A founder and a researcher studying interview notes and an early product on a laptop.

Validation & Market Fit

What is the company trying to prove next?

A roadmap can be full while the next real decision is still unnamed. The useful question is which commitment of time, money or reputation the company is about to make, and which observation would justify it.

Marcus Startup Advisor17 min read

Contents

The next commitment is the article

Founders are rarely short of activity. A technical choice, a hiring plan, a sales hire and a fundraising conversation can all be open in the same week. The activity feels like progress because each item has an owner and a date. The decision underneath them is quieter. What, exactly, would have to be true before that time or money is well spent?

Write the commitment in one sentence before you argue about the evidence. “Hire two salespeople” is a commitment. “Spend the next quarter rebuilding the product on a new platform” is a commitment. “Open a second market” is a commitment. “Show that someone other than the founder can sell this offer to this buyer” is not yet a commitment. It is the proof the commitment may depend on. Those two sentences get confused because the work looks the same on a calendar. They are not the same decision.

This is an evergreen operating question, not a claim about a search trend. While drafting, this site’s Search Console and Keyword Planner data were not used, so this article does not invent a monthly search volume or call the topic trending. The language founders already use is enough to name the intent: what to prove next, which experiment to run, and which assumption the next spend depends on. The reader is a founder, or the person advising one, who is about to commit people or cash and wants a narrower question than “is the startup working?”

A full roadmap hides that question. It lists outputs. It rarely says what those outputs are allowed to decide. If the answer is “we will learn”, the company has not yet chosen a proof. Learning is a consequence of a test. It is not a test.

Evidence and inference do not belong in the same column

Start with what someone in the company has actually observed. A customer renewed a paid contract. Five people in the same role described the same workaround, unprompted, and named what it cost them last month. A prototype was used to finish a real task without the founder in the room. A buyer asked for a price and then went quiet. Those are observations. They can be incomplete, and they can still be written down without decoration.

Then write the inference on a different line. The belief that a neighbouring segment will behave in the same way is an inference. The belief that a technical rebuild will shorten sales is an inference. The belief that a polite meeting was demand is an inference. Neither column is shameful. The damage starts when the inference is filed as if it were the observation. A decision memo that cannot tell them apart will accept almost any story, because the story and the record have become the same paragraph.

This distinction is editorial judgement about how to use evidence. It is not a statistic, and it is not a claim that founders are careless. It is a practical response to a common pressure. Teams reward narrative. A narrative can be true and still be one step ahead of the record. The record is what you can point to. The narrative is what you think it means. Investors, co-founders and your own memory will all prefer the narrative once it has been repeated. The memo exists so the next commitment has to survive contact with the record.

If you only have inferences, say so. “We have not yet seen a buyer commit” is a usable fact about the state of the evidence. It is more useful than a softer sentence that sounds like progress. The point of the column is not to humiliate the idea. It is to stop the company spending the next quarter as if the inference had already been checked.

Work backwards until you hit the first unsupported dependency

Take the commitment you are actually considering and ask what has to be true for it to be rational. Then ask which of those conditions you have already seen, and which you are hoping. Stop at the first one you have not seen. That is the experiment. Everything upstream of it is a later problem, even if it is more exciting to discuss.

Suppose the commitment is a sales hire. The hire is rational if someone other than the founder can explain the offer, if a defined buyer will take a meeting for a commercial reason, and if the company can deliver what is sold without destroying the margin. If you have not yet seen the buyer take the founder seriously, a second seller will not answer that. If buyers do take it seriously but delivery is bespoke and uncosted, the hiring question is premature in a different way. The commercial question is the one in A useful product is not automatically a viable company. Hiring into an uncosted promise scales the confusion.

Suppose instead the commitment is a rebuild. A rebuild is rational if the current product is blocking a behaviour you have already seen users try to complete. It is not rational merely because the codebase embarrasses the team. Embarrassment is real. It is not, by itself, evidence that customers are waiting on the new architecture. Name the blocked behaviour. If you cannot name it, the rebuild is a preference, and preferences can wait behind the unresolved dependency.

This way of working is slower in the meeting and faster in the quarter. It will disappoint anyone who wanted the roadmap to bless every current workstream. That is the point. Parallel work is justified when the streams do not depend on the same unanswered question. It is waste when three teams are building answers to a question the company has not yet earned the right to ask.

