Same users, different habits: AI search isn't Google with extra steps
Panel data on 16,310 matched users shows the same people search Google and prompt LLMs differently, which breaks keyword-based brand tracking.
Key takeaways
- iPullRank matched 16,310 users across Google and LLM platforms for a month and found sharply different query behaviour.
- 87% of Google queries run four words or fewer; LLM prompts are far longer and more synthesis-oriented.
- Keyword rank and LLM citation are separate metrics measuring separate user behaviours, not two views of one signal.
- Multilaterals and financial brands optimised for Google's short-query regime may be invisible in LLM synthesis answers.
- Brand visibility audits need prompt-level sampling, not keyword-volume inference, to track LLM-era research behaviour.
Eighty-seven percent of the time, a Google query runs four words or fewer. iPullRank's panel study, matching 16,310 users across two Datos datasets to compare the same people's behaviour on Google and on LLMs over a single month, found that prompts look nothing like that. Users who type "best CRM for small business" into Google type something closer to a memo into ChatGPT. The keyword and the prompt are not two versions of the same query. They are two different cognitive acts, performed by the same person, often in the same hour.
This matters more than it sounds, because most brand-tracking infrastructure built over the past two decades assumes the opposite. Rank trackers, keyword tools, share-of-voice dashboards: all of it is built on the premise that search behaviour is stable and that what varies is only the result. iPullRank's panel data says the behaviour itself is not stable. The same user switches modes depending on the platform, which means a brand's "visibility" in Google and its "visibility" in ChatGPT are not two readings of one signal. They are two separate instruments measuring two separate things.
Why the keyword list stops working
Marketing teams at financial-services firms and industrial groups have spent years building keyword taxonomies: the terms a prospective buyer types when researching a payments processor, an ESG rating, an industrial automation vendor. Those taxonomies were built from Google query logs, under the four-word regime. They were never designed to anticipate a user writing "compare the total cost of ownership between X and Y assuming a mid-size manufacturing deployment," which is the kind of thing iPullRank's data suggests now happens routinely in LLM interfaces.
A brand optimised for the four-word query is optimised for a behaviour that a meaningful share of its own audience has partly abandoned, at least for research-stage tasks. The keyword list still matters for Google, which still handles the overwhelming majority of global query volume. But treating prompt behaviour as a subset of keyword behaviour, as most SEO and content strategies currently do, means building for the wrong unit of analysis in the fastest-growing channel.
The multilateral and philanthropic case is sharper than the corporate one
Consider how this plays out for an institution like a UN agency or a policy foundation. A journalist or policy researcher Googling "UNDRR disaster risk report 2024" is doing keyword retrieval: they know roughly what exists and want the document. The same researcher, six months later, asking ChatGPT "summarise the main disagreements between UN agencies on disaster risk financing since 2020" is doing something else entirely: synthesis across sources, with the LLM selecting which institutional voices get folded into the answer and which do not.
Multilaterals and policy bodies have historically measured their digital reach through citation counts, document downloads, and search rank for their own terminology. None of those metrics capture whether an institution's position gets selected as one of the two or three sources an LLM chooses to synthesise. That is a different competition, fought on different terrain: structured, extractable, well-attributed content rather than keyword-matched landing pages. An institution that writes for Google's four-word query and an institution that gets cited in a synthesis answer are, per iPullRank's framing, optimising for different user intents entirely, not different formats of the same intent.