Wander: Making the internet feel worth exploring again

Search works when I can already name what I want. I missed the other kind of internet experience: stumbling into an unfamiliar archive, an odd interactive site or someone's small obsession, then following a trail I couldn't have planned.

StumbleUpon was an early influence. My question was whether discovery could feel less like a feed and more like opening one door at a time, with the person free to spend their attention on the destination. Time inside Wander isn't the goal.

So the first version is a simple loop: arrive, press Wander, open the website if it interests you, come back when you're ready, and continue. The product couldn't ask someone to define their tastes before they had experienced it.

My starting hypothesis was that wandering could be the onboarding. I wanted someone to experience a Discovery before learning any controls or being asked for an account. I think of a street magician who shows you the trick before selling you a ticket: the product should give a reason to care before it asks for any commitment.

The first screen keeps the choice small, a short premise and one button. After that, the interface adds a concept only when it becomes useful.

I think of Wander as a landscape of possible journeys rather than a list of websites. A destination matters on its own, but so does where it might lead next. That pushed me to look at relationships, variety, repetition and dead ends instead of treating a bigger collection as automatic progress.

It also needs a precise description of what the product runs today. Each Discovery is a curated record: a canonical URL, title, description, content type, topic labels, quality and obscurity scores, domain and publication status. The runtime doesn't store a semantic graph, search for paths or use embeddings. Shared topics are a rough relationship signal, and the graph I analyze is built offline from those labels, as a diagnostic.

"Never show me that as new again" couldn't be a screen-level preference. It had to survive trails, reloads and optional account creation, so it lives in the API and the database. A skipped destination is still remembered as seen, but isn't treated as a dislike. Starting a new trail changes the path, not the memory, and when no unseen destination remains, the product says so.

Wander From Here accepts any public URL as a starting point. If it matches an indexed Discovery, its topics are used. Otherwise the server derives a few topic hints from the URL's host and path, with a small hand-written map of related words. It doesn't fetch the page. It's a starting-point heuristic, not an understanding of arbitrary web content.

Not that the current engine is validated. Interesting destinations don't automatically make an interesting journey, and a tidy graph isn't evidence of discovery. I closed each experiment rather than tuning until a preferred result appeared. Whether Wander's paths feel coherent, surprising and worth following is still a human question, and the next section is where I asked it.

I ran two studies to see what happens when people meet Wander, and kept their findings separate.

Alpha01 was a first signal. It was a small consented pilot with an optional debrief, built around a hypothesis I wrote down beforehand: someone clicks Wander ten times and wants an eleventh. Firstvisit01 tested the premise without me. It was remote and unmoderated, with consent before exploring and an optional questionnaire after the fifth Discovery. It ran in Production and closed on October 27, 2026.

The research complicated my assumptions without settling the design. Several people started without any guidance, but short sessions and thin questionnaire coverage left basic comprehension and perceived value unresolved. I didn't respond by adding instructions to the interface. I want to understand what participants meant before I change the experience.

The clearest decisions came where I could separate a visible defect from an open product question. I fixed demonstrable reachability problems, and left preference and comprehension questions to research, rather than letting the speed of AI-assisted building turn every uncertainty into another feature.

The discovery promise depends on the system remembering what it has shown and on the doors leading somewhere safe. Those boundaries took real implementation work, because convenience could otherwise turn into unwanted identity linking, hidden tracking or lost history.

A local review found the sign-in and Wander From Here dialogs clipping their controls on short screens, and a recoverable account error that could leave keyboard focus in the wrong place. I fixed the dialog scrolling and made focus return to the right place. Four of nine reproduction probes had failed; all nine passed afterward, along with focused automated browser checks.

That fix is local and not deployed yet, and it isn't an accessibility certification. Physical devices, assistive technology and the hosted product still need checking.

Wander is publicly available, and v1.0 is still in progress. The strongest evidence so far is that a small first pilot reported worthwhile discoveries, while the later first-visit study showed the experience still needs more evidence about comprehension and value.

Not a perfect algorithm or a feature-complete roadmap. A trustworthy destination, an understandable first action, dependable memory, essential accessibility, and evidence that people find something worth their time.

I started with a belief that the internet could still surprise us. Building Wander has made me more careful about what counts as evidence for that belief. The product is real and available. Whether it consistently earns someone's curiosity is what I'm still working to learn.

The longer-term idea is whether relationships between discoveries could reach beyond websites, to creative work, music, events, places or experiences offline. That's an exploration, not a roadmap or a validated need, and none of it is built.

Before extending the model I'd need to learn what counts as a meaningful connection across different kinds of things, how people can steer without turning wandering into search, how to represent place and time responsibly, and how to protect privacy when discovery leaves the web. The current topic graph is already too thin for websites alone, so the first product has to earn those questions. For now, the web is Wander's only tested surface.