One agent that runs conversion as a weekly loop. Every Monday it reads the funnel, sees what competitors changed, finds where yours leaks and picks one thing to test. Then it writes the variant, builds it and sets the test up to be trusted. Nothing changes until you've said yes. Two weeks later it reads the result.
Every Monday the agent pulls last week's funnel straight from the sources. Not one conversion rate, but every step: visits, who reached a room and a price, who started to book and who paid — split by phone and desktop, and by where the visitor came from.
The site
Booking page
Checkout
Bookings
Then the agent walks through the competitors' booking flows, the way a guest would — the room list, the price, the steps to pay. It keeps last week's copy of every page, so what changed shows up on its own. The OTAs are on the list too, because that's where the guest compares.
Booking flows · since last Monday
A funnel doesn't leak evenly. Most of the money goes at one step, on one device. So the agent lays the four steps side by side and finds the biggest drop — then splits it by phone and desktop, and checks whether it's the page, the form or the load time that's doing it.
Where the week leaked
A funnel moves for reasons that have nothing to do with the page. School holidays, a rate change, a new phone model, a competitor's offer. The agent keeps watch on all of it, so a drop isn't blamed on a button when it was the season — and so a real chance isn't missed.
Signal feed
This is where it all meets. The funnel, the competitors, the leak and the signals are laid side by side, and the agent picks one thing to test this week. One, because two changes at once can't be told apart afterwards. It also sets the smallest effect worth finding — if the test can't see that, it's not worth running.
This week's test
Before anything is built, the test is written as a sentence you can be wrong about. Because we saw this, we believe that change will do this, for these visitors, and we'll know by this number. A test without one of those five parts is a guess with a dashboard.
Test card · wk 38
The variant is the control with exactly one thing different. Everything else the week turned up is sorted into two other piles: plain fixes that don't need a test — a broken date picker, a field nobody uses — and good ideas that wait their turn. A test only answers one question at a time.
What changes this week
Now the right agents are brought in. One knows pages and forms and why people stop, one knows how to run a test that can be trusted, one knows popups and when they help rather than annoy, one knows the flow after the click, one knows why people decide, and one makes sure every step is measured.
Agents on the job
This is where the change becomes words. The price line and the sentence under it, the button, the trust line next to it, the popup for the visitor about to leave, and the form with the fields that are left. Short, concrete, and nothing the hotel can't stand behind.
Writing the variant
A test that stops the day it looks good is how most teams fool themselves. So the sample size and the end date are set before it starts, from the baseline and the smallest effect worth finding. Whole weeks, at least two, because Tuesday's visitors don't book like Saturday's.
Test · total-price-first
The week is done. The variant, the fixes and the queue sit ready exactly as they'll look on the site — nine drafts, each marked test, fix or queued. Nothing changes on the site until someone says yes.
Keep these dates for 24 hours?
Fri 9 – Sun 11 Oct · Double room · SEK 3,590 total. We'll hold the price, no card needed.
Hold my dates No thanksAfter your yes, the test goes live on Monday morning and the fixes go out beside it, outside the test. Then the hardest part: nothing is decided until day 14. The agent checks every day that both versions work, that the tracking fires, and that the guardrail holds — and reports without calling a winner.
21 September – 4 October
On day 14 the answer is read exactly as it was written on day one: the primary metric, at 95 %, on the full sample. Then mobile on its own, because that's where the leak was. A winner goes to everyone. A loser is a fact worth as much — most tests lose, and the ones that win rarely win by much.
Day 14 of 14 · booking page → booking
The week after, step 1 reads a funnel that has one leak fewer, and the next test in the queue moves up. Every result — win, loss or nothing — goes into a library with the hypothesis, the number and the reason. After a year the site has been tested fifty times, and it's built on the hotel's own guests, not someone else's best practice.