- Ask a language model for a price and you get the midpoint of comparable products it has seen. That is a market average, not your price, and it is wrong in a predictable direction.
- Field experiments found that a $9 price ending raised demand in all three tests, with a stronger effect on new items than on items the retailer had sold before.
- That finding matters precisely because you are pricing something nobody has bought from you yet, which is the condition where buyers have least information.
- The Van Westendorp method turns four plain questions into a defensible range, and drafting it is a job AI does well.
- Price the first batch to learn rather than to optimise. A price you can explain and revise beats a number a model produced.
You have made forty of something. You know what the materials cost, roughly what your time is worth, and nothing else. The question of what to charge has no data behind it, because the thing has never been sold, and the temptation to ask an assistant is strong precisely because it will answer immediately and with confidence.
Pricing a new product is the one pricing decision with no data behind it, so it is worth understanding what that confident answer is made of before you use it.
What is a model actually doing when it suggests a price?
It is recalling what similar things cost and returning something near the middle. It has no access to your costs, your capacity or your buyer, so it reconstructs a plausible market price from descriptions it has seen, which is a different quantity from the one you asked for.
This produces a bias with a consistent shape. For a commodity, where your product genuinely is one of many, the estimate tends to be reasonable and sometimes slightly high, because the training material over represents branded retail listings rather than the cheapest sources. For anything differentiated, it tends to be low, because the distinguishing features that justify your premium are exactly the information the model does not have and cannot weigh.
So the failure mode is not random error. It is regression toward the category average, which systematically punishes the thing that makes your product worth buying. If you make a slower, better version of a common object, an assistant will price it like the common object.
None of this makes the tool useless. It makes it a research assistant rather than an oracle, and the distinction decides how you use the hour.
Does the ending of the number still matter?
Yes, and the evidence is unusually specific about when. Eric Anderson and Duncan Simester ran three field experiments in which price endings were manipulated, and reported that a $9 price ending increased demand in all three.
The second finding is the one worth your attention here. The increase in demand was stronger for new items than for items the retailer had sold in previous years. The authors read this as evidence that $9 endings work harder when customers have limited information, which also explains why retailers do not use them on everything.
Read that against your situation. A product nobody has bought from you before is the maximum information gap case. The buyer has no prior price to compare against, no memory of what it used to cost, and nothing to anchor on except the number you chose. The published abstract in Quantitative Marketing and Economics adds a third result, that the effect appears weaker when a "Sale" cue sits alongside the price, which is a useful warning against launching a new product on a discount.
Two cautions before you append a 9 to everything. The experiments were conducted in a US mail order context in the early 2000s, and ending conventions differ by market and by category: a studio ceramicist selling at 45 is making a positioning statement that 44.99 undoes. And the effect operates on information poor buyers, which means it fades as your product becomes known, exactly as the data showed.
Turning four questions into a range
The most useful pricing method for a small seller with no sales history is almost forty years old and requires no software. Van Westendorp's price sensitivity meter asks four questions and reads the answers as curves that cross.
The questions are variations on a single theme: at what price would this be too expensive to consider, at what price would it be expensive but still worth it, at what price would it feel like good value, and at what price would it be so cheap you would doubt the quality. Plotted as cumulative frequencies, the intersections of those curves give an indifference price point where equal numbers call it cheap and expensive, an optimal price point where equal numbers hit their upper and lower limits, and a range bounded by marginal cheapness and marginal expensiveness.
| Intersection | Which lines cross | What it tells a small seller |
|---|---|---|
| Point of marginal cheapness | Too cheap and expensive | Lower bound, below which quality doubt starts |
| Indifference price point | Expensive and cheap | Where opinion splits evenly, often near market norm |
| Optimal price point | Too cheap and too expensive | Fewest people ruled out at either extreme |
| Point of marginal expensiveness | Too expensive and cheap | Upper bound of the acceptable range |
The honest caveat, which the method's own documentation carries, is that Van Westendorp was solving for price rather than for demand. It tells you what a range of acceptable prices looks like. It does not tell you how many units you will sell at each point, and treating the optimal price point as a revenue maximum is a misreading that has cost people money.
Where does AI genuinely help with this?
In every part of the exercise except choosing the number, which is also the rule that holds when an agent negotiates that price on your behalf. The method is mostly drafting, collection and arithmetic, and those are the jobs machines do without complaint.
Writing the four questions for your product. The wording has to be specific to the item and free of leading language, and getting five drafts to choose between takes a minute.
Finding whom to ask. Thirty to fifty responses from people who plausibly buy this category is enough to see where the curves cross. An assistant is good at helping you list the places those people already are.
Doing the plotting. Cumulative frequencies and four intersections is spreadsheet work, and describing what you want in words is faster than remembering the formula.
Assembling the comparable set. This is the strongest use. Ask for twenty comparable products with their prices and what differentiates them, then verify every one, because this is also where a model will quietly invent a listing. Reading competitors' prices is legitimate; the question of how you may do it at scale is a separate one we worked through in what the law allows when you read a rival's prices repeatedly.
Arguing the other side. Give it your proposed price and ask it to make the strongest case that you are too expensive, then the strongest case that you are too cheap. Both arguments are useful and neither requires the model to know anything you do not.
