Redemptive AI

AI that gets out of the way.

Sojourn is a research project with users. The question we are testing is whether AI can put people in front of scripture and then fall silent. This page is the scoreboard, updated as we ship, and it shows the numbers that embarrass us alongside the ones that don’t.

The premise

Most AI products are built to hold your attention. We think that is exactly backwards for this one. Scripture is the thing worth your time, and a good guide walks you to the door and stops talking. So we measure Sojourn by one thing: how much of your time with it you spend reading the Bible.

That gives us an unusual pair of goals. We want time in scripture to go up. We want conversation with the AI to go down, or at least to get shorter and more direct on the way to a passage. Those pull against each other, and publishing both is the only way to keep ourselves honest about which one is winning.

We hold that loosely. A reader who needs to talk for a while before they are ready to open Job is well served by a product that lets them. What we are watching is the shape of the whole thing: whether the conversation is a hallway or a room.

How we are doing

Six measures, taken from how the app is actually used. Each one carries the direction we are trying to move it and how it has moved since we started counting. Underneath them sits the one claim on this page that is finished.

Time spent in scripture

7.7%

↑we want this up

Of the time a reader is actually active in the app, the share spent with a verse or a chapter open rather than in conversation. This is the number the whole project is for.

up from 0.6% to 10% across the weeks, going the right way.

Visits that reach scripture

30%

↑we want this up

The share of app visits where the reader opened a verse or a chapter at all. A visit that ends without scripture is a visit where we talked and nothing else happened.

up from 25% to 29% across the weeks, going the right way.

Talked, never opened a verse

62%

↓we want this down

Of the visits where a reader wrote to Sojourn, the share that never reached scripture. This is our failure number, and it is the one we most want falling.

down from 75% to 50% across the weeks, going the right way.

Turns before a verse opens

2turns

↓we want this down

How many messages a reader typically exchanges before the first verse opens. The lower this goes, the faster the AI is handing off and going quiet.

Steady from week to week.

Peeks that become real reading

40%

↑we want this up

Of the verses a reader peeks at, the share where they keep going into the full chapter instead of closing the sheet. We want the glance to turn into real reading.

down from 67% to 25% across the weeks, going the wrong way.

Typical length of a visit

3.0min

·context only

The median time a reader is active in the app in one sitting. This one has no target on purpose. A long visit spent entirely in conversation does nobody much good, and a short visit that ended in a chapter did its job.

down from 3.0 min to 1.7 min across the weeks.

100%

Done

Scriptural accuracy

Every verse anyone reads in Sojourn is the World English Bible text, character for character.

This is not a number we watch and hope stays high. The AI is not permitted to write scripture at all. It can only emit a marker naming a reference, and the app resolves that marker against the Bible database it ships with and renders the real verse. If what the model pointed at does not line up with the actual wording, the citation falls back to a bare reference rather than show anything uncertain, and a reference the database does not recognise is dropped. There is no path through the app by which a sentence the AI wrote could reach a reader as scripture.

Week by week

The headline number, tracked weekly. This is the share of active time in the app that is spent with scripture open rather than in conversation.

Share of active time spent in scripture

0.6%

12%

7.5%

5.0%

12%

10%

Aug 17Aug 24Aug 31Sep 7Sep 14Sep 21

Week by week the share of active time spent in scripture has climbed from 0.6% to 10%. Each bar is one week, scaled against the best week so far rather than against a full hundred, because at these levels a bar drawn to scale would be invisible. We would like this chart to eventually make itself embarrassing to draw this way.

A few visits in particular

Averages hide the interesting part. These are single visits, picked because their shape says something an aggregate cannot. We hold no record of what anyone wrote, so what follows is the sequence of what happened and nothing else: every step below is an action the app recorded, with the scripture ones marked.

One message, three chapters

31% in scripture 12.6 min active 1 message September 20

went back to a saved thread›read Psalm 31›came back›read Romans 12›started a new thread›asked›read Psalm 51

This is the shape we are building for. The reader opened a conversation they had saved, went straight into Psalm 31, left and came back, read Romans 12, and only then typed anything at all. The single message they sent led to a third chapter. Almost a third of the visit was spent in the text, with the AI barely involved.

The hallway working as intended

49% in scripture 10.3 min active 4 messages August 25

Three messages, then James 1. One more message, then 2 Corinthians 9. Nearly half this visit was spent reading. It is the pattern the product is supposed to produce: talk long enough to find the passage, then stop talking. Worth saying plainly that the talking came first and mattered.

A good visit our own scoreboard scores at 1%

1% in scripture 9.2 min active no messages September 10

went back to a saved thread›read John 14›came back›read Hebrews 12

No messages whatsoever. The reader came back to a saved thread, read John 14, left, came back, read Hebrews 12. By any honest reading this is the best kind of visit there is, and the number at the top of this page counts almost none of it, because the reader closed both chapters by leaving the app, which we have no way to see. This one visit is the clearest argument for fixing that.

