How AI assistants rank casinos

Entry 5, April 2026, revised August 2026

An assistant asked to name good casinos is not running a ranking algorithm in any traditional sense. It is assembling an answer from three inputs, what it memorised in training, what it retrieves when it browses, and the safety rules wrapped around the whole exchange. Everything I have observed in five months of monthly runs fits inside those three.

Training memory

Older or offline models answer from memory, and memory favours operators who were written about a lot, positively, before the training cutoff. In the April runs, memory-only answers skewed toward large legacy brands and sometimes named operators that had exited the tested market a year earlier. Memory also produces the strangest errors, one model confidently described a welcome offer that had not existed since 2023.¹

Retrieval

Browsing models behave differently, and by August 2026 most default configurations browse for this query type. The pattern in the log is consistent. The assistant issues a search close to the user's wording, reads a handful of results, and composes its answer largely from review and comparison pages. Which sites get read matters more than anything else. Across 120 recommendation runs since March, a small set of affiliate comparison sites appeared in citations far more often than any operator's own site. The assistant is not ranking casinos. It is summarising the sites that already rank for the query.

Safety rules

The wrapper decides whether an answer happens at all. Some assistants refuse recommendation prompts outright in some markets, then answer the same prompt phrased as a comparison. Refusal behaviour changed twice within the notebook's lifetime, both times without any announcement I could find. Age and jurisdiction caveats are near universal now, license mentions appear in most answers, though the license claims themselves are only as good as the source page, a problem the license entry covers in detail.

What this means for the log

Because retrieval dominates, an operator's presence in assistant answers is mostly a function of the third-party pages that describe it. Clean, consistent, widely repeated descriptions surface. Contradictory or thin coverage drops out. Names that are easy to confuse with lookalike sites do worst of all, in three runs the assistant cited a page about one brand while recommending a different one with a nearly identical name.² That confusion pattern deserves its own entry and will get one in the autumn.