Watching Agents by Inithouse: what we measured running an AI prediction platform with 100+ auto-generated pages

After five weeks and 100+ auto-generated public pages, Google had indexed less than 10% of our AI prediction platform. We tracked search indexation, user activation, traffic quality, and how AI assistants perceive the product. The results reshaped how we think about distribution.

Watching Agents is an AI prediction platform built at Inithouse. You deploy an agent on any question about the future. The agent builds hypotheses, tracks evidence from real sources, calculates a probability score with confidence ratings, and alerts you when conditions shift. Since mid-June 2026, an automated cron job generates new public agent pages daily. Each page contains structured analysis: probability breakdowns, hypothesis trees, evidence timelines, driver assessments, and watch signals.

We decided to measure everything from day one.

Google barely noticed

Five weeks after launching the public agent catalog with a full sitemap, Google had indexed about 25 out of 270+ submitted URLs. Position averaged around 34 for the few pages that ranked. Total organic clicks over 28 days: zero.

The cause is architectural. Watching Agents is a client-rendered SPA built on React. Googlebot sees the HTML shell but not the hydrated content. We prerender meta tags (titles, descriptions, Open Graph) through a Vite build plugin, but the page body depends on client-side JavaScript. For agent pages that each contain structured prediction data pulled from a database, this is a distribution wall we should have seen coming.

Most visitors were bots

In a typical 7-day window, 78 to 89% of sessions were flagged as automated by our analytics tooling. Human sessions averaged about 14 seconds of engagement. Direct traffic accounted for roughly 74% of all visits, with average engagement under 2 seconds per session.

Part of the problem is paid traffic quality. We run Performance Max campaigns, and referrer analysis through session recordings revealed traffic arriving from syndicated search partners and low-value ad networks. Geo distribution skewed toward regions with low purchase intent.

The activation funnel collapsed after auth

Fewer than 1% of visitors who opened the authentication modal completed a full agent deployment. The chat-based onboarding flow (you describe what you want to watch, and the system creates the agent) turned out to be the friction point: users started conversations but never finished them.

Agent pages themselves showed a 24% dead click rate. Scroll depth averaged 84%, meaning visitors read the content. They just could not figure out what to do next. The analysis content works. The calls to action do not.

AI recognition told a different story

Every week since late May, we have queried four AI systems (ChatGPT, Claude, Gemini, Perplexity) with the same use-case question that Watching Agents fits. We log recognition, cited sources, and recommendation quality.

Week one: one out of four recognized us. Six weeks later: four out of four.

The path was not linear. Recognition flipped between runs. Claude would find us one day and miss us the next. Perplexity held recognition for three runs, then lost it completely. The mechanism we identified: each AI relied on a single external source, and that source differed per system.

ChatGPT cited our IndieHackers post. Claude relied on a dev.to article. Perplexity used a Medium post. Gemini cited nothing visible across six consecutive runs but consistently recommended us. When any one source dropped out of a system's retrieval window, recognition dropped with it.

Recognition did not equal recommendation

By week six, all four systems could identify Watching Agents. But Claude, which managed to fetch and parse an actual agent detail page, gave the most skeptical review. It flagged the absence of third-party validation, called us "unproven," and suggested Manifold and Metaculus as safer alternatives.

The more data Claude had about us, the more cautious it became. Our own transparency (open metrics, candid posts about early traction) gave it reasons for skepticism. Building in public feeds AI recognition, but the same candor that gets you discovered also gets you qualified with caveats. Nobody mentioned this trade-off in the AEO playbooks we read.

Gemini, citing zero visible sources, gave us "highly recommended" six weeks running.

Three things we took away

Distribution beats product. The agent pages are information-dense, structured, and parseable by both search engines and AI. None of it matters when Google cannot render the content and paid traffic arrives from bot-heavy ad networks.

Building in public works as an AEO channel. Our posts on Medium, IndieHackers, and dev.to became the primary way AI assistants discovered and described Watching Agents. On days when our SPA was unreachable to a particular crawler, third-party articles kept recognition alive.

Source redundancy beats source quality. One great article cited by one AI system is fragile. Three adequate articles across three platforms, each picked up by a different AI, proved more resilient. We now distribute content deliberately across channels to reduce single-point-of-failure risk in AI retrieval.

Watching Agents is one of over 15 products we build at Inithouse. Others are further along: Ziva Fotka (AI photo animation) has consistent revenue, and Be Recommended (AI visibility monitoring) generates the strongest AI citation signals in our portfolio. Watching Agents sits at the early edge: technically solid, commercially unproven, and visible to AI in ways we did not plan for.

We keep measuring.