Pick a crate of lager, shandy or cider and we will knock 50% off a box of crisps. Take the crate
Recommendations

Six things it uses. Nine it refuses to.

The recommendation engine, written out in full, because a recommender that will not explain itself is indistinguishable from one that is doing something you would object to.

6 signals used
9 refused
0 third-party data
1 switch to stop it
Both lists in full

What the engine sees, and what it is not allowed to.

The right-hand column is the interesting one, and the last line of it is the one that matters most.

Uses

  • What you have actually bought here, and when
  • What you have said you like, in your own preferences
  • What is genuinely in stock this week
  • What is in season on this farm right now
  • What people who bought the same three things also bought
  • Which of the three businesses you have never used

Refuses to use

  • Anything inferred about your health
  • Anything inferred about a pregnancy
  • What you looked at on any other website
  • Data bought from a broker, ever
  • Your precise location. The parish is as fine as it gets
  • Your age, beyond the yes-or-no of an age check
  • Your name, gender or anything guessed from either
  • Whether you have ever complained
  • How much we think you would tolerate paying
The last line: no personalised pricing. Ever.
The red line

The last line: no personalised pricing. Ever.

The natural extension of a recommendation engine is a pricing engine. Once a system knows that somebody buys a jar of chilli jam every six weeks without fail, the commercially rational move is to stop discounting it for them, and to discount it heavily for somebody who has never bought one.

That is legal, it is widespread, and it is the single fastest way to destroy the trust of a customer base built on a television programme about honesty. A price here is the price. It is the same price for a member who has spent four thousand pounds and for somebody arriving from a search engine for the first time.

What changes with membership is a published, flat discount — five, ten or fifteen per cent depending on tier — which is the opposite of personalised pricing because everybody can see the rule.

0
personalised prices
3
published discount rates
1
price per product
How it is built

The six signals, weighted.

SignalWeightWhat it actually does
Your own order history Heaviest Suggests the thing you buy running out, and stops suggesting the thing you returned
Your stated preferences Heavy Six checkboxes in Faffing. Beats every inferred signal, always
Stock Absolute Nothing out of stock is ever suggested. This is a veto, not a weight
Season Heavy In August it suggests the tomatoes. In February it does not pretend to have any
People who bought the same three things Light Enough to surface a jar nobody would have found. Not enough to steer
Businesses you have not used Light Fires once, ever. If you ignore it, it does not come back
Cold start

Why the cold start is a person rather than a model.

On a first visit the engine knows nothing, and a recommender that knows nothing will confidently recommend the best-seller, which is how everybody ends up with the same jar of chilli jam.

So the first suggestion anybody gets is not a recommendation at all. It is a quiz — five questions, written by the person who runs the shop counter, based on the five things she actually asks people who stand in front of the shelf looking lost.

It outperforms the model for the first three orders. After that the order history takes over and the quiz answers become just another stated preference, weighted heavily and never overwritten.

Four rules

Explainable, every time

Every suggestion carries a one-line reason. "Because you bought this in March."

Dismissible for good

Dismiss a suggestion and it does not return. Not in a fortnight, not ever.

No third-party data

Nothing bought, nothing brokered, nothing from another site.

One switch off

Turning it off leaves you with the shop in the order the shop is in.

Wondering

Suggesting, questioned.

Because a suggestion without a reason is asking you to trust a black box, and because writing the reason forces us to notice when the reason is embarrassing.

Entirely, in one switch. What you get instead is the shop in its ordinary order, which several people prefer.

No. Three million people follow the farm shop and we have no idea which of them are members, which is by choice rather than incompetence.

Only to accounts that have passed an age check, and never as a first suggestion to somebody who has never bought any.

Everything we hold about you.

Field by field, with where it came from and how long it lasts.

Next along the route

On its way

Carry on
Basket Book the pub Plot a field Join the list Bother us

Snooping, briefly.

We use a small number of cookies to keep your basket upright and to work out which pages people actually read. We do not sell anything about you to anyone, because that would be a rotten way to behave.

Read the detail