The Discovery Gap: 97% of New Products Never Surface in ChatGPT
A 112-product study found AI models recognize products by name 99% of the time but surface them in category searches just 3–8% of the time — and on-page GEO optimization did nothing to close the gap.
Ask ChatGPT “What is Notion?” and it answers perfectly. Ask “What are the best note-taking apps?” and Notion might not appear at all. That gap — between recognition and recommendation — is the subject of the most sobering GEO study to date.
The experiment
Amit Prakash Sharma (IIT Patna, 2025) selected 112 products from the top 500 on Product Hunt and ran 2,240 queries against ChatGPT (gpt-4o-mini) and Perplexity (sonar with web search). Each product got three direct queries (“What is [product]?”) and seven discovery queries (“What are the best [category] tools?”).
The result: recognition ≠ recommendation
| Metric | ChatGPT | Perplexity |
|---|---|---|
| Direct recognition | 99.4% | 94.3% |
| Discovery visibility | 3.32% | 8.29% |
| Gap ratio | 30 : 1 | 11 : 1 |
| Products ever discovered | 6 of 112 | 31 of 112 |
The models almost always know the product exists. They almost never surface it unprompted. For a business, only the second number matters — and it is brutal.
The GEO surprise
Sharma built a composite GEO score from the levers the literature recommends: statistics density, citations, technical terminology, authoritative language, structured data, content depth. If on-page GEO worked for discovery, high-scoring products should have surfaced more often.
They didn’t. The correlation was statistically insignificant: r = -0.108 (p = 0.256) on ChatGPT, r = -0.102 (p = 0.286) on Perplexity. On-page optimization simply did not predict whether a product broke into visibility.
“GEO optimization functions as a multiplier rather than a catalyst. If you’re not being discovered at all, there’s nothing to multiply. You can’t multiply zero.” — Sharma (2025)
What actually predicted discovery
For Perplexity — a web-search model that crawls live — several off-page signals were significant:
| Predictor | Correlation (r) | p |
|---|---|---|
| Referring domains | +0.319 | <0.001 |
| Unique subreddits (cleaned) | +0.405 | 0.001 |
| Reddit mentions (cleaned) | +0.395 | 0.002 |
| Dofollow ratio | +0.238 | 0.012 |
| Product Hunt upvotes | +0.225 | 0.017 |
| On-page GEO score | -0.108 | 0.256 (n.s.) |
For ChatGPT’s knowledge-cutoff model, nothing was significant — discovery was effectively random, dominated by what was already dense in the training data. The authors call this authority concentration: established names are over-represented in training text, so the rich get richer.
The takeaway
The staged strategy the data supports: build discoverability first — backlinks, referring domains, genuine community presence — and treat on-page GEO as the multiplier you apply after you’re being found. For the full four-study picture, see our reference hub on generative engine optimization.
§ Frequently asked questions
How big is the AI discovery gap? +
In Sharma's 2025 study, products were recognized by name 99.4% of the time on ChatGPT but appeared in category-style discovery queries only 3.32% of the time — a 30-to-1 gap. Perplexity was better but still an 11-to-1 gap (94.3% vs 8.29%).
Did GEO optimization help products get discovered? +
No. The composite on-page GEO score had no statistically significant correlation with discovery on either engine (r = -0.108 and -0.102). The signals that predicted discovery were off-page: referring domains and genuine Reddit community presence.
Technology correspondent · MSc, Computer Science
Marcus Okonkwo covers artificial intelligence and search for DataBackedNews, focused on what the measurements actually show once the marketing is stripped away. He reads the papers so you don't have to — and cites them so you can.