What Graham records about Airbnb, and what he does not

The useful case here is not a slogan about hustle. It is a dated account with numbers, written by someone who was in the room. On 10 December 2020, Paul Graham published The Airbnbs. The same essay is on his own site. He wrote it around Airbnb’s listing, to explain what he thought had been unusual about the founders during Y Combinator. It is his recollection and his interpretation. It is not an Airbnb financial statement, and it is not a method that every marketplace can copy.

Graham’s account of the early target is specific. For Brian Chesky, Joe Gebbia and Nathan Blecharczyk, “ramen profitability” was 4,000 US dollars a month: 3,500 dollars for rent and 500 dollars for food. They taped that goal to the bathroom mirror. Graham’s advice, as he tells it, was to start growth in the hottest subset of the market rather than to average the whole idea. Asked where demand already was, the founders knew from searches: New York. They went there. A large part of making listings better was photography. Joe and Brian rented a professional camera and took the pictures themselves.

Two observations in the essay matter more than the camera. When they came back, Graham asked what had surprised them about hosts. The surprise was how many hosts were in the same position the founders were in. They needed the money to pay their rent. Graham places that in the recession, which he describes as the worst in decades, and says it had hit New York first. The photographs were a tactic. The trip also changed what the founders understood about the supply side. That is the part teams usually drop when they retell the story as “do things that do not scale”.

He is also careful about the shape of the early numbers. In late January 2009, about three weeks into the batch, the figures began to creep up. He says it was hard to know whether that was growth or random fluctuation. By February he judged it to be real. The fee figures he gives for the first three weeks of February are 460, 897 and 1,428 US dollars. He writes that Brian emailed him on 22 February to say they were ramen profitable, and that Brian’s reply to Graham’s warning about the following week was seven words: “We are not going to slow down.”

Airbnb fee revenue in the first three weeks of February 2009
Week 1460Week 2897Week 31,428
Airbnb fee revenue in the first three weeks of February 2009
PeriodUS dollars of fees
Week 1460
Week 2897
Week 31,428

Paul Graham reports these as the fee figures for the first, second and third weeks of February 2009. They come from the founders’ numbers, as he records them, not from an audited filing. Source: Paul Graham, The Airbnbs, 10 December 2020. The bars use these disclosed values only. Nothing has been estimated or smoothed.

Those three bars are not a model of Airbnb. They are the three numbers Graham reports, drawn to scale against each other and repeated in the table underneath so the chart cannot drift away from the source. There is no trend line, no extrapolation into March, and no claim that the slope continued. A chart that added later weeks would be fiction.

What that record is allowed to decide

Read as evidence, the essay supports a narrow set of claims. During that winter, the founders had a numeric survival target and they knew it. They chose New York because searches, not a brainstorm, said demand was already there. They did the unscalable work of taking the photographs. The weekly fees Graham records rose across those three February weeks, from 460 to 897 to 1,428 dollars, and he reports that the founders themselves announced ramen profitability on 22 February. He also records that January’s movement was ambiguous. The founders did not treat the first uptick as proof.

The essay does not support the claims people often hang on it. It does not show that professional photography is the cause of Airbnb’s later scale. It does not show that visiting customers is always the right experiment. It does not give a conversion rate, a host count, or a take rate. It does not prove that every marketplace has a single city where effort will compound. Graham is explicit that this was a way to get growth started in something like Airbnb, by concentrating on the hottest pocket. “Something like Airbnb” is his boundary, not an invitation to generalise to every software company.

It is also one person’s later account. The fee figures are the figures he says the founders gave him. A founder using this essay in a board meeting should describe it that way. “Graham reports that the weekly fees were these amounts” is accurate. “Airbnb proved that doing things that do not scale produces a billion-dollar company” is a different sentence, and the essay does not earn it. The IPO is outside the evidence he is presenting. Survival of the company for a decade is outside it too. Plenty of earnest teams did concentrated early work and did not become Airbnb. The essay is not a denial of that. It is a description of one case in which the early proof was concrete, local and numeric.

