A template privacy policy describes an app you did not build. Write yours feature by feature
We rewrote the privacy policy for our delivery apps one feature at a time. Every honest sentence about how long we keep something turned into work. Here are the four questions we ask per feature, and what answering them found.
You are getting an app ready for Google Play. The listing asks for a privacy policy URL. You paste your company name into a generator, publish the result, and move on. The store accepts it. Nobody reads it.
That document is not useless. It is worse than useless, because it is written from the outside in. It lists categories: "location data", "device identifiers", "usage information". Your app does not collect "location data". Your app reads one position when a customer taps one button on one screen, and a different position every 30 seconds while a rider is online. Those are two different promises with two different risks, and the template covers both with a sentence that cannot be wrong.
A sentence that cannot be wrong cannot make you change anything. That is the whole problem.
What we did instead
Over the last two weeks we rewrote the privacy policy that covers our delivery apps, ASO Delivery, ASO Rider and ASO Restaurant, one feature at a time. Not one sitting. Roughly two dozen separate passes, each covering a single thing the product does: the date of birth, rider ID photos, the profile photo, flagged rider positions, photos on a problem report, the notifications tab, the busy-areas map, nearby restaurants. Each pass was reviewed before the next one started, and several came back with corrections.
The output is a policy. The useful part was never the policy. It was that a feature has nowhere to hide once you have to write four sentences about it.
The four questions
For every feature that touches a person, answer these in writing, in plain language, with numbers:
- What exactly is collected, and what causes it. Not the category. The trigger. "Each time you tap 'I am at this address now', we read your phone's current position." If you cannot name the tap, you do not know what your app does.
- Who sees it, and when do they stop. Access has a start and an end. In our app a rider sees a customer's name, number and address only while carrying that order, from when the food is ready until the delivery finishes or is cancelled. Riders who do not take the order never see who the customer is. Write the end, not just the start.
- How long it is kept, and what deletes it. This is the question that creates work. "We keep it for 14 days" is a claim that something runs on day 15. If no job deletes it, the sentence is false the day you publish it.
- What happens when the account is deleted. Not in general. For this specific piece of data. Files in storage, rows in a table, and copies held by another company all have different answers.
What the questions found
Retention sentences need machinery, and the machinery is never instant. We keep a flagged rider position for 14 days. The deletion runs once a day, so in practice a record can live up to 24 hours past its 14 days. The honest sentence says so. The tidy sentence does not, and the tidy sentence is the one a template gives you.
Some data leaves your control, and you have to admit it. When a rider taps Navigate, the delivery point goes to whatever maps application is on their phone, and may stay in that app's history. We cannot delete it there. Our map tiles come from Stadia Maps, which sees the part of the map being viewed and keeps its request logs for roughly 7 to 14 days. Crash reports go through Firebase, and Google keeps them for 90 days before removal starts. None of that is a scandal. All of it belongs in the policy, because a promise you cannot keep is the one that damages you.
Error text carries personal data by accident. A crash report records what went wrong in the code. The error message can contain whatever the app was handling when it failed, which might be an order number, an address or a phone number. We do not attach an account to a report, and we still had to say that an error message can include that.
Deletion needs a route for someone who no longer has your app. Deleting an account from inside the app is the easy half. Someone who removed the app, or changed phone, still has an account. We built a page on the website where you sign in the same way as in the app and confirm. During review it turned out the page's hidden steps were being shown anyway, because a CSS rule overrode the hidden attribute. A privacy feature can fail silently and look finished.
A second language is a second copy to maintain, and ours is partial. Our customers read French. The policy itself is an English document. Five of the recent changes also carry a French paragraph: the busy-areas map, nearby restaurants, photos on a problem report, the notifications tab, and deleting an account without the app. The passages on the date of birth, rider ID photos and flagged rider positions are in English only, and the French deletion paragraph sends the reader to the English text for the detail. A French paragraph that lags its English one is wrong for the people who rely on it.
How to run this on your own product
- Take your screen list, not your database schema. Walk the app the way a user does, one screen at a time.
- Answer the four questions per feature, in the policy itself, before you think about tone.
- When an answer is "I do not know", that is a ticket, not a wording problem.
- Put a number in every retention sentence, then go and find the thing that enforces it. If it does not exist, build it or change the sentence.
- Have someone else read each pass, looking for claims the code does not support. Several of ours came back with exactly that.
- Treat the policy as part of shipping a feature. A feature that changes what you collect and does not change the policy has made your policy false.
What this does not solve
It does not make you compliant with any particular law. This is a product practice, not legal advice, and a lawyer should still read what you end up with.
It is also not free. Two dozen passes, each reviewed, plus five French paragraphs, is real time that did not go into features. And each French paragraph is a second copy of an English one, a maintenance cost we now carry forever.
What you get for that is a document where every sentence can be checked. For ours, the feature-by-feature reviews of the source and the policy are complete. Release checks remain open: the first scheduled runs of the newest deletion jobs, the new phone builds, checks on real devices, and the checks we run before switching a new feature on.
We ran this on our own product before recommending it to anyone else. That is the order we prefer.