I Reverse-Engineered How ChatGPT Picks Software (Most Founders Have No Idea)

“What project management tool should I use for a remote engineering team?”

The answer to that question used to live on Google’s first page. Now it lives in ChatGPT’s response. And if your product isn’t in that response, you’re invisible to a growing segment of buyers.

I’ve spent the last two years studying how LLMs form product recommendations. What I’ve learned is both frustrating and encouraging. Frustrating because you can’t game it. Encouraging because the path to visibility is straightforward, just not easy.

What LLMs Actually Know

LLMs know what they’ve been trained on. This includes:

• Reddit discussions and comments
• Product documentation and websites
• Review sites like G2 and Capterra
• News articles and blog posts
• Technical forums and Stack Overflow
• Books and academic papers

They don’t have access to private information, recent events (unless using retrieval augmentation), or anything not in their training corpus. If your product is absent from these sources, that explains why ChatGPT recommends your competitor instead of you.

The Recommendation Algorithm

When you ask an LLM for a product recommendation, here’s roughly what happens:

1. Query Understanding

The model interprets what you’re actually asking. “Best CRM for startups” triggers different associations than “Enterprise CRM with Salesforce integration.”

2. Knowledge Retrieval

The model accesses patterns in its training data related to the query. This isn’t like searching a database. It’s more like remembering based on what was most frequently and confidently discussed.

3. Confidence Weighting

Products mentioned positively, frequently, and recently get higher confidence. Products with mixed or outdated mentions get lower confidence.

4. Response Generation

The model synthesizes a response, often citing specific features, use cases, or comparisons it learned from training data.

What Signals Matter Most

Not all mentions are created equal. Here’s what influences LLM recommendations:

Authenticity

LLMs are trained to recognize patterns of genuine versus promotional content. A glowing review that sounds like marketing copy carries less weight than a nuanced discussion that mentions both pros and cons. This is why creating AI-friendly content requires a different approach than traditional marketing copy.

Authentic signals include: - User complaints alongside praise - Specific use case descriptions - Comparisons with competitors - Technical details and workarounds

Specificity

Vague mentions don’t stick. “Tool X is great” is less memorable than “Tool X saved us 20 hours/week on reporting and integrates seamlessly with Snowflake.”

The more specific the praise, the more likely it forms strong associations.

Frequency

Products mentioned in dozens of threads form stronger patterns than products mentioned once. But volume without quality doesn’t help. Ten genuine discussions beat a hundred spammy posts.

Recency

Modern LLMs increasingly use retrieval systems that access current information. But even base model recommendations favor recent mentions. A product heavily discussed in 2023 but silent since will lose ground.

Source Authority

Mentions in high-quality sources carry more weight. Reddit threads in established subreddits matter more than random forums. G2 reviews matter more than unknown review sites. This is why finding the right subreddits is so important for your AI visibility. Building this kind of presence is the core of Reddit marketing for B2B SaaS.

Why You Can’t Game It

Some founders ask: “Can’t I just create fake Reddit accounts and pump up mentions?”

No. For several reasons:

Pattern detection: LLMs are trained on massive datasets that include examples of astroturfing. They learn to recognize promotional patterns and weight them accordingly.

Platform enforcement: Reddit and other platforms actively detect and remove fake content. Coordinated campaigns get accounts banned.

Negative signals: Even if fake content survives, it creates patterns that don’t match authentic discussion. The lack of critical comments, the promotional language, the suspicious timing. These become signals of inauthenticity.

Model updates: Even if you fool the current model, next quarter’s update might downweight everything you planted.

The Honest Path to LLM Visibility

Testing Your LLM Visibility

Run these tests monthly:

The direct query: Ask ChatGPT and Claude “What’s the best [your category] for [your target customer]?” Are you mentioned?

The comparison query: “What’s the difference between [Your Product] and [Competitor]?” Is the description accurate?

The problem query: “I need to [job your product does]. What should I use?” Does your product surface?

Document the responses. Track changes over time. This is your GEO baseline. For a more systematic approach to monitoring, see our guide on tracking AI brand mentions.

The Compound Effect

LLM recommendations compound. A product that gets mentioned generates more discussion, which generates more training data, which generates more recommendations.

The reverse is also true. Silence compounds into invisibility.

The SaaS companies winning LLM visibility in 2026 started building their presence in 2023 and 2024. They weren’t optimizing for LLMs specifically. They were building authentic community presence that happened to feed LLM training.

You can’t shortcut this. But you can start now. If you want to understand how this fits into the broader picture, read about the differences between GEO and traditional SEO and where to focus your investment.