Two assistants agreed on the top plumber 4.2% of the time. One agreed with itself 7% of the time.

Divide 1,487 queries by 50 metros and 10 verticals and you get roughly three prompts per cell ^[9]. That arithmetic sets a ceiling on how much of the report can be read as category insight. The finding that personal injury law, dentistry and auto repair show different citation behaviour ^[6] is pooled across markets by necessity, because no single metro-vertical pair has enough runs behind it to separate a pattern from one query's luck. The headline figure survives better, since it pools everything: about 62 of the 1,487 identical prompts produced the same top business in both engines ^[10].

The comparison worth sitting with is inside one engine, not between two. Steady Demand's benchmark puts Gemini's top-match stability at roughly 7%, against about 90% for Google's local pack ^[3]. Cross-engine agreement came in at 4.2% ^[1]. Two different assistants therefore agreed about 60% as often as Gemini agreed with its own earlier answer ^[13]. The gap between ChatGPT and Gemini is only a little wider than the gap between Gemini and Gemini.

The grounding figures point the same direction. Re-running queries word for word, cited sources matched about 40% of the time ^[4], which means roughly 60% of the evidence base moved between runs ^[11]. The report calls that grounding drift ^[4].

This complicates the study's own prescription. Reporting each engine separately ^[14] is necessary, but it does not produce a per-engine score either, because the per-engine answer is not stable enough to be a score. What can be measured is a rate: share of N runs in which a business is named, with N large enough to have an error bar. The study's design cannot deliver that at vertical level with three prompts per cell ^[9], which is a limit on the research, not on the method.

The asymmetry in sources is the part that translates into work. Gemini leaned on business websites; ChatGPT drew more on Reddit and directories ^[5]. Steady Demand reads that as a signal that clear on-site service information matters more in Gemini, while listing accuracy and public discussion matter more in ChatGPT ^[5]. That is inference from citation counts, not a tested intervention, and the mix changes by category ^[6], so the two work orders it implies are bets rather than fixes. National platforms including Angi, the Better Business Bureau and Reddit recur across metros ^[8], which at least gives a short list of things worth auditing regardless of which engine you are chasing.

One boundary is worth stating plainly, because the report states it: the study counts citations and top-name outcomes, not conversions or ranking quality ^[7]. Nothing here establishes that being named by an assistant produces a booking. Anyone moving spend on the strength of a 4.2% divergence is buying instrumentation for a channel whose demand has not been measured in this data.