MarketBeat — OpenAI IPO creative

Two pen ads have been live since 31 July. Same format, same voice, same window, same account. The only visible difference is which offer they point at — and the gap is large:

The IPO angle pulled 68% more clicks and converted 33% worse. It's a click magnet that under-converts — which lines up with the two ICPs on file: high-yield speaks to the conservative, dividend-focused buyer; IPO speaks to the younger, aggressive one. The first is who actually downloads. Caveat: two ads, and the copy differs as well as the offer, so this is a strong signal rather than a clean A/B — but it's 8,063 conversions of signal, and every round-1 concept sits on the IPO side of it.

Every round-1 concept is IPO-angle. This holds the hook skeleton of round 1 V1 exactly and swaps only the angle and the offer, so if the offer really is the lever, this one ad says more than six more IPO variants would.

All three round-1 pen ads are “The 3 most dangerous [words/assumptions/bets]”, and all three dinner scripts are a brother-in-law regret story. That's two tests wearing six hats. With new-creative tests moving to $500/day, near-identical creatives split the budget without splitting the question. Two genuinely different openings:

The pen barrel sits over the line it's reading. In round 1 it hides up to three words at a time — on a format whose entire job is getting copy read, that's working against itself. The pen is now 10% narrower. Left is round 1, right is the new build, same moment in the read:

The end card: the 1440 reference closes on a signup screen with a “Sign-Up Now!” pill — it asks for the click. Round 1 just stopped on a scene and asked for nothing.
The length: round 1 ran 46–59s against a 35s reference, because the scripts are full narratives. The trims drop the middle beat of each — the flourishes, not the argument — landing at 38–46s. Both changes are in the new cuts; round 1 is untouched beside them.

The third pen ad, and both DCT_207 image directions. Nothing here changed.

Option A matches the reference: the image slot is a plain editorial photo and the headline lives in the ad unit, so nothing is baked in. Option B sets the headline on top. Native's whole trick is not looking designed, so A is the on-format answer and B is the hedge.

Checked Notion for MarketBeat concepts still sitting in brief stage. Three had real content — and one of them, DCT_206_Dangerous, is the copy above: V1/V2/V3 word-for-word. But its Style field says “Text Over Video”, and we'd built pen ads. So the briefed version had never been made.

Before building a text-over-video cut, we pulled the Foreplay reference attached to that brief. It's a pen ad. The Style field is simply wrong — the reference is the format we'd already built. Building to the label would have meant building the wrong thing to satisfy a bad database field. So we compared ours to the reference instead, and it showed two things we were off on: it runs 13 seconds (ours were 20–26s, because the copy is roughly twice as long), and a real human hand holds the pen — which is most of why it reads as someone filming a page rather than as an ad.

First hand attempt was better than a floating pen but still read as a sticker. Five tells against the reference: too small, no contact shadow so it floated above the paper, lighting mismatch (studio-white cutout on a warm daylight page), no forearm (cropped clean at the wrist), and a near-vertical grip instead of a ~35° writing angle.

Two passes to fix it. v2 regenerated the asset larger, with forearm running off the bottom-left and a proper writing angle, then warm-graded it to the page and added a contact shadow — but the background removal left a white halo around the arm, which against a beige page reads as a cutout outline and arguably looked worse. v3 erodes the alpha, feathers the new edge and de-spills the remaining white fringe. That's the one to look at.

Zoomed in, v1–v3 all had the same fault: the pen is upside down. The reference isn't a writing grip at all — it's a pointing grip, hand below the barrel, fingers wrapped round it like holding a stick, tip up. Ours used a writing grip (thumb-and-index pinch, barrel through the web of the thumb) but inverted so the nib pointed up — i.e. a person writing with the pen the wrong way round. The earlier note about a “~35° writing angle” was a misread of the reference, and the prompt baked that error straight into the asset.

The reason a composite is hard is that the reference is one photograph — hand, pen, paper and shadow all share a single light. So we also tried generating the entire frame in one shot. It's markedly more photoreal than anything composited: real window light, a real shadow cast onto the page, wood desk, shallow depth of field.

But it can't ship as a pen-read ad. The text on that page is model-generated, so we can't set the copy, the price or the offer — and controlling copy exactly is the entire reason this pipeline exists. It's here as the look benchmark for what we're aiming the composite at, and it's a strong candidate as a native static in the DCT_207 mould, where no verbatim script is needed.

Luiz, 24 Aug: “the pen ads look good. However, there are a lot of changes I'd make to the voiceover ads.” Three notes, all on DCT_205:

1. The image must match what's being said. His example was the line “brags about the money he made on Peloton” — the B-roll should be a cocky guy showing his phone at the dinner table; ours looked like everyone wanted to see what was on the phone. He's right, and it was structural: round 2 had six stills for a 46-second script, so each one had to cover ~8 seconds of narration and could only be mood-adjacent. 2. Cadence. The 1440 reference cuts roughly every two seconds, dissolves between shots, and every shot zooms in or out. 3. Captions sometimes lag the voice. He captions with his own tool, so the ask was: fix the B-roll, don't spend time on captions.

What changed in the build. Each script was broken into beats (V1: 26, V2: 21, V3: 19) and a still was generated for that line — “his face turned red like a tomato” is a flushed close-up, “you're the one holding the bag” is him peering into an empty paper bag, “Suppliers / Backers / Infrastructure” is a wafer, a handshake, a server aisle. Shots are timed to the word the beat starts on, dissolve over 0.4s, and alternate zoom-in / zoom-out. Trimmed cuts reuse the same stills, so the full and trimmed versions of each script now match shot-for-shot.

Captions: per Luiz's note, the trafficking set is caption-free — he adds his own. A captioned twin of each cut is here too, with the timing lead fixed, in case it's quicker to use as-is. (Fixing it also surfaced a wrap bug in round 2: long caption lines were being clipped at the frame edge.) Recurring characters — the brother-in-law (slicked hair, blue shirt) and the narrator (grey, dark crewneck) — are prompted consistently but are AI stills, so faces drift a little between shots.

Karaoke captions with the lead fixed and the pill kept above Meta's bottom-third crop. Optional.

Luiz, 24 Aug, on round 5: “It's wayyy better now.” Two more notes. The VO drifts as the read goes on — he generated his own ElevenLabs take at 1.13× speed and sent it (“give it to Claude and the result will be better”). And when the script gets to the seven stocks, show the landing page.

So this cut uses his audio file untouched (51s vs our 57s — the faster pace also tightens the average shot to 2.1s), re-timed shot-for-shot to his read, and the end card is now a capture of the live offer=7aistocks page inside the phone frame, starting on “MarketBeat put all seven in a report” and holding ~10s to the end. His read closes “free for the next 24 hours” rather than “it's free today”, so the last beat was re-cued to match.

One thing to decide: the landing page prints today's date in its body copy (“free today, Tuesday, August 25th”). A baked-in screenshot is stale the day after it's captured. Options: recapture on the trafficking day, crop the date line out, or keep the HTML mock end card. This cut uses the 25 Aug capture.