I make brands impossible to scroll past.

The obvious shoot is overhead flatlays on marble. Every restaurant within three blocks already has that, and it makes a 40-year-old family kitchen look like a franchise that opened last spring.

So we killed the studio idea. The angle we kept was heat — steam, char, hands, the pass. Nothing styled to death, everything shot at the temperature it is actually served at. I brought one light and used the kitchen's own tungsten for the rest, because the warm cast IS the brand.

The unlock was shooting during a real service instead of a closed set. Harder, uglier, and the reason the pictures look like a place instead of a catalogue.

The constraint was time: whatever I designed had to be repeatable by a salesperson with a phone after I left.

So the idea could not be a look that depends on me. It had to be a rule. We tested three: golden-hour only (unreliable in this climate), a portable backdrop (too slow), and one fixed position against the workshop's black wall with a single light on a stand.

The black wall won because it does two things at once — it kills the sky problem and it makes every car in the inventory look like it belongs to the same brand.

Everyone shoots the gear because the gear is what was paid for. But nobody joins a gym because of a leg press.

We listed what members actually described in their reviews: the 6am crowd, knowing the front desk person's name, the walk home after. None of that is equipment. So the whole shoot became faces and the minutes either side of the workout, and the equipment is only ever out of focus behind someone.

I argued for keeping sweat and blown-out hair in the final selects. That was the fight. It is also the reason the pictures do not look like stock.

The differentiator was not the furniture, it was how the light moves through their rooms — they design around windows. A single hero photo cannot show that.

So instead of one perfect frame per room, we shot the same locked-off frame at 8am, 2pm and 8pm. The carousel format does the work: swipe, swipe, and you have understood the entire pitch of the studio without reading a caption.

It is a slower, more expensive way to shoot one room. It is also the only version anyone remembers.

Retention graphs are brutal reading: the drop was not gradual, it was a cliff at 0:03. That is not a content problem, that is an opening problem.

So before spending a cent on production, we cut nine different first-seconds from footage we already had — a face mid-sentence, a hard sound effect, a text-only frame, an unfinished action, a number. Everything after second three stayed identical across all nine, which is the only way to know what you are actually measuring.

The winner was the unfinished action: someone starting a movement, cut before it completes. People stay to see it land.

The mistake was treating captions as an accessibility add-on rather than as the actual script.

We rebuilt the edit muted, on a phone, at arm's length — which is the real viewing condition — and anything that stopped making sense got recut. Captions became the primary layer: two to four words at a time, one idea per card, timed to the cut rather than to the speech.

Sound design then went back on top as a bonus, not a dependency. Every edit now works twice.

The temptation with long footage is to cut a highlight reel. Highlight reels perform badly because they have no single idea.

Instead I transcribed the whole thing and read it as text, marking only the sentences that would survive with no context at all. There were fourteen. Each of those became its own edit with its own hook, and everything else was thrown away.

The test I use: if the first sentence needs the second sentence to make sense, it is not a short.

The whole genre has trained viewers to be suspicious. So the credibility came from matching, not from flattering: same lens, same position, same time of day, same grade.

Counter-intuitively we made the before shots look better than the client wanted. A fair before makes the after believable, and believable is what converts.

The transition is a straight cut on a beat, not a wipe. Wipes read as a sales technique.

The discount angle is the one every restaurant runs, which is why it costs the most to win with.

We wrote three angles instead: scarcity (the room only has eleven tables), ritual (what the regulars order without looking), and price. Four hooks under each, one format, so twelve statics that differ in exactly one variable at a time.

Ritual won and price came last. The client's instinct was the losing angle, which is the entire argument for testing rather than debating.

Most teams solve this by designing for 9:16 and then cropping, which puts the logo in a safe zone on one platform and under a UI element on another.

We designed the safe area first — a single grid where the type never leaves the middle 60% and the brand mark sits on a corner anchor that exists in all four ratios. Then the artwork was built outward from it.

