The Architect Lab
ProcessFeedback loops2026.09

We agree a number. Then it runs until it hits, or we tell you it won't.

Most agencies run campaigns. A campaign has a start, an end, and a report at the bottom that explains what happened. Ours have a target and no end date, which changes what everyone is arguing about.

What a loop is

§01 / Five parts
01The number

One thing we are moving. What a sale costs, how many people watch past three seconds, where you rank.

One only
02The target

The figure that ends it. Written down and agreed with you before we spend anything.

Agreed day 0
03What we try next

Ten more versions of whatever is currently winning. Never ten unrelated ideas.

Automated
04How we judge it

The same rule every round. Nobody votes on what they personally like, including us.

Automated
05When we stop

Target hit, or three rounds with no movement. If a loop has no stop rule it is not a loop, it is a habit.

Enforced

Three ways it ends

§02 / Only three
HitIt worked

The number is reached. The winner goes live, we bank it, and the next loop starts from there rather than from scratch.

KilledIt won't work

Three rounds, no movement. We stop, tell you plainly, and move the money to something with a better chance. This is a normal outcome, not a failure.

ReframedWrong question

The number will not move because the problem is somewhere else. Usually the offer, the price or the landing page. We say so and start a different loop.

Every loop we run has all three endings written into it before it starts. That is the only reason we can promise the second one out loud. Agencies that cannot say "this isn't working" will spend your money forever.

One running

§03 / A fortnight, sped up
Round 01 Target 30% watched past 3s
    Target hit. This one goes live, the rest are binned.
    Round1 of 3
    Best so far—
    Spend used€0

    Round one is four different openings on the same body. Whatever wins becomes the thing round two builds from. Round three tunes the opening on that winner.

    Twelve attempts, roughly €240 of test spend, one fortnight. If round three had come back flat we would have killed it and told you, and the €240 would still have been the cheapest thing you bought that month.

    Which loops run for you

    §04 / By package
    LoopNumber it movesFeedbackPackage
    Opening testWatched past 3 seconds48 hours02 · 03
    Version explosionCost per sale14 days02 · 03
    Copy against a fixed imageCost per sale48 hours02 · 03
    Fatigue watchClick-through decayDaily03
    Budget reallocationAccount cost per saleDaily03
    Landing pageConversion rate14 days03
    Offer wordingCost per sale and order value14 days03
    HoldoutWhat the ads were actually worthMonthly03
    Rank gapGoogle position14 days04
    Internal linksPosition, pages 4 to 1521 days04
    AI answer visibilityNamed across 20 buyer questionsWeekly04
    Source placementCited where the models look30 days04

    The audit tells you which of these are worth running before you buy any of them

    Where the AI actually sits

    §05 / And where it doesn't
    It does the counting

    Pulling accounts apart, scoring every attempt the same way, spotting decay before a person would. This is why the audit takes a day instead of a fortnight.

    It does the volume

    Versions, crops, captions, first-frame swaps. Rendered rather than re-edited, so testing ten things costs about what testing one used to.

    It does not decide

    A person reads what the numbers mean and writes the next brief. Models are good at ranking and bad at knowing when the whole question is wrong.

    It does not make the good thing

    The one video a month is made by a human who is good at it. Everything else is built from that.

    §06 / Start with the audit

    Pick a number.
    We'll tell you if it's reachable.

    Book twenty minutes Email us the account See the packages