The practical transfer is the shape of the proof, not the tactic. They named a threshold before they celebrated. They distinguished a flicker from a run of weeks. They used an existing signal, search demand for New York, to choose where to spend themselves. They let contact with hosts change their explanation of why the supply side existed. A founder can copy that discipline in a business that will never need a camera.

A decision memo you can finish this week

The memo is short because it has to be used while the decision is still open. Six lines are enough.

The commitment. One sentence, with a verb and a cost. “We will hire a seller at this salary” or “We will spend six weeks on this migration.” If the sentence has no cost, it is a wish.

The observation. What has already happened, who saw it, and when. No adjectives that smuggle in a conclusion. “Three logistics managers showed us the spreadsheet they used last Tuesday” is an observation. “The market is desperate for this” is not.

The inference. What you believe those observations mean, written so that a colleague can disagree with it without rewriting the facts. The disagreement belongs here. If the team cannot disagree with the inference, you have probably hidden it inside the observation.

The unsupported dependency. The first condition the commitment needs that the observation does not yet cover. This is the only thing the next experiment is allowed to be about. If you list five, you have not chosen.

The threshold, set before the result. A number, a behaviour, or a refusal condition. Graham’s account is useful here because the threshold was almost domestic. Four thousand dollars a month, written down, taped to a mirror. Most companies do not need that particular number. They need some number or behaviour that was chosen before the week went well.

The decision the result is allowed to trigger. Proceed, change the offer, stop, or gather a missing fact. If every result leads to “continue”, you do not have a test. You have a ritual.

This memo is an editorial tool. It is not a research instrument with a published error rate, and it should not be described as one. Its value is that it forces the commitment, the record and the threshold into separate sentences. Teams that already do this in their heads do not need the form. Teams that keep sliding from a story into a hire do.

A labelled hypothetical, not a case study

The following company is invented, so that the memo can be shown in full without borrowing a real firm’s numbers. Northline is a two-person team in Manchester. They have a scheduling tool for independent physiotherapy clinics. Eight clinic owners have said the diary is a mess. Two have forwarded the calendar they actually use. None has been offered a price. The team wants to hire a developer and a seller, and they want to raise a small round to pay for both.

Commitment: spend four months of salary on a generalist seller and a second developer. Observation: two clinics shared a real calendar, and the pattern of double-booked rooms was visible in both. Inference: clinic owners will pay a monthly fee to stop those clashes, and someone who is not a founder can arrange the conversations. Unsupported dependency: no buyer has been asked for money, and no one has watched a non-founder run the conversation. Threshold, chosen on Monday, before any more calls: four clinics, in one city, agree to a paid six-week pilot at a stated price, with the practice manager as the buyer, or the team does not hire. Decision rule: if four buy, hire the seller and keep the second developer on a short contract tied to the pilot’s delivery work. If two buy, change the offer and do not hire. If none buy, stop the raise and rewrite the problem.

Nothing in that paragraph is evidence about physiotherapy software. The prices are not market data. The only thing the example is allowed to illustrate is the order of the sentences. Northline’s mistake, before the memo, was treating “owners agree the diary is painful” as permission to hire. The pain is a reason to design a paid test. It is not the test. A team that wants the commercial standard for that test can use the questions in the validation stage of the startup journey, and then the separate question of whether the price can carry delivery.

Set the threshold before the result arrives

Graham’s January comment is the part worth copying. The numbers had started to move, and he says it was hard to tell growth from fluctuation. That sentence is a better operating rule than the February celebration. A single good week is cheap to produce and expensive to believe. A threshold chosen afterwards will be set just below whatever happened. The company then “hits” it, and the commitment goes through on a result that was never required to clear a bar.

Write the bar in the same note as the method. Who will you approach. What will you count. What will you refuse to count. A compliment is not a yes. An email address for a launch note is not a purchase. A pilot that is free because the founder was embarrassed to name a price is not evidence of the price. Those distinctions are argued more fully in Compliments are not a purchase order. They belong in the threshold, not in a later argument about whether the test “really” failed.

