AI search knows your product but rarely recommends it

Ask by name and AI recognizes a product almost every time. Ask by category, as in "recommend some good new products in this space," and only 6 of 112 showed up in the answers. A discovery-gap study released in December 2025 measured 112 top Product Hunt products across 2,240 queries and found a 30-fold gap between brand recognition (99.4%) and actual recommendation (3.32%)^[1]. The more striking result is that the GEO optimization score, built to target AI visibility, had no correlation with recommendation rates at all^[1].

The evidence is a single-author preprint that Amit Prakash Sharma released in December 2025 (The Discovery Gap)^[1]. The sample is 112 products drawn at random from the 2025 Product Hunt top 500 (at least 200 upvotes, a working English-language site), and the measurement is 2,240 queries run through the ChatGPT (gpt-4o-mini) and Perplexity (sonar) APIs between December 15 and 20, 2025. Scoring is a plain string match for whether the product name appeared in the answer. So the limits are baked in from the start: the sample skews to Product Hunt (developer tools, productivity, AI), only two models were tested, and the GEO score is regex-based^[1].

We build a product that repeatedly measures brand mentions in AI answers, so we have an interest in this topic. To avoid pulling the result toward "GEO is useless," the figures are reported as written and the author's reading is carried alongside. Yesterday's post, Measuring GEO's revenue impact with a marketing mix model, covered a model that enters occurrence probability as a multiplicative factor. This post covers the stage before it, where that factor is near zero for many new products.

Recognition and recommendation are different questions

Knowing a name and recommending it are different tasks for an AI. The study splits queries in two. Direct queries name the product outright, in three forms like "What is [product]?" and "Tell me about [product]," testing whether the model knows the product exists. Discovery queries ask by category, in seven forms like "What are the best [category] tools launched in 2025?" and "Recommend some new [category] products," testing whether the product surfaces on its own when the user is not looking for it specifically.

The gap in success rates between the two is the central number of this study.

Table 1. Visibility rates and gap by query type. Source: Sharma 2025, Table 4.1^[1].

ChatGPT recognizes 99.4% when asked by name, but by category only 6 of 112 products appeared, and only 26 times across 784 discovery queries (Table 4.1). Reading brand recognition as performance hides this gap entirely. Customers usually seek a recommendation without knowing the product name, so the number closer to a real new touchpoint is the discovery rate. This split sits in the same place as separating Match from Selection in the stage model of brand mentions in AI answers.

Why new products are at a disadvantage

The author frames three structural reasons new products struggle in discovery.

First, the recency wall. A product launched after the training cutoff does not exist in the world of a knowledge-cutoff model. That is a wall no optimization can clear.

Second, authority concentration. Training data is weighted toward incumbents. There is far more text about Slack than about any competitor launched this year. The rich get richer in AI visibility.

Third, query architecture. A direct query wakes a fact-retrieval mechanism ("what is X?"), while a discovery query wakes a recommendation mechanism with different biases. The two paths behave differently even inside the same model.

Together these mean a new product starts from an unmeasured zero. The distinction in unmeasured zero versus true zero becomes practical right here.

When GEO optimization does not work

The most counterintuitive result in the study is that the GEO score, built to target AI visibility, had no correlation with recommendation rates. It was a composite over six axes, statistics, citation density, technical terminology, authoritative phrasing, structured data, and content depth, each scored 0 to 100 and averaged (based on the Aggarwal framework)^[1]. The correlation was r = -0.108 (p = 0.256) for ChatGPT and r = -0.102 (p = 0.286) for Perplexity, statistically zero in both^[1].

This conflicts with Aggarwal et al.'s 2024 study^[2], which coined the term GEO and reported that signals like citations, statistics, and technical terminology raise visibility in generative answers by 30 to 40%. The author resolves it this way. Aggarwal measured optimization effects on content already appearing in search-augmented answers, while this study's data captures the earlier stage, products not yet discovered at all.

So the author treats GEO as a multiplier, not a catalyst. If you are already being discovered, optimized content can make you appear more often, but if discovery itself is zero there is nothing to multiply. Zero times anything is zero. Yesterday's GMMM entered occurrence probability as a multiplicative factor for market-level exposure, and this study shows that factor is near zero for most new products. That is why the difference between GEO, SEO, and AEO should be read as one sequence of stages.

