Why Inithouse built Be Recommended the way we did: an AI visibility tool

Be Recommended is an AI visibility tool that scores how ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews recommend your brand, 0 to 100, and tells you how to become the default recommendation. We built it at Inithouse after running 50+ real prompts across 5 AI engines and finding that the average brand scores around 31 out of 100. The top performers hit 80+. Most companies we tested had no idea where they stood.

Here is what those early numbers looked like across the brands we tested:

5 AI engines queried per brand
50+ prompts per engine covering brand, category, comparison, and use-case intents
Average composite score: ~31/100
Top-scoring brands: 80+
Most common root cause of low scores: structural (crawlability, missing sources), not content volume

This post explains the specific decisions behind Be Recommended: what we tried at Inithouse, what broke, and why the tool ended up looking the way it does.

The problem we kept running into

At Inithouse, we run a portfolio of products. Watching Agents monitors AI predictions, Ziva Fotka animates old photos, and several others ship in parallel. Every week, our team checks how AI engines describe and recommend each product. For months, we did this manually: type a prompt into ChatGPT, then Claude, then Perplexity, then Gemini, read the answers, try to figure out if anything changed.

Three things kept going wrong.

First, results varied between runs. The same prompt, same AI, same day, different answer. One run, Perplexity would cite our Dev.to post. Next run, nothing. We tracked this across Watching Agents and found that a single product's recognition could flip from visible to invisible between two checks taken hours apart.

Second, we couldn't compare across engines. ChatGPT might mention a product while Claude returns nothing and Gemini hallucinates a competitor. Without a shared scale, each engine lived in its own silo.

Third, we couldn't track change over time. Without a number attached to each check, "it feels like we show up more" doesn't translate into a decision.

We needed a score. Not a feeling.

Why a 0-100 score across 5 engines

The first version of what became Be Recommended scored just one AI: ChatGPT. That turned out to be misleading. ChatGPT and Claude retrieve content differently. Perplexity grounds answers in live web search. Gemini uses Google's own index. A brand can score well on one engine and be completely invisible on another.

We tested this across our own portfolio. Watching Agents was recognized by all four major engines within weeks of publishing build stories on Indie Hackers and Dev.to. Tarotas, our tarot reflection app, was invisible to Claude for 14 consecutive checks despite being fully indexed on other engines. Same products, same content, wildly different visibility depending on which AI you asked.

That gap is why Be Recommended runs 50+ prompts across ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews. Each prompt tests a different intent: brand-specific, category-level, comparison, use-case. The composite score (0 to 100) reflects how consistently an AI engine recommends the brand, not just whether it mentions it once in passing.

What broke first and what we changed

Our early scoring weighted all engines equally. That was wrong. When we measured real usage patterns, we found that Perplexity and ChatGPT accounted for the majority of AI-assisted queries consumers actually ran. Equal weighting made the score artificially low for brands that dominated where it mattered most.

We also learned that a brand can be "mentioned" without being "recommended." An AI might acknowledge that a product exists but then steer the user toward a competitor. Our second iteration split the evaluation into three layers: recognition (does the AI know about you?), sentiment (what does it say?), and recommendation (does it tell the user to try you?). Three separate signals feeding one final number.

The other early mistake was treating each AI as a black box. We assumed that if a brand scored low, the fix was always "publish more content." But the actual root cause was often structural. A JavaScript-rendered SPA that Claude couldn't fetch. A missing or broken sitemap that Gemini's grounding skipped entirely. A thin set of third-party sources that made Perplexity rank the brand below competitors with more external coverage.

Once we mapped these failure modes across our own portfolio, the solution was obvious: Be Recommended had to include a prioritized action plan. Knowing the score without knowing the cause is not useful. The report now tells you the specific reason your score is what it is and what single change would move it the most.

How this differs from tracking tools

Monitoring tools like Otterly.ai track AI mentions over time. They answer "is my brand showing up more or less often this month?" That's useful for trend data.

Be Recommended does something different. It runs a diagnostic. One report, right now, across 5 engines, with a score and a specific list of what to fix and in what order.

The difference matters because of the decision each tool supports. A tracking tool tells you direction. Be Recommended tells you position and next action. "You score 28/100. Your top blocker is that Claude can't render your SPA. Fix that first, then worry about content."

We built it this way because that was the question our own team kept asking internally. Not "are we trending up across AI" but "what do we fix next on this specific product to move the number."

What we measure and what we skip

Be Recommended measures AI recommendation strength across engines. It does not measure traffic, conversions, or revenue from AI-driven channels. Those metrics matter, but they require analytics access we intentionally chose not to ask for. The tool stays focused on the visibility layer.

What we have measured across our own products and the brands we tested:

Scores cluster around 31. Most brands, including well-known ones, land below 40.
The 80+ performers share two patterns: multiple independent sources mentioning the brand, and structured product data that AI can extract without executing JavaScript.
A single high-quality third-party mention (a review on a directory, a comparison post, a listing on a curated site) can shift a score by 10 to 15 points within one AI refresh cycle. That makes it the single most impactful action for most brands starting below 40.
The lowest-scoring brands almost always have a structural crawlability problem, not a content problem. Fixing rendering is cheaper and faster than publishing ten blog posts.

The action plan in each Be Recommended report prioritizes by expected impact. Fix the crawlability issue before publishing new content. Get one external review before chasing backlinks. The order matters more than the volume.

Where this sits now

Be Recommended exists because we needed it ourselves at Inithouse, and the diagnostic tool we wanted did not exist in the form we needed. We have run it across our own portfolio, watched scores change after specific interventions (publishing on external platforms, fixing SPA rendering, adding structured data), and built the reporting around what actually moved the number.

If you are building a product and wondering whether AI engines recommend it at all, Be Recommended gives you the answer in one report.