I started posting AI-generated puzzles on February 1st. Two series: AI-SPY, which is a flat lay scene filled with objects to find, and Hidden Objects, where items are structurally blended into decorated surfaces and artistic compositions. Both made with Nano Banana Pro in Adobe Firefly. Both posted daily alongside my usual content.

I tracked every post. Impressions, likes, replies, engagement rate, detail expands, profile visits. All of it. And the story the data tells isn't what I expected.

The puzzles don't reach as many people. But the people they reach actually care.

That's the engagement rate across all 24 puzzle posts. Likes, replies, reposts, bookmarks, and detail expands divided by impressions.

My non-puzzle organic content over the same period? 8.3%.

That means puzzles drove 29% more engagement per impression than everything else I posted. Not a slight edge. A measurable, consistent gap across two weeks of data.

The like rate tells a similar story: 4.24% on puzzles versus 2.90% on non-puzzle posts. Nearly half again as high. But the metric that surprised me most was replies. Puzzle posts generated replies at 1.6 times the rate of my other content. People weren't just tapping a heart and moving on. They were stopping to comment which objects they found, tagging friends to try the puzzle, coming back to check if they missed something.

That behavior matters more than the numbers suggest. Replies carry the most weight in X's ranking algorithm. A like is a signal. A reply is a conversation. And conversations get rewarded with distribution.

Here's the thing nobody tells you about engagement rate: it can go up because your reach goes down.

Average impressions on my puzzle posts: 403. Average impressions on non-puzzle organic posts over the same window: 575. The puzzles reached fewer people.

But every quality signal improved. More likes per impression. More replies per impression. More detail expands, which means people tapping to look closer at the image. The algorithm wasn't pushing puzzles wide. It was pushing them to the people most likely to actually engage.

This is what tightening audience fit looks like in real data. Engagement growing faster than impressions doesn't mean you're going viral. It means you're finding the right people. And for a creator building a community rather than chasing reach, that's the more valuable signal.

I'd rather have 400 people who stop, look, and reply than 600 who scroll past.

Hidden Objects Level .010. A single image with five objects blended into a decorated composition. Nothing flashy. No thread. No hook beyond "Can you find all 5 hidden objects?"

427 impressions. 30 likes. 15 replies. 3 reposts. 12 detail expands.

22.7% engagement rate.

For context, that's nearly three times my non-puzzle engagement average. From one image and one line of text. The replies were people calling out specific objects they found, debating whether the crown counted because it was tucked into shadows, and tagging people to try it themselves.

I've posted images that hit thousands of impressions. Threads that took hours to research and write. The single post that generated the deepest engagement this month was a puzzle that took me ten minutes to create.

That's not a fluke. The top five puzzle posts by engagement rate are all above 12%. Three of them are above 15%. The format consistently outperforms because it does something most social media content doesn't: it gives people a reason to stay on the image.

The most interesting thing about puzzle engagement isn't in my analytics dashboard.

It's in how people behave. Someone lands on the image. They read the caption. They start scanning for objects. They find three. They zoom in. They find four. They scroll down to the replies to see if anyone mentioned the fifth. They come back later to check.

That's dwell time. Time spent on a single piece of content. And while X doesn't surface that metric to creators, it's one of the strongest signals the algorithm tracks internally. Detail expands, which do show up in analytics, are a proxy for this. Puzzles averaged 4.8 detail expands per post versus 15.0 for non-puzzle content in raw numbers, but non-puzzle posts also had 43% more impressions to draw from. The per-impression rate tells a different story: people who see a puzzle are significantly more likely to tap into it.

There's also a behavioral pattern I've watched play out in the replies. Someone finds the objects, comments their count, and tags a friend. The friend replies. Now you have a reply chain on a single image post, which is algorithmically identical to a conversation. The puzzle creates its own engagement loop.

One reply I keep coming back to: someone called it "a nice break from the scroll." That's the whole thesis. In a feed full of hot takes and promotional threads, a puzzle asks for a different kind of attention. Not agreement. Not outrage. Just quiet focus for thirty seconds.

I ran two series because I didn't know which approach would land.

AI-SPY is a traditional I Spy format. Photographic flat lay scenes cluttered with objects, and you scan through the chaos to find specific items. Hidden Objects takes a different approach. Objects are structurally integrated into decorated surfaces, baroque scrollwork, ornamental carvings, painted patterns. The item IS part of the image until you recognize the shape.

The numbers give Hidden Objects a slight edge. 11.6% engagement rate versus 10.1% for AI-SPY. More replies per post on average. And the recent Hidden Objects posts are trending upward while AI-SPY has been more consistent but flat.

The reply behavior is different too. AI-SPY replies tend to be confirmations. "Found them all." Hidden Objects replies are more specific. "The crown is in the shadows." "Where's the fish?" People engage differently when objects are genuinely camouflaged versus placed in visual clutter.

Both series stay in the rotation. But Hidden Objects is where the difficulty ceiling is higher and the engagement trend is pointing up.

I'm not pivoting to puzzle content. But the data is clear enough that puzzles earn a permanent slot in my content rotation.

The difficulty problem is still the biggest challenge. AI wants to present objects, not hide them. Every puzzle is a prompt engineering negotiation where I'm trying to convince the model to do the opposite of what it's trained to do. Some days I get a 22% engagement rate. Some days the objects are sitting in plain sight and the puzzle barely functions.

That inconsistency is actually the interesting part. Each puzzle is a live experiment in how AI models interpret the concept of "hidden." And documenting that process, the successes and the failures, has turned out to be content in its own right. The thread I posted about I SPY prompt methodology hit 568 impressions and 56 engagements. The article I published on the research behind Hidden Objects prompted one of my best days on X this year.

The process becomes the content. The content generates the engagement. The engagement funds more experimentation.

I'm pushing into new art styles for the Hidden Objects series. Testing whether the algorithm rewards puzzle consistency over time. And still trying to answer the question I started with: can AI image generation produce a genuinely challenging puzzle?

The answer so far is: sometimes. And "sometimes" is enough to keep going.

You don't need a viral post. You need a reason for someone to stop.

A puzzle is one of the simplest reasons possible. It doesn't require your audience to agree with a take, feel an emotion, or click a link. It just asks them to look closely at something for a few seconds longer than they normally would. And those few seconds compound into engagement signals that the algorithm rewards.

If you're an AI creator looking for content that performs, consider what in your workflow already asks people to pay attention. A before-and-after. A spot-the-difference. A detail zoom. Anything that turns passive scrolling into active looking.

The numbers back it up. 24 posts, 12 days, 10.7% engagement rate. The craft is in the process. And the process is worth sharing.

All puzzle images generated with @NanoBanana in @AdobeFirefly. Analytics pulled from X Creator Dashboard, January 16 – February 12, 2026.