Lab: Observed behaviour over 82 cycles in London

American Express is supposed to be hard to use abroad.

In London, that doesn’t line up with what happens.

Since the Marriott Bonvoy Brilliant American Express Card moved to a $25 monthly dining credit in September 2022, we’ve run two cards continuously against normal spend. From November 2022 through March 2026 that produced 82 monthly cycles, $2,050 available, and $2,039 captured.

Nothing was set up to make that happen. No tracking, no routing, no attempt to force the credit.

It’s the full history. Every charge, every credit.

The credits don’t post as $25. They arrive in pieces, often $3.58, $7.62, $10.22, amounts that line up exactly with individual transactions. Once those are matched back, the pattern starts to show.

The underlying spend is ordinary.

Pret, Itsu, Gail’s, Joe & The Juice, Nando’s. Pubs you end up in without planning it: The Bolingbroke, the Ship in Mortlake, White Horse on Uxbridge Road, Churchill Arms, the Waterfront by the river.

There are a couple of repeat outliers that don’t look like restaurants at all, a school-adjacent café and a hockey club bar/café.

Across the dataset, roughly 15–20 distinct venues appear with any regularity. They fall into a small number of functional categories:

The split isn’t even. Most months complete through a small number of cafés and pubs, with chains filling in the rest. Other venues appear occasionally but don’t persist.

It isn’t the type of place.

Two pubs that look identical can behave differently. Wetherspoons is the cleanest example. Sometimes it triggers, often it doesn’t. Meanwhile other pubs, no more “restaurant-like” than that, clear almost every time.

A small number of venues never developed a pattern. Not many, but enough to notice, and repetition didn’t change them.

Credits don’t post immediately. Most appear a few days later, with a clear clustering around day four.

At the start that delay creates uncertainty. You don’t yet know what has worked. After a few repetitions it stops mattering.

What changes is concentration. The same names begin to recur, attached to the same small charges, and then they dominate. By mid-2023 most months are completing through a short list: sub-$25 transactions at Pret, Gail’s, the same café, the same pubs.

Other places come and go, depending on the underlying spend.

The spend itself doesn’t get shaped around the credit.

The only adjustment happens at the margin, at the point of payment. If a place had cleared reliably before, the Brilliant card comes out. If it hadn’t, something else does.

Across the full run, the shortfall was $10.94.

Not engineered, not optimised. The same places kept triggering.

Each month starts with a $25 target. In practice it’s completed by a small, stable subset of transactions.