What assistants recommend

A fixed set of prompts was run weekly, at the same time on the same day, against four consumer AI assistants, across eleven professional-services categories, for twenty-six consecutive weeks from February to August 2026 — 264 runs in total. No personalisation, no memory, no prior context carried between sessions; each run started clean. The study exists because most published advice on appearing in AI-generated answers is transposed from search engine optimisation on the assumption the mechanisms are similar, without anyone having actually watched what these answers contain over a sustained period. XONIK has a commercial interest in the answer, which the report states plainly rather than leaving implicit.

Two hundred and twenty-six distinct firms were named at least once across the study. That number alone overstates how meaningful "appearing in AI answers" is: 94 of those firms were named in only one or two of the twenty-six runs, and being named once produced no guarantee of being named the following week. Only 33 firms — about one in seven — appeared in seventeen or more of the twenty-six runs, and just 7 were named by all four assistants on the same question in the same week. The four assistants disagreed with each other more than they agreed: 167 of the 226 named firms were named by exactly one assistant.

The report tested six characteristics against the gap between consistently-named and incidentally-named firms, and most of what's commonly recommended turned out not to be the answer. Technical crawlability, directory listings and schema markup were near-universal in both groups, meaning they're a threshold to clear rather than a lever that moves anything once cleared. What separated the two groups was third-party corroboration — being cited in writing the firm didn't control, such as trade press or industry publications — present in 86% of consistently-named firms against 9% of incidentally-named ones. Publishing original research showed the second-largest gap, 71% against 6%, and is plausibly the mechanism behind the citation gap rather than a separate lever, since firms that publish something worth citing tend to get cited.

Thirty-one firms in the wider set underwent deliberate structural visibility work — entity consistency, passage-level restructuring, directory listings — tracked before and after six months. The inputs moved substantially: entity consistency rose from 52% to 89%, extractable passages from 44% to 81%. The outcomes barely moved: citation rose from 9% to 14%, and appearing in nine or more runs rose from 12% to 17%. The report states this finding runs against its own commercial interest in recommending structural work, offered as a reason to trust it, while flagging that the sections built partly on the same underlying data are a reason to check them independently.

How much the question itself matters was also tested: four prompt variants run alongside the baseline retained between 39% and 62% overlap with the baseline's top five firms, meaning a visibility measurement built on one phrasing is a measurement of that phrasing, not of visibility generally. The report's own conclusion is that structural work is necessary and achievable but not sufficient — the ceiling sits with corroboration, which is produced by other people writing about a firm, not by anything a firm can do to its own site alone.