Also write the stop. Founders resist this because a stop feels like a verdict on the company. It is a verdict on one commitment. Airbnb’s founders, in Graham’s telling, were not running a test of whether travel marketplaces could exist. They were trying to reach a stated monthly figure in the city where demand had already shown itself. A failed week would not have proved that strangers will never share a home. It would have proved that this effort, in this window, had not yet produced the figure on the mirror. The decision available at that point is revise the effort, change the pocket of the market, or stop spending. It is not “the idea is dead” and it is not “the idea is proven”.

Timeboxes help only if they are attached to the threshold. “We will know in three weeks” is meaningless if every outcome is a reason to continue. “We will know in three weeks because four paid pilots is the bar, and we will not hire if we are short” is a decision. The length of the window should match the behaviour, not the team’s anxiety. A procurement cycle that really does take a quarter should not be judged in a fortnight. A consumer signup that can happen today should not be given a quarter of polite waiting.

Proceed, revise, or stop

Proceed when the observation now covers the dependency you named, at the threshold you set, and the commitment’s cost still makes sense against that result. Proceeding is not a licence to add three further bets that were not in the memo. The original commitment goes ahead. The next dependency gets its own memo. This is how a company avoids turning one piece of good news into a hiring plan, a rebuild and a raise.

Revise when the problem still looks real but the offer, the buyer or the pocket of the market was wrong. Revision is a new test, with a new threshold, not a quiet edit of the old one. If Northline’s clinic owners will not pay the pilot but will pay for a done-for-you tidy of next month’s diary, that is a different product and a different cost. The team should write it down as a different commitment. They should not call it the same test with a better explanation.

Stop when repeated contact undermines the inference you needed. Stopping a path is a success if it happens before the salaries and the raise. It will not feel like one. The memo helps afterwards, because the team can see which observation changed their mind. Without that, a stop becomes a mood, and moods get reversed by the next encouraging conversation.

Gather more evidence only when the test itself was inconclusive for a reason you can name. Too few of the right buyers. A broken prototype that never let the behaviour occur. A price that was never actually offered. “Gather more” is the outcome founders choose when they do not want the other three. Make it expensive to select. Say what the missing fact is, how you will get it, and what you will do if it does not arrive. An open-ended research phase is usually the old commitment in a softer coat.

None of these outcomes requires a large sample or a claim about a market. Early proof is local. Graham’s New York is the reminder. The hottest pocket is allowed to be the whole test. A result there does not automatically transfer to the next city, the next buyer or the next price. Transfer is the following question. It deserves its own threshold rather than a footnote.

What to do on Monday

Open the commitment you are closest to making. If you cannot say it in one sentence with a cost, you are not ready to collect evidence for it. You are still choosing. Choosing can take an afternoon. It should not take a quarter of building.

Under it, list the observations you would be willing to show a sceptical co-founder. If the list is mostly inferences, you have a research task, not a scale-up task. Pick the earliest missing condition and design one bounded attempt to see it. Use a real buyer, a real price or a real behaviour, and a bar you could fail. Put the bar where the team can see it before the results come in. The bathroom mirror was not a metaphor in the essay. It was a place the number could not be quietly revised.

Then decide, in advance, which of the four outcomes you are allowed to reach. Tell one person who is not the author of the plan. A threshold that only the founder remembers will move. A threshold that has been read back by someone else is harder to edit on Friday.

If the result is strong, spend only what the memo authorised. If it is mixed, change the offer rather than the story. If it is weak, stop that path while the company is still small enough to stop. The quality of the week is not how busy it looked. It is whether the company can point to the sentence it was trying to prove, the record it accepted, and the decision it actually made.

The advisory work around this is deliberately narrow. A Focused Session is for the commitment already on the table, not for a general endorsement of the roadmap. Bring the sentence, the observations, and the spend you are about to approve. The useful outcome is a proof you can fail, and a decision you can explain afterwards without rewriting the week.

Sources

  1. Paul Graham, The Airbnbs, Y Combinator, 10 December 2020 Graham’s own account of Airbnb during Y Combinator’s winter 2009 batch.
  2. Paul Graham, The Airbnbs, paulgraham.com The same essay on Graham’s site, including the New York photography passage and the February fee figures.
  3. Marcus Startup Advisor, A useful product is not automatically a viable company The commercial question that often comes after the next proof.
  4. Marcus Startup Advisor, Compliments are not a purchase order How to tell a kind reaction from a commitment.

customer research · MVP