AI Visibility Tracked Over 60 Days — Real Data from a SaaS Product
Most companies have no idea how AI models represent their product. This case study tracks AI visibility on a real SaaS product — ImportKit — over 60 days using Bersyn scans across ChatGPT, Claude, Perplexity, and Gemini.
The Setup
ImportKit is a SaaS product for importing data between platforms. Before the study began, it had no intentional AI visibility strategy. Scans ran across ChatGPT, Claude, Perplexity, and Gemini every week to measure how AI represented it to buyers.
Baseline (February 12, 2026): ImportKit scored 0.7 out of 10 on AI visibility. Only 3 buyer conversations were scanned. Zero mentions across all four AI models. When buyers asked AI "what is the best data import tool," ImportKit did not exist.
What Changed
Starting February 13, structured content was published targeting the specific conversations where ImportKit was absent:
- Comparison pages ("ImportKit vs [competitor]")
- Category definition content ("What is a data import tool?")
- FAQ content matching buyer questions
- Documentation with clear product positioning
- Third-party presence (Reddit discussions, dev.to articles)
Each piece was anchored to ImportKit's verified product identity and targeted a specific gap identified by the scan.
The Results
| Date | AI Visibility Score | CCI Conversations | Key Change |
|---|---|---|---|
| Feb 12 | 0.7 / 10 | 3 | Baseline — invisible to all models |
| Feb 14 | 1.2 / 10 | 5 | Perplexity picked up first content within 2 days |
| Feb 17 | 2.4 / 10 | 8 | Gemini started mentioning ImportKit |
| Feb 19 | 3.1 / 10 | 9 | ChatGPT included ImportKit in one comparison |
| Feb 21 | 3.3 / 10 | 10 | Peak score. Active publishing period. |
| Feb 26 | 3.1 / 10 | 10 | Score stabilized after publishing paused |
| Mar 5 | 3.1 / 10 | 10 | Plateau confirmed — no new content, no score change |
| Apr 10 | 3.1 / 10 | 10 | Score held at plateau for 7 weeks |
Score increase: 0.7 → 3.3 (4.7x improvement) in 9 days of active content publishing.
Five Findings from the Data
1. Different AI Models Respond at Different Speeds
Perplexity picked up new content within 2 days. It uses live web retrieval, so fresh content appears fast.
Gemini responded within a week, likely because it blends Google Search with training data.
ChatGPT took over two weeks to reflect new content. It relies more heavily on training data, which updates on a longer cycle. Research shows ChatGPT pre-selects brands from training data when users ask buying questions — content needs to reach training data to influence ChatGPT recommendations (Lee, 2026).
Claude was the slowest to change. It appeared to give the most weight to training data over recent web content.
2. Score Directly Correlates with Publishing Activity
The score rose steadily during active publishing (Feb 12-21) and immediately plateaued when publishing stopped (Feb 22+). Seven weeks later, the score has not changed.
This is not a "publish once and done" situation. AI visibility requires sustained content activity, similar to how traditional SEO requires ongoing optimization.
3. CSI (Category Presence) Is Harder to Move Than CCI
The CCI score (buyer conversation coverage) responded well to targeted content. The CSI score (broad category association) stayed flat at approximately 1.4 out of 10 throughout the entire period.
Broad category association appears to depend more on training data and third-party mentions than on first-party content. Getting AI to categorize a product correctly may require third-party validation — Wikipedia, review sites, industry publications, Reddit discussions.
4. Structured Content Outperforms Marketing Content
The content that moved the score most was structured and specific: comparison tables, FAQ answers that matched buyer questions, clear category definitions.
This aligns with research findings: comparison structure is the strongest predictor of AI citation (Cohen's d = 0.43), significant across all intent types (Lee, 2026). Marketing-style content with vague claims had no measurable impact on AI recommendations.
| Content Type | Impact on AI Visibility |
|---|---|
| Comparison tables with specific criteria | High — strongest citation predictor |
| FAQ sections matching buyer questions | High — direct answer extraction |
| Category definitions in first paragraph | High — query-term coverage signal |
| Documentation with specific capabilities | Medium — structured and extractable |
| Marketing copy ("best tool ever") | Zero — AI ignores sentiment claims |
| Blog posts in first-person/opinion tone | Negative — first-person tone reduces citation probability |
5. Third-Party Mentions Carry Disproportionate Weight
A single Reddit thread explaining what ImportKit does appeared to influence AI recommendations more than multiple first-party blog posts. AI models weight authentic third-party descriptions higher than marketing content.
Research confirms this pattern: ChatGPT injects brand names from its training data into 32% of fan-out queries when users ask buying questions (Lee, 2026). Brands that appear in Reddit threads, review sites, and community discussions become part of that training data entity map.
Implications for Any Product
Based on the data from this 60-day study:
| Action | Priority | Expected Impact |
|---|---|---|
| Publish comparison pages with tables | Highest | Strongest citation signal (d=0.43) |
| Remove first-person/blog tone from content | High | First-person tone is the strongest negative predictor |
| Add FAQ sections matching buyer questions | High | Direct answer extraction by AI |
| Include specific statistics with sources | High | Pages cited by 3+ AI platforms have 7x statistics density |
| Publish weekly — do not stop | High | Score plateaus as soon as publishing stops |
| Build third-party presence (Reddit, reviews) | High | Gets brand into AI training data entity maps |
| Track each model separately | Medium | What works for Perplexity may not work for ChatGPT |
| Use deep H3 subheadings | Medium | Significant in all position bands |
| Target ~2,000 words per page | Medium | Cited pages average 2,150 words vs 1,415 uncited |
Methodology
All data was collected using Bersyn, an AI visibility platform that scans ChatGPT, Claude, Perplexity, and Gemini with buyer-intent questions. Scans ran weekly on Mondays. Scoring methodology uses CCI (Conversation Coverage Index) and CSI (Category Saturation Index) combined into an overall AI visibility score on a 0-10 scale.
Run a free scan at bersyn.com to see where any product stands.