I've been running an automated YouTube Shorts channel alongside the three directory sites in this experiment. The Shorts are AI-generated, low-cost, and designed to test whether video distribution adds anything to the organic search play. After 19 published game-comparison videos, the analytics have started to say something specific enough to act on.
Here's what changed in my titles, and why.
The default template wasn't working
The original Shorts followed a template like: "This indie game costs $9.99. It has more Steam reviews than that AAA title." The structure is fine. The problem was my first few videos named no one: "This indie game costs less than a major FPS — and has more reviews."
Median views for those: 5. Not 500. Not 50. Five.
The underlying idea (small indie, big AAA, review-count comparison) is defensible. The framing was gutless. I wasn't giving the viewer anything to recognize. "Major FPS" could be anything. "This indie game" tells them nothing. They have no reason to click because they have no context for why the result would be surprising.
I changed the template. Named both games. Within a week I had a video at 162 views and one at 373. The structure was identical — same archetype, same hook rhythm, same thumbnail layout. The only change was specificity: Terraria vs Cyberpunk 2077 instead of "cheap indie vs big budget RPG."
The middle case was the most instructive
I tested a hybrid: name one game, genericize the other. "Terraria Costs $9.99. It Has More Steam Reviews Than a $70 AAA Game." Terraria is named; the opponent isn't.
That video landed at 77 views — better than unnamed, worse than both-named by roughly half.
The lesson wasn't "name more things." It was that the viewer's surprise depends on two reference points, not one. Terraria being cheap is only interesting relative to something they already expect to be expensive. "A $70 AAA game" could be anything, so the comparison doesn't land. Cyberpunk 2077 is a specific cultural object they have a price expectation for. The gap between those expectations is what produces the click.
What this forced into the code
I have a machine-generated daily directive that tells the Auto-Script routine what to generate. After this data, I added the following as a hard gate:
R6 HARD GATE: the title MUST name TWO real, recognizable proper nouns — a famous indie game AND a currently-famous AAA game. Named-vs-named drove 162-373 views; stripping the names collapsed the same template to 77; naming nobody died at 5.
This isn't a style suggestion in a prompt. It's an invariant the directive enforces before the video script is accepted. The archetype selection already lives in code, so adding a verification step for the two-named-noun requirement was a small extension.
The build-in-public videos I was also running had a median of 13 views and have since been marked dead in the directive. They failed for a different reason — no recognizable reference points at all, not even a game title to anchor to — but the failure mode is adjacent.
The side effect: a pre-verified matchup queue
Naming two specific games introduced a new problem: I need to verify the review-count premise before publishing. "Lethal Company has more Steam reviews than the latest Call of Duty" is only true if you check the actual Steam store page, today. Review counts drift. Franchise titles have SKU splits (which CoD entry counts?).
I built a matchup queue in docs/yt-matchup-queue.md — a prioritized list of indie-vs-AAA pairs with plausibility notes. Each entry is flagged as SHIPPED, INVALID, or pending. One entry (Schedule I vs Hogwarts Legacy) was initially plausible and later marked INVALID after verification: Schedule I had ~307K reviews against Hogwarts Legacy's ~368K. The premise failed; the video was skipped.
Today's script pulled the next valid pair from the queue: Lethal Company vs the latest Call of Duty on Steam. Before it publishes, the script cites the live review count with a Steam store URL. The quote in the video will say "as of July 2026," not a number I committed two weeks ago.
The naming requirement forced me to build verification infrastructure I should have had from the start. Concrete claims are useful precisely because they're checkable.
Part of an ongoing 6-month experiment running three AI-curated directory sites. The technical claims here are real; this article was AI-assisted.
Top comments (0)