Dynamic pricing for a small shop, and the line not to cross

Two things get called dynamic pricing and only one of them is a normal way to run a shop. The first is moving a price because the world moved: stock is low, the season turned, a supplier raised a cost, a competitor changed a list price. Shops have done that forever and nobody objects.

The second is showing different prices to different people based on what you know about them. That is a newer activity, it is technically easy now, and it is being examined by regulators in a way the first one is not. Conflating them is how a small seller talks themselves into something they would not have chosen deliberately.

What did regulators actually find?

That the raw material for person level pricing is already assembled and sold, and that shoppers do not know. The Federal Trade Commission ordered eight intermediary firms to hand over material on their pricing products in July 2024 and published preliminary findings the following January.

The list of signals it found in use is worth reading slowly, because most of it is not what people picture when they imagine personalised pricing.

Alongside the obvious inputs of location and demographics, the FTC reported that behaviours ranging from mouse movements on a webpage to the products left unpurchased in a cart can be tracked and used to tailor pricing. The examples it gave are concrete: promotions targeted by skin type and skin tone, and higher priced baby thermometers shown first to people profiled as new parents.

The scale detail is the one that matters for context. The intermediaries whose documents the FTC examined, among them Mastercard, Accenture, PROS, Bloomreach, Revionics and McKinsey, worked with at least 250 retail clients between them. This is not a hypothetical capability being trialled. It is infrastructure with customers.

Does any of that apply to a shop with one person running it?

The capability does, cheaply, which is the uncomfortable part. You do not need an intermediary to show a different price to a returning visitor or to somebody who abandoned a cart yesterday. That is a cookie and a conditional, and any competent storefront can do it.

Which means the question is not whether you can, but whether you should, and the honest answer for a small shop is almost always no. Not primarily for legal reasons, though those exist, but because the discovery cost is asymmetric. If a customer finds out you charged them more than their friend for the same item, the explanation never lands well, and small businesses trade on exactly the trust that destroys.

Why does an automated repricer fight your discount badge?

Because in the EU the reference price for any announced reduction is the lowest price you charged in the previous thirty days, and your own repricer keeps lowering it.

The rule is straightforward on its own. A price reduction announcement must indicate the prior price, and the prior price is the lowest price at which the item was available in the 30 days before the reduction. Goods that deteriorate quickly and goods on the market for less than 30 days are treated differently. It applies in shops and online alike.

The Court of Justice then closed the gap most retailers were standing in. In a case decided on 26 September 2024 concerning advertised discounts on fresh produce, the Court held that a percentage discount, or a promotional statement that stresses how advantageous a price is, must be calculated on that same 30 day prior price rather than on any higher reference. Saying "20% off" is an announcement of a reduction, even without a struck out figure beside it.

Now put an algorithm in the middle. A repricer that responds to stock and competition will lower your price several times a month, sometimes briefly. Each of those lows becomes the number your next promotion is measured against. A shop that cut to 39 for two days in week one cannot advertise 20% off 59 in week four, because the prior price is 39. The tool that improved your margin day to day has quietly shrunk every headline discount you can honestly claim.

Read the last two rows together, because they are the counterintuitive result. Personalised reductions and loyalty prices are excluded from the EU disclosure obligation, so the regime that governs your public discounts says nothing about them. That is not a permission. It is the reason the other regulator is looking there.

One more asymmetry worth keeping in view. The 30 day reference rule constrains what you may claim about a reduction, not what you may charge. Nothing stops you selling at 39 whenever you like. What you cannot do is describe 39 as a saving against a figure you were not really charging, which means the rule polices your marketing copy rather than your pricing decisions. Shops that treat it as a pricing constraint end up avoiding sensible markdowns for no reason, and shops that treat it as nothing at all end up with a badge that is quietly untrue on every seasonal campaign.

What can a small shop safely automate?

Anything driven by facts about the product rather than facts about the person. That single test resolves most cases without needing a lawyer.

Stock level. The oldest and best signal. Slow moving stock ties up cash and shelf space, and a rules based markdown ladder handles it without drama. This is also the one where a forecast genuinely helps, and the shape of that problem is covered in our piece on what stock forecasting can and cannot tell a small shop.

Cost and margin. When a supplier price or a shipping cost moves, the price should follow. Automating that is bookkeeping rather than pricing strategy, and it prevents the slow margin erosion that comes from forgetting.

Time and season. Predictable demand curves, end of season clearance, day of week patterns for perishables. All of it is about the product and the calendar.

Competitor list prices, with one caution. Matching a published price is ordinary competition. Using a shared pricing service that ingests your competitors' data and recommends prices back to all of you is a different structure, and the fact that a third party algorithm made the decision does not make coordinated pricing acceptable. If a tool's selling point is that everyone in your category uses it, that is the feature to ask hard questions about.

How often should prices actually move?

Less often than the software would like. A repricer can update hourly, and for most small catalogues that produces churn without benefit while destroying your discount headroom.

