How AI assistants choose casino sources

Entry 13, August 2026

Every browsing run in this notebook leaves a citation trail, and by August the trail is long enough to say something about selection. When an assistant answers a casino question, which pages did it actually read, and is there a pattern in what it picks? The short answer from five months of logs, the pattern is strong, boring, and worth writing down precisely because it is boring.

Shape matching comes first

The clearest regularity is that assistants pick pages shaped like the question. Ask an ordered question, which sites are best, and the citations are dominated by ordered comparison pages. Ask a factual question, is this operator licensed, and register pages and review sites with a visible license field appear instead. Ask about a bonus and terms summaries win. The systems seem to prefer a page that has already done the assembly work over several pages that would each need reading in full.¹

Repetition beats authority

A claim that appears on one authoritative page gets cited sometimes. The same claim spread across four or five modest pages, worded differently but agreeing, gets repeated almost every run. In the citation logs this shows up as clusters, the assistant reads three pages that say the same thing and then states it without hedging. Where sources disagree, some models flag the conflict, most quietly follow the majority. I have not seen a run where a single high-authority page overrode three agreeing lesser ones.

What never gets read

Some absences are as consistent as the presences. Pages that need JavaScript to show their content are almost never cited, the fetched version is presumably empty. Forum threads appear rarely and only for complaint-type queries. Operator sites themselves are cited less than newcomers to these logs expect, mostly for terms pages and payment tables, almost never as evidence of their own quality. An assistant looking for a judgment wants a third party to have made it.²

Freshness, within limits

Dated pages do better than undated ones, and 2026-dated material shows up disproportionately in this quarter's runs. But freshness is a tiebreaker, not a gate. Stale comparison pages from 2024 still surface when nothing newer covers the query, and the assistant repeats their contents without noting the age. Only one model in the panel volunteered that its main source looked outdated, once, in July.

Why this matters for the log

Source selection is upstream of everything else this notebook measures. Named operators, license claims, safety framing, all of it inherits from a handful of pages chosen in the first seconds of the run. The next quarter's prompt set adds a block that asks assistants to justify their sources directly, to test whether the stated reasons match the observed picks.