48 startups, 4 known by name, 28 recommended by category

A developer ran 48 AI-native products through a two-part recognition test in a single afternoon: a model with no web access was asked "What is [brand]?", then a search-grounded model was asked the category question a buyer actually types, with the brand name withheld ^[15][16]. The model described exactly 4 of the 48 correctly, and 28 of them appeared in the category answers, many at number one ^[1][3]. The gap between those two numbers is the operational point, because the two halves are fixed on completely different timescales.

The 44 misses came back with near-identical text. For Decagon, the model returned "I do not have reliable information about the software product named Decagon (decagon.ai) and cannot provide accurate details about its features, use cases, or pricing" ^[2][5]. Decagon is a funded customer-service AI company and Mercor is a talent marketplace with a valuation most founders would envy; neither could be described from its name ^[6]. The four that passed were ElevenLabs, Suno, Runway and Cursor ^[4]. That is 8 percent recall by name against 58 percent presence by category, roughly seven times as many appearances when the question is about function rather than brand ^[1][2][3].

The mechanism is unglamorous. A model's memory of names is built from training data, which is mostly the open web talking about you: Wikipedia, hundreds of G2 and Capterra reviews, a Crunchbase profile, Reddit threads, press. A product shipped last quarter has a homepage and maybe a Product Hunt launch, so there is almost nothing to recall ^[8]. Recall therefore tracks funding and press, both of which take years, which makes it a lagging indicator rather than a verdict on the product ^[7].

Three things stack on top. Training cutoff means anything published after the last run is invisible to memory until the next one ^[9]. Generic names actively hurt: Sierra, Wonder, Peek, Ray, Brew, Marx and Nora were all in the sample, and asked "what is Brew," the model has nothing to separate the email tool from the drink ^[10]. And a JavaScript-only site sends an empty body to crawlers including GPTBot and PerplexityBot, which do not execute JS, so the one source unambiguously about you says nothing ^[11]. The first two are slow work; the third is an afternoon, and it feeds the retrieval path that is already producing results ^[17].

Those results, as reported: Decagon, unknown by name, ranks first for AI customer service automation next to Ada, Intercom's Fin and Zendesk AI ^[12]. Lindy ranks first for AI automation platforms, ahead of Zapier and Make ^[13]. Sierra places fifth for customer experience platforms ^[14]. Half the sample was invisible by name and visible by function simultaneously, which is 24 of the 44 unrecognised products, or 55 percent of them ^[18][4].

Treat the numbers as one operator's single pass. The write-up does not name the models used, does not report repeat runs, and the category prompts were written by the tester rather than harvested from real buyers ^[19]. Search-grounded rankings are the volatile half of this, so a placement shown once is not a placement shown to be stable.

What to watch: whether fixing server-side rendering measurably changes category placement on the same prompts, whether the products in the 28 hold position across repeat runs, and whether any of the 44 crosses into name recall after the next training refresh ^[9][3][1].