Thirty two minutes, seventeen messages, no scripture

0% in scripture 31.9 min active 17 messages September 19

asked 4 times›came back›went back to a saved thread›came back›asked 3 times›came back›asked twice›came back›asked 4 times›came back›asked 4 times

The failure case, and it happens plenty. Half an hour of conversation, a saved thread reopened partway through, and not one verse opened the whole time. Something kept this person here, and it was us doing the talking. We do not know what they were carrying and the log will never tell us, but on this page that counts as a visit we got wrong.

What we have tried

Every idea gets built and put in front of people to see what happens. Here is the log: what we thought, what we shipped, and what the numbers said afterward.

Sep 16, 2026Shipped

Search should answer with questions

Keyword search over the Bible is the worst version of this. Someone typing "envy" is carrying something around. If the search returns the questions underneath the word, already answered and already attached to a passage, the reader can skip the conversation entirely and land in the text.

What shipped: A search tab that takes a word and returns the questions beneath it, each arriving with the passage that speaks to it and a way to walk the full chapter.

Visits that reach scripture: 29% before, 32% after. Moved the way we hoped.

Sep 11, 2026Shipped

Give people a first line so they stop composing one

A blank prompt makes people explain themselves. They write a careful paragraph, we write a careful paragraph back, and scripture is still two exchanges away. Offering a concrete opening line should collapse the throat-clearing and get to a passage sooner.

What shipped: Two tappable suggestions under the empty chat, one naming a feeling and one asking a question about scripture, reshuffled each time.

Turns before a verse opens: 2 turns before, 1 turn after. Moved the way we hoped.

Sep 10, 2026Shipped

Let people carry a passage out of the app

A conversation that ends when the app closes has not changed anyone's week. If a reader can send the whole thread to their own inbox, the passage outlives the session and turns up again somewhere we cannot see.

What shipped: An "Email this to me" option at the foot of every conversation, sending the thread and its scripture.

Another one we cannot score in-app on purpose. What happens in someone's inbox is outside our instrumentation, and we would rather ship it than only build what we can measure.

Sep 4, 2026Shipped

A verse on the home screen, before the app is even opened

The best version of this product is one you do not open. A widget puts a passage in front of someone during the same glance they were going to spend on notifications, with no conversation involved at all.

What shipped: A home screen widget carrying a verse chosen from the topics a reader has been sitting with.

Deliberately unmeasured by the numbers above, which only see time inside the app. A reader served entirely by the widget looks like a reader who left, and we think that is the right kind of loss.

Aug 19, 2026Shipped

A citation should be a door, not a footnote

If the AI names a verse, most people will not go look it up. But if the verse itself is tappable and opens into the passage around it, the reading happens in the same breath as the answer. The reference is the handoff.

What shipped: Every scripture reference in a reply became tappable, opening a peek of the verse with a way through to the whole chapter.

This is the interaction everything else on this page measures, so it has no before to compare against. It is the baseline everything else gets measured from.

NextNot built yet

Record when a chapter is closed, too

The deepest reading in Sojourn is invisible to us. A reader who opens a chapter and stays there until they put the phone down produces one event and then silence, so our own scoreboard scores that visit as though nothing happened. Until the app reports the close, every scripture figure on this page is too low, and the readers we are serving best are the ones we can see least.

Not built yet. When it ships, the numbers above will rise for reasons that have nothing to do with anyone reading more, and we will say so here plainly.

How these numbers are made

A visit is everything one reader does with less than 30 minutes of silence between actions. Time is credited action by action rather than end to end, and a single action can earn at most 5 minutes, so a phone left face-up on a table does not read as devotion.

Time in scripture is measured from the moment a verse is tapped until the reader closes the peek or moves on. Peeks are exact: every tap is closed by an action we see.

Chapter reading is undercounted, badly. The app records that a chapter was opened but nothing that records it being closed, so when a reader opens a chapter and then leaves the app we cannot tell ten seconds from ten minutes. Those reads currently score zero. That is very likely the deepest reading anyone does in Sojourn, and it is missing from the number above. Instrumenting the close is the next thing we are fixing, and when we do, the scripture share on this page will jump for reasons that have nothing to do with anyone reading more.

We publish shares only. Every figure here is a percentage, a median, or a duration. Sojourn is young and a raw number of readers would say more about our launch than about whether the idea works, so we leave it out and let the ratios carry the argument.

The last week is usually partial. The rightmost bar covers however much of that week had happened when these numbers were taken, so it moves around and often reads low. Wait for it to fill in before reading anything into it.

The sample is small. These figures come from over a hundred visits, and a single unusual week moves them visibly. Watch which way they are moving more than where they land.

Our own accounts are excluded. The people building Sojourn use Sojourn, and none of that reaches these numbers.

Measurement starts where the app starts reporting. Nothing before August 18, 2026 is counted, because earlier releases recorded conversations but not what readers did with scripture. Counting those would read as a failure that was really just a blind spot.

Measured from August 18 to September 24, 2026. Last updated September 27, 2026. The underlying figures are published as metrics.json.