For AI assistants, recognition precedes recommendation

For the businesses in this test, the obstacle appeared before an assistant could weigh them against a competitor: it usually could not identify them. Recognition came before recommendation. Across 48 answers to direct recognition questions, 44 were explicit admissions of not knowing the company, with 3 uncertain replies and just 1 confident identification.

On 10 September 2026, with web search disabled, 4 assistants in their strongest available configurations answered from memory about 12 operating businesses selected nonrandomly from companies previously covered by the publications and 2 clients of Maryna Butovych. The businesses spanned 6 countries.

The test was small and nonrandom, and because it was not repeated, it cannot establish an industry pattern, nor can it establish the consistency of replies from systems that do not answer identically twice. The sample ran from a Ukrainian denim retailer and a German children’s furniture brand to engineering and energy auditing, an electric-scooter service, retail furniture and mattresses, an ERP supplier, an engineering IT agency, a numerology web application, a dental studio, a barbershop, a detailing studio and a shop selling wooden products.

MiniMax M3 identified denim.ua as a Ukrainian online retailer of American denim brands, the sole confident match in the recognition round, and that single reply supplied a business identity at the most basic level a buyer would need, namely what the company sold and where it operated.

A familiar name can point somewhere else

ROOTY exposed a different kind of absence. GPT 6 associated the name of the Ukrainian wood-products shop with an album by Basement Jaxx, giving the name a recognisable referent while failing to identify the business being asked about. A model can have an association ready for a name without knowing the company that bears it.

Kimi K3 described other name collisions involving a video-game franchise, a myth and the word for pineapple; GPT 6 also pointed to a Ukrainian weapons system when discussing another company’s name. These answers were specific, but their specificity concerned the wrong subjects.

Maryna Butovych, who manages small companies’ digital presence from Poltava, had encountered the distinction in her own client checks. In the publication’s questionnaire of 26 August 2026, she described an assistant recognising a detailing studio without recommending it, while ROOTY was not recognised as a shop and its name was treated as a music album. Her studio example concerns a business the assistant could identify but did not suggest; her shop example concerns a business it failed to identify at all.

A request for names often ended without any

The second round asked for specific companies by location and category, generating 28 answers to 7 prompts across the same engines. These were ordinary requests for businesses a user might actually contact, including children’s furniture in Germany, ERP for small manufacturers in Italy and building energy audits in Munich.

Those categories produced names such as Pinolino, Flexa and BioKinder for furniture; Zucchetti, TeamSystem, Passepartout and Odoo for ERP; and TÜV SÜD and Drees & Sommer for energy auditing. The names supplied in these categories belonged to large, widely described brands, and that pattern held wherever an answer contained names at all. Elsewhere, nobody was named by any assistant for detailing in Poltava, barbershops in Valencia or shops selling wooden products in Ukraine.

Among the sampled businesses, only denim.ua appeared in a category answer, again from MiniMax M3, linking the sole clear recognition to the sole appearance in that round. The prominent names therefore did not amount to a comparison of the sampled companies: almost all the businesses being checked remained outside the answers altogether.

A live website and model recall answer different questions

All the sample businesses had functioning websites, yet the assistants were answering with access to those sites switched off. Web search introduces a different source. It supplies material retrieved at the moment of the question, rather than whatever the model can recall unaided, and the two conditions therefore answer two different questions about the same company. That is why a response from memory and a response informed by a live page describe different kinds of presence, even when the business name in the prompt stays the same. This test concerned recall.

The result was an absence of recognisable business identities in most answers, alongside occasional associations with unrelated subjects and a narrow set of named companies in category replies. The two rounds asked different questions, and this test cannot say which answer in one round caused which answer in the other. What it does show is that in these conditions the recognition question mostly returned no business at all, and the category question mostly returned either nothing or a large brand. The studio that one assistant knew but did not suggest, reported by the expert, is a reminder that absence from a recommendation and absence from memory are two separate states.