Three pricing shapes, and which one fits what you made
Most first time sellers reach for cost plus because it is the only method with a formula. It is also the method most likely to leave money on the table, because it prices your inefficiency rather than your customer's benefit.
The three approaches are worth knowing by name, because the right one depends on a property of your product rather than on your preference.
Cost plus. Take your landed cost, add a margin, done. It fits when your product genuinely is interchangeable and the buyer is comparing on price alone, and it has the virtue of never selling below cost. Its weakness is that it is blind to demand: if people would happily pay three times your number, cost plus will never find out.
Competitive. Position relative to the comparable set, above or below depending on where you want to sit. This is the default for anything in a crowded category and it is where an assistant is most useful, because assembling and summarising the set is the bulk of the work. Its weakness is that it imports somebody else's cost structure as your benchmark, and theirs may be very different from yours.
Value based. Price against what the buyer avoids or gains. A tool that saves a tradesperson an hour a week is priced against the hour, not against the plastic. This is the method with the most upside and the most work, because you have to know the buyer well enough to name the alternative they are weighing you against.
The practical move is not to choose one. It is to compute all three, which takes an afternoon, and then notice where they disagree. Cost plus gives your floor. Competitive gives the context. Value based gives the ceiling you could defend if you chose to. A price that sits comfortably inside all three is a safe launch; a large gap between them is telling you something about the product's positioning that is worth resolving before launch rather than after.
What if your product is genuinely the first of its kind?
Then the comparable set is not other products like yours. It is whatever your buyer does today instead of buying from you, and that is almost never nothing.
Somebody buying a specialist storage box for a hobby is currently using a shoebox, an expensive generic container, or living with the mess. Those three alternatives have prices, including the price of the annoyance, and they form the real reference set. An assistant is reasonable at helping you enumerate substitutes you had not considered, which is a much better question to ask it than what the price should be.
Start from landed cost, which is not what you think it is
Most first prices are built on an undercount. Materials and time are the visible costs; the ones that decide whether the price works are the ones that arrive later and get attributed to bad luck rather than to the price.
A workable floor includes materials, your time at a rate you would accept from someone else, packaging, the shipping you absorb, payment processing, the marketplace or platform percentage, and an allowance for returns and breakage. That last one is the most commonly omitted and the most dangerous, because return rates on a new product are unknown and are rarely zero. A product priced with no return allowance at a 12% return rate is a product sold at a loss you will not notice for two quarters.
Work out that floor before you look at any comparable prices, and write it down. The reason for the order is psychological rather than analytical: once you have seen what similar things sell for, you will unconsciously reverse engineer your costs to fit, and the number you produce will be the market's rather than yours.
If the floor lands above the market range, you have learned something valuable and unwelcome, and the options are to change the product, change who you sell it to, or not make it. Pricing below your floor to match a competitor is a decision to fund their customers, and it is worth naming it that way.
Price the first batch to learn
The first price is not a commitment, it is an experiment, and the most useful thing it can produce is information rather than margin. This reframing removes most of the anxiety, because the question changes from "what is the right price" to "what will I know in six weeks".
Start at the higher end of your defensible range. Coming down is easy and has an obvious mechanism; going up on an existing product irritates the people who already bought it and requires an explanation. The asymmetry is one sided enough to be a rule.
Change one thing at a time, and leave it long enough to mean something. Enough traffic to be worth reading takes longer than most sellers allow, and the temptation to adjust weekly produces a dataset nobody can interpret. We went through how many variants a test needs before a winner is real in our piece on testing ad variants, and the same arithmetic governs a price test with the same uncomfortable conclusion about small numbers.
Watch the conversion rate on the product page rather than the units sold. Units move with traffic, which moves with everything else you are doing. Conversion on the page is closer to the thing your price is actually affecting.
Decide in advance what you will do if nothing sells. A price that is too high and a product nobody wants look identical for the first month, and the way to tell them apart is to ask the people who looked and left, not to keep cutting until something happens. Repeated cuts teach your audience to wait, and the timing of a markdown on stock that has not moved is a decision with its own logic that should not be made in the first six weeks of a product's life.
What about letting software adjust the price over time?
That is a different activity with different constraints, and it is worth keeping separate in your head from setting a launch price. Automated repricing operates on a product with a demand history; you are trying to create that history.
When you do get there, the constraints are legal as much as technical. What a shop may automate about its own prices, and where the line sits between responding to a market and coordinating with one, is genuinely bounded, and we set out where that boundary runs in our guide to what pricing software is allowed to do. The short version is that the mechanism matters more than the intent, and that shared repricing tools can produce outcomes that regulators treat as agreements.
For a first product none of that applies yet. The useful discipline is simpler: know your floor, know your range, pick a point you can defend in a sentence to a customer who asks, and write down what would make you change it.
A sentence worth being able to say
The best test of a price is whether you can explain it out loud without apologising. Not justify it with a cost breakdown, which no buyer wants, but state what the thing is and what it costs in the same breath.
If you cannot do that, the problem is usually not the number. It is that you have not decided who the product is for, and an unclear buyer produces an unclear price. That is the part no assistant can do for you, and it is also the part that, once settled, makes the number obvious. The model can build your comparable set and draft your survey and check your arithmetic, and if you are setting up the shop itself the mechanics of plans and how pricing is structured are worth getting right early. What it cannot do is tell you what you are selling and to whom, which is the only input that was ever decisive.