Boring, structural, and it removed a whole category of last-minute rework.

You cannot win on the property when the property is not yours to choose. So the creative stopped selling houses and started selling the agent's track record.

The headline became a result — days on market — with the house demoted to background texture. It is closer to a recruitment ad than a property ad, which is exactly why it stood out in a feed of identical kitchens.

We tested the honest number against a rounder, better one. The specific number won, as it usually does.

Retargeting creative usually pretends to be a first impression. It never is — the viewer knows they were on the site, and the pretence is what makes the ad feel creepy rather than useful.

So the copy owns it plainly, and the creative shows the actual product they looked at against a black ground with nothing else in frame. No lifestyle, no model, no discount in the first touch.

The discount only appears at day five, which meant we stopped paying a margin to people who were going to buy anyway.

Forty-six ad sets on a modest budget means no ad set ever leaves the learning phase. The account was not badly optimised, it was badly shaped.

The hard part was arguing for less: fewer audiences, broader targeting, more budget per set. It feels like giving up control, and for the first eight days the numbers look worse, which is why most people revert on day three.

We agreed the read date in advance and wrote it down. That single agreement is why the restructure survived long enough to work.

Most launches fail in the planning meeting, not the execution — because nobody agrees what the number is, so every channel gets judged by a different one.

We spent the first session picking a single north-star metric (first-time purchasers) and explicitly demoting everything else to diagnostics. Reach, engagement and CTR stayed on the dashboard but under a line labelled 'not decisions'.

That line is the most useful thing I have ever put in a report.

Splitting a year of budget into twelve equal pieces is an accounting decision that has quietly become a marketing one at almost every small business I have worked with.

We pulled three years of their own sales data — not industry benchmarks — and found the demand curve was much spikier than the spend curve. So we underspent deliberately in the flat months and banked it.

The uncomfortable part is that the flat months then look like failure on a monthly report. We changed the report to a rolling twelve weeks so nobody had to defend a bad-looking March.

The awkward thing about managing a channel is that recommending its removal reduces your own invoice.

I pulled twelve months and looked only at last-non-direct touch plus a holdout week where we simply switched it off. Nothing changed when it was off, which is about as clean a signal as this work ever gives you.

Saying so out loud is the whole job. A retainer that survives on a channel nobody can defend is not a client relationship, it is a countdown.

A report that needs explaining is a report that will be ignored the month you are on holiday.

We asked a different question: what decision does this person make each week? There were three — how much to spend, what to shoot next, and whether to keep a channel. Everything that did not feed one of those three was deleted.

Six numbers survived. Each one has a plain-English sentence underneath it saying what to do if it moves. That sentence is the actual product.

Everyone wanted a creative strategy. It would have been irresponsible to give them one, because every decision would have been made against numbers we already knew were wrong.

The audit found the boring, common causes: a duplicate tag firing on the thank-you page, no UTM convention, and a checkout redirect losing the session. None of it clever. All of it decisive.

Spending three weeks on plumbing when the client is paying for growth is an uncomfortable conversation. It is also the only honest sequence.

Institutional memory is the cheapest performance gain available and almost nobody keeps it.

The format matters more than the tooling: one row per test, written hypothesis before launch, the metric agreed in advance, and a conclusion in one sentence written the day it ended. If the conclusion cannot be written in a sentence, the test was badly designed.

The hypothesis-before-launch column is the one that does the work. It stops post-hoc storytelling, which is the main way marketing teams fool themselves.

Retention graphs are normally used as a post-mortem. Read across thirty videos instead of one, they become a shot list.

We overlaid the curves and found the drops landed on the same three things every time: establishing wides, logo stings, and any shot without a person in it. The rises landed on faces and on hands doing something.

So the next shoot day was planned from the graph: no establishing shots at all, twice the hands, and the logo moved to the end where it costs nothing.