Gemini 4 Hydrogen, the model Google made the default in AI Mode on November 12, could not name one DEJAN product when it relied on its own memory. Dan Petrovic, DEJAN’s founder, published the test on November 18. The same model described the agency accurately as soon as it received search results or the company home page.

The finding matters because of how AI Mode works. Google first searches on the user’s question, passes the results along, and the model composes the reply and selects which brands to recommend. If the model holds little brand knowledge of its own, the retrieved results carry most of the weight. For a search team, that makes organic visibility the input to AI Mode visibility, not a separate project.

DEJAN asked Hydrogen what dejan.ai is about with search switched off. It opened by saying that, based on the URL, the site appeared to belong to a company specializing in AI, then guessed at SEO and digital marketing. In a second round it reasoned from the domain and the name. Petrovic’s log-probability tooling showed the model split evenly between two openings for its one-sentence description, which suggests a guess rather than recall.

The product test was sharper. DEJAN asked each model for ten products or services tied to the site, repeating the request ten times with search off. Hydrogen named none of DEJAN’s products in any of those lists. Its answers were generic items such as SEO consulting and technical audits, which fit almost any consultancy. Two larger Google models did better: Argon recalled Algoroo and LinkBERT, and Carbon recalled LinkBERT.

Grounding changed the result. With search results available, Hydrogen ran three queries and listed the DEJAN App, Algoroo, Flux, AI Rank, Chisel, and Tree Walker, and it identified the company as an Australian AI SEO and GEO agency. When the home page was supplied via the URL context feature, the three models produced nearly identical descriptions. Argon’s and Carbon’s product lists shifted from research topics to commercial services.

Petrovic reads the memory answers as a positioning problem. Years of published models, open datasets, and free tools taught the models to file DEJAN under research, not under the consulting and software it sells. Every one of Carbon’s 19 alternative descriptions included the word “research,” according to the post. That is a useful warning for any brand whose public footprint skews toward content it gives away.

Treat the evidence with care. This is one brand, tested by the brand itself, with a handful of runs per condition, all on November 18 at a low thinking setting. DEJAN also sells Tree Walker, the tool used for the analysis, and invites readers to run their own domains through it. Nothing here shows how Hydrogen treats brands with heavier press coverage, and the post does not measure how often AI Mode actually triggers a search for brand queries.

Even so, the mechanism is plausible and testable. A model that knows a name but not the offer will repeat whatever the retrieved pages say about the offer. That raises the cost of thin product pages, vague positioning copy, and third-party profiles that describe a company inconsistently.

Teams should check what their own domain returns in AI Mode for brand queries, then compare it against what the home page and top organic results say. If the retrieved pages never state what the company sells and for whom, fix that copy before spending effort on any AI-specific tactic.

Based on research published by Dan Petrovic on the DEJAN blog on November 18, 2026.