Ads are a research instrument

A product can be described in two sentences that are both true and that land as entirely different offers. One names the problem the reader already feels. The other describes the mechanism, which in this category means offering to analyse your last game, roughly what every chess tool already says. Put both in front of people who have never heard of you, and what comes back indicates which sentence names a problem someone already believes they have. That is a narrow thing to learn, and it can be asked before a line of product code exists. Include the category-standard claim as a control, so there is something to measure the alternative against.

Finding out what people want has an ordering problem at its centre. Construction, the expensive part, comes first. Finding out whether anyone wanted the thing constructed comes last, when the cost is sunk and everyone has a stake in the answer being yes. Most of the founder canon exists to reverse that order, so that something is known before the building starts: the landing page for a product that does not exist, the concierge MVP where the automation is a person doing the work by hand.

Paid distribution is the most direct version of that move and the least polite. Nobody in an auction is doing you a favour. Ask chess players what they would pay for and you get generous, free opinions from people with no stake in being right. Put the same claim on a card in an auction and responding costs attention, which is the scarce thing here. What comes back is a behaviour that cost the reader something, from people the auction decided were close enough to the problem to be shown the card.

A hypothesis with a budget attached

An ad is a hypothesis with a budget attached: the creative is the variable, a line of text on a card someone scrolls past, and the auction is how it gets delivered. At this stage the only thing legible is the sentence, which description of the problem strangers recognise and which words they already use for it. Someone can notice a problem instantly and still not care much about it. The problem they long ago stopped expecting anyone to solve may get no click at all.

That makes it a research instrument that also acquires users, which is what makes the expense easy to misfile. Any users it brings in are real and worth keeping. One question sorts out which activity is being run: if a campaign returned no signups and one clear answer about which framing people respond to, was the money well spent? We would say yes.

Short at one end of the loop

None of this is new to advertising. What changed is the distance between noticing that a framing is worth testing and having it live. An agent working on the real repository can stand up a promise and put it into the auction without waiting on a design pass or a deploy window, which makes the question cheap to ask, though the ad spend still pays for the answer.

What the instrument cannot measure

An ad can show which wording gets a response, and that does not tell you which product to build. Fast, cheap tests measure what is easy to measure, and the click is the easiest thing to measure. A promise can do well in the auction and still describe a product nobody keeps using, and from where the click was counted, those two outcomes look the same.

There is a worse version of the same failure. A framing can do well because it promises more than the product delivers, and the auction will keep rewarding it for as long as it takes people to find out. Ranking framings on clicks can therefore hand you the promise you are least able to keep, sitting at the top of the report.

The instrument is also noisier than it looks. Attribution is approximate at best. Small samples produce confident-looking differences that evaporate on a second run, and the temptation to call a test finished on the day it agrees with you grows with how much you liked the hypothesis going in. Two variants can diverge for reasons that have nothing to do with the copy: the hour they ran, the audience the auction happened to find, whatever else competed for the same attention that week.

• Which of two descriptions of the same problem strangers recognise: answerable, quickly.
• Which words people already use for the problem: answerable, with care.
• Whether a promise survives contact with the product: answerable, on a slower clock.
• Whether the thing is good enough that people keep using it: not answerable here, at any budget.
• Whether the market is large: not answerable here; an auction reports on the people it found.
• What to build instead: never answerable. A test ranks the options put in front of it and produces no new ones.

Marketing as a line in the research budget

Reclassified this way, ad spend moves out of customer acquisition cost and into the research budget. A research budget goes to questions where the answer is not known and might be unwelcome. An acquisition budget goes to whatever is currently converting, which is a question already answered. Confusing the two is how a team spends a year scaling its confidence in something it never tested. There is no sales organisation here discovering demand one conversation at a time, so the two of us ask narrow questions in public instead, and take the answer when it is no.

What the instrument returns, at its best, is a ranked list of sentences people will respond to. A fair portion of any such list describes products that should not exist, ideas that are interesting for one sentence and tedious for a month.