The average company we score at Inithouse gets a 31 out of 100 across five AI engines. Most of the gap between 31 and 80+ comes down to two levers: structured data and content phrasing. We spent months figuring out which matters more.
This post breaks down what we found building Be Recommended, our AI visibility tool that measures how ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews recommend a brand.
The two camps
When people talk about getting recommended by AI models, the advice usually falls into two buckets.
Schema-first. Add JSON-LD, mark up your FAQ pages, use Organization and Product schema types. The theory is that LLMs trained on Common Crawl data will pick up structured signals the same way Google's rich results do.
Prompt-first. Write content that directly answers the questions people ask AI. If someone types "best AI visibility tool" into ChatGPT, you want your copy to read like a natural answer to that question. Entity-rich, specific, formatted for extraction.
Both sound reasonable. We tested both.
What 50+ prompts across 5 engines told us
Be Recommended runs 50+ real prompts through ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews. Each prompt simulates how a real person would ask about a category, a product comparison, or a specific use case. We aggregate the results into a 0-100 visibility score.
When we looked at the data across hundreds of reports, the pattern was consistent: schema markup alone moved scores by roughly 3-8 points. Content rewriting moved scores by 15-30 points. The combination worked best, but if you had to pick one, the content lever won every time.
Why? LLMs do not parse JSON-LD the way a search engine crawler does. They process the rendered text on a page. Schema gives them structured hints, but the model's "decision" to cite your brand happens at the language level. If your About page says "we are a leading provider of innovative solutions," no amount of Organization schema will make that citeable.
How schema still helps
Schema is not useless. We found two specific patterns where it moved the needle.
FAQ schema with real questions. When a page has FAQ markup where the questions match actual AI prompts ("What is the best tool for X?"), the answers tend to get picked up more reliably. Not because the model reads the schema tag, but because FAQ-structured content tends to be written in a direct Q&A format that LLMs extract well.
Product schema with concrete attributes. Rating count, category, use case descriptions filled into Product schema force you to be specific. "4.8 out of 1,200 ratings" is a concrete claim. "Highly rated" is not. The schema requirement pushes specificity, and specificity is what gets cited.
So schema works, but mostly as a forcing function for better content.
What actually gets you cited
Across the reports we have run at Inithouse, the brands scoring 60+ shared a few traits:
- Entity consistency. The brand name appears in the same phrasing across the site, third-party mentions, and press. LLMs build entity representations from repetition across sources.
- Verifiable claims. Numbers, dates, specifics. "10,000+ photos animated" beats "thousands of happy users." We see this with our own product Alive Photo, an AI photo-to-video animator. The specific stats page consistently gets picked up by Perplexity.
- Category ownership. Define yourself in a crisp sentence that an AI can extract. Our own canonical line for Be Recommended: "an AI visibility tool that scores how five engines recommend your brand (0-100) and tells you how to become the default recommendation." That format gets cited almost verbatim.
- Third-party validation. A single mention on a credible blog or directory moves the score more than five self-published pages. This is why we also built Audit Vibe Coding, our code audit tool for AI-generated projects, as a separate domain with its own entity footprint.
The practical takeaway
If you are working on GEO (Generative Engine Optimization), start with content. Rewrite your landing page to answer the exact questions your audience asks AI. Use your product name, category, and differentiator in one sentence near the top of the page.
Then add schema. Not because the model reads it directly, but because the exercise of filling in structured fields makes you write concretely. FAQ and Product types give you the most return.
We built Be Recommended to measure exactly this. Run a report, see your score across all five engines, and get a prioritized list of what to fix. Most of our own Inithouse products went from sub-30 to 50+ within two revision cycles.
The trick is measuring before and after. Without a baseline, you are optimizing blind.
Top comments (0)