What predicted AI search visibility

What predicted visibility instead of the GEO score was traditional SEO signals, and only on Perplexity. Referring domains emerged as the strongest predictor (r = +0.319, p < 0.001)^[1]. Referring domains and dofollow ratio were collected via DataForSEO, and the dofollow ratio was also significant (r = +0.238, p = 0.012). Among Product Hunt metrics, Product of the Day rank (r = -0.286, p = 0.002, a higher rank meaning more visibility), upvotes (r = +0.225, p = 0.017), and rating (r = +0.187, p = 0.048) were significant.

Table 2. Factors significantly correlated with Perplexity discovery visibility (p < 0.05). Source: Sharma 2025, Appendix B^[1].

Because Perplexity searches the web in real time, pages that rank well in traditional search are more likely to be cited in its answers too. Building referring domains through a backlink strategy and SEO fundamentals for the AI era is the groundwork for AI visibility.

The ChatGPT and Perplexity difference

Even on the same discovery query, ChatGPT and Perplexity behaved differently. Perplexity's discovery rate was 2.5 times ChatGPT's in the same study (8.29% versus 3.32%)^[1]. The more important difference is predictability. On ChatGPT not a single correlation was significant. Discovery looks essentially random. On Perplexity seven correlations were significant. Discovery follows a recognizable pattern.

The author calls this the architecture divide. Web-search models give you levers to pull, while knowledge-cutoff models leave only waiting. To raise visibility on a cutoff model there is little to do directly except wait for the training data to refresh. So the place to spend resources is the web-search model.

Community signals only showed up when measured accurately

Reddit signals looked uncorrelated at first, but emerged once a measurement error was cleared away. In the raw data Reddit mentions were uncorrelated with visibility (r = +0.052, p = 0.586). The author noticed something odd. Generic-name products like "Cursor," "DROP," and "Solar" had enormous Reddit mention counts yet no discovery success. The Reddit search was scraping posts about mouse cursors, music drops, and solar panels.

Removing the 52 products (46% of the sample) where generic names bred false positives changed the analysis. Reddit mentions came in at r = +0.395 (p = 0.002) and unique subreddits at r = +0.405 (p = 0.001), both significant^[1]. GitHub stars and Hacker News points, by contrast, were insignificant before and after cleaning.

The lesson is not that the signal was absent but that the measurement was wrong. Trust one dirty number as is and you reach the opposite conclusion, "community has no effect." That is why you have to measure repeatedly under the same conditions.

What this study did not see

The author grants several limits. The sample comes from one place, Product Hunt, skewing toward developer tools, productivity, and AI. Consumer products and enterprise software may differ. Only ChatGPT and Perplexity were tested, so Claude or Gemini may show other patterns. The Reddit cleaning cut 46% of the sample, lowering the statistical power of the community analysis. The GEO score is regex-based, so a more sophisticated content analysis might catch signals this one missed.

To that we add one question of our own. What set apart the 6 products that achieved ChatGPT discovery from the other 106? The author leaves this as future work. And this post rests on a single-author preprint before peer review. The figures are as written, but we do not read one study's result as a general law.

Measuring first shows the path

Putting this into practice, the first job is to measure whether you are being discovered right now. Separate brand recognition from discovery, and start by not being reassured by near-perfect recognition on direct queries. Next comes the author's staged approach. Build the SEO foundation first through referring domains, genuine community conversation, and platform success, then layer GEO optimization on once you clear the discovery threshold. Run GEO before measuring and there is no base to multiply, so you see no effect. Spend resources on the web-search models that have levers.

The measurement tool we build repeatedly asks non-brand questions like "recommend something in this space" per engine, recording how often a brand appears in the recommendation answers at the level of question, engine, and time. It is the place that first shows, as a number, whether you are being discovered. When that number is near zero, the groundwork of raising entry and rank in the recommendation census comes before content polishing. Reduced to one line, this study says the cart cannot pull the horse.