Weekly is enough for most physical goods, with exceptions for genuinely perishable stock and for categories where competitors move constantly. The reason to slow down is not caution for its own sake. It is that every downward move sets a new floor for the next thirty days of promotions, so frequent small cuts convert into permanently smaller advertised discounts. Deciding how often you reprice is really deciding how much promotional headroom you keep.

There is also a customer perception cost that nobody models. A shopper who saw 45 on Tuesday and 52 on Thursday does not conclude that your algorithm is responsive. They conclude the price is arbitrary, and the next thing they do is wait, which is the opposite of what the repricer was bought to achieve.

How do you know your price is wrong?

Not from a model and not from a competitor. From your own conversion rate at a given price, measured over a period long enough to mean something.

The two failure signals point in opposite directions and both are common. A product with strong traffic and weak conversion, where the traffic is arriving on the right search terms, is usually priced above what that audience expects. A product that sells steadily and never generates a single question, complaint or comparison is often priced below what people would have paid, and the tell is that nobody hesitates.

Neither observation is worth acting on from one week of data. Small catalogues produce small numbers, and the temptation to reprice on a quiet Tuesday is how a shop ends up with a price ladder nobody designed. Set a period, hold the price for all of it, and compare like with like. A price test that runs alongside a promotion is not a price test.

Worth noting that the traffic mix changes what a conversion rate means. If a growing share of your visitors arrive from an assistant that already told them your product suits them, they convert better regardless of price, which is the pattern behind the finding that AI referred shoppers convert better than other retail traffic. Read your conversion rate by channel or you will credit a price change for a traffic shift.

What does a price rise actually cost you?

Less than most small sellers fear, and the fear is expensive. Underpricing is the more common error in businesses run by the person who makes the thing, because the maker knows how long it took and prices the hours rather than the value.

The arithmetic is worth doing once, properly. If you raise a price by 10% and lose 10% of unit sales, revenue is roughly flat and your margin per unit is materially better, so profit rises and you do less work. The break even loss for a 10% rise is larger than instinct suggests, and it grows as your margin shrinks. A shop on thin margins gains the most from a rise and is usually the most frightened of it.

What a rise does cost is the goodwill of existing customers who notice, which argues for timing rather than for avoidance. Raise prices when something changes that the customer can see: a new supplier, better packaging, a version change, a cost increase you can name. Announcing a rise with a reason is a normal business communication. A rise that appears silently is the one that reads as opportunism, which is the same trust question that runs through the personalisation argument above.

What about pricing advice from a model?

Useful for structure, unreliable for numbers. Asking a model what a fair price is for your product produces a confident figure derived from nothing about your costs, your market or your customers.

Where it earns its keep is in the reasoning around the number. Working out which competitor set is genuinely comparable, listing the attributes that justify a premium, drafting the bundle options, checking whether your price ladder has an obvious hole in it: these are structured thinking tasks with your own inputs, and a model is a decent partner for them. The judgement of whether 44 or 48 is right stays with you, because only you know what the item costs you and how many you have.

The bigger risk is subtler and worth naming. Ask a model to set prices across a catalogue and it will produce a plausible, internally consistent ladder that quietly reflects whatever pricing conventions dominate its training data. That is fine if your business is conventional and actively harmful if your positioning is not, because it will regress you toward the middle of a market you may have deliberately chosen not to sit in.

Is there a version of personalisation that is fine?

Yes, and it is the one shops have always used: an offer somebody can see, understand and qualify for. A member price, a first order discount, a returning customer voucher, a volume break.

The distinction is not personalisation versus none. It is whether the customer knows the rule. A loyalty price is a published rule that happens to produce different prices for different people. Surveillance pricing is an unpublished inference that does the same thing. One a customer can explain to a friend; the other only makes sense from your side of the screen.

Applied honestly, that test also settles most of the marketing adjacent cases. A discount code in an abandoned cart email is fine, because it is a visible offer with a reason. Silently raising the displayed price because the visitor's behaviour suggests they will pay it is not, and the fact that a tool can do it automatically is not a reason to.

The practical setup

Three things, in this order, and none of them require a pricing product.

Record every price change with a timestamp, per item. This gives you the 30 day reference figure on demand and turns compliance from a research project into a lookup. Second, compute your advertised discount from that recorded low rather than from your list price, and have the calculation happen where the badge is generated rather than in somebody's head. Third, decide your repricing cadence deliberately and write it down, because the default of whatever the tool does is a decision made by a vendor about your promotional strategy.

All three of those are small pieces of logic that sit next to your catalogue, which is the point at which the shape of your shop starts to matter. On a platform where price history is not exposed and the discount badge is rendered by a theme you cannot reach, they are a support ticket. Where the storefront is code you control, they are an afternoon, and the same rule change that costs other sellers a month costs you a morning. That is the unglamorous everyday case for building a shop you can change yourself, and pricing rules are where it shows up most often, because they move on a legislature's schedule rather than a product roadmap's.