miss either one and you are reassuring yourself.

Two things to state plainly

The evidence base for this method is a single case. Every conclusion in the handbook was reverse-engineered from one due diligence report. Its value is in taking one failure apart in enough detail. Use it as a checklist, not as a law.

I corrected the numbers in this piece once. The first draft said “five months, 36 updates, sold for $695”, copied from the handbook, and all three were wrong. Checking against the original fact base: 39 updates, not 36; $695 is lifetime revenue, not a sale price, and the asking price was $7,000; and it had not sold as of the audit. The launch date conflicts too, with the changelog saying March 1 and the platform profile self-reporting April. I use the former. I am leaving the correction here because it is exactly what this method asks for: downgrade the evidence level of secondhand material by one, including a handbook I wrote myself.

I have not used it to make a real decision. The skill was written today, the only subject it has been run on is me, and the verdict was that I shouldn’t build a product. Whether it works will be clear after it kills or clears a real direction. When that happens I will write up the process, including whatever I got wrong.

Failed products and their founders appear here only in aggregate, never by name. The figures come from a publicly available due diligence report.

Get new posts by email

One or two a month.Or subscribe by RSS. Reading in Chinese?中文版