A prompt experiment testing 300 queries across leading large language models has produced something genuinely useful: structured data on how query type affects brand citation rates in AI-generated responses. The findings break queries into three categories - brand, soft-brand, and non-brand - and measure how each one influences which brands surface in LLM outputs. It is the kind of empirical grounding that AI visibility strategy has been lacking.
The headline finding is not entirely surprising, but the scale of difference between query types matters. Branded prompts - those that explicitly name a company or product - predictably generate more brand-specific responses. But the more interesting territory is what happens with soft-brand queries, where a user signals intent related to a category or use case without naming a specific brand. This is where most commercial discovery actually happens, and where the battle for AI visibility is really being fought.
The Three Query Types and Why They Behave Differently
Branded queries - those containing a specific brand name - tend to produce responses that validate, summarise, or contextualise that brand. The LLM is essentially retrieving and synthesising what it already knows about the named entity. This is valuable for brand reputation management and for ensuring your brand information is accurately represented, but it does not drive new discovery. People who already know your brand are asking these questions.
Non-brand queries - generic category or problem-based prompts - represent the widest top-of-funnel opportunity, but also the hardest to win. LLMs draw on a broad set of training data and cited sources, and brands without significant authority signals in a given topic area are rarely surfaced. The research reinforces what practitioners have observed anecdotally: authority in a topic domain, not just brand recognition, drives citation here.
Soft-brand queries sit between the two and may be the most commercially significant segment. A user asking "which CRM is best for small agencies" or "what should I look for in a UK accountant" is actively in consideration mode. LLMs generating responses to these prompts are effectively making shortlist decisions on behalf of users. Brands that appear here - without being explicitly named - have earned genuine AI-driven recommendation. That is the goal of effective GEO work.
Gemini's Brand Mention Volume Is an Outlier Worth Watching
One of the more striking findings in the experiment is the volume of brand mentions generated by Gemini compared to other LLMs. The research flags this as a notable result, with Gemini producing a higher frequency of brand citations across query types. This is worth noting for UK marketers given that Gemini is widely understood to underpin Google's AI Overviews - the feature that, based on available reports, appears above organic results for a growing share of commercial queries in the UK, though the precise proportion varies by query type and continues to change as the feature rolls out. Whether Gemini's citation behaviour in the experiment translates directly to AI Overview appearances is not yet established.
If Gemini is more inclined to surface brand names in its responses, then brands that have been properly structured and cited within Google's ecosystem - through well-optimised pages, structured data, and authoritative external references - stand to benefit disproportionately. The inverse is also true. Brands with weak entity signals in Google's knowledge graph are more likely to be absent from AI Overview citations even when Gemini is actively generating brand-heavy responses.
For practitioners, this suggests that Google-specific entity optimisation - ensuring your brand is consistently represented across Google Business Profile, Wikipedia presence where applicable, structured schema on your site, and high-authority third-party mentions - remains a priority distinct from broader LLM optimisation work. Gemini's behaviour is not identical to ChatGPT's or Perplexity's, and treating all LLMs as a single target is a strategic shortcut that will cost visibility.
What This Means for How You Build an AI Visibility Programme
The three-query-type framework gives AI visibility work a practical structure that many brands are still missing. Most GEO efforts either focus narrowly on getting cited in generic non-brand searches - which is difficult and slow - or assume that existing brand strength will carry over automatically into LLM responses. Neither approach is grounded in how these systems actually behave.
A more effective approach maps your priority query types to the appropriate content and authority-building tactics. For non-brand queries, the focus should be on topic authority - publishing substantive, well-structured content that answers the specific questions users are asking in your category, at sufficient depth and originality that LLMs treat it as a citable source. For soft-brand queries, the emphasis shifts to comparative positioning: ensuring your brand appears in contexts where it is being evaluated against alternatives.
For branded queries, the priority is accuracy and completeness. LLMs can and do misrepresent brands - getting products wrong, citing outdated information, or omitting key differentiators. Auditing what each major LLM says about your brand in response to direct branded prompts should be a standing process, not a one-off exercise. Discrepancies between LLMs are common and often reflect which training data or cited sources each model has weighted most heavily.
Connecting AI Visibility Data to PPC Strategy
There is a practical connection between this research and paid search that is easy to overlook. If LLMs are generating brand citations in response to soft-brand and non-brand queries, the users who then search for those brands directly are increasingly influenced by AI-generated recommendations they received moments earlier. This creates a dynamic where branded paid search conversions may be partly attributable to AI visibility - a channel that sits entirely outside your current attribution model.
This matters for how you interpret Performance Max and Smart Bidding performance. PMax campaigns that include brand terms will be capturing some of this AI-influenced demand, but the attribution will show it as brand search, not as a downstream effect of LLM citation. As AI search usage grows - and early data suggests brand citation rates in LLM responses are already commercially significant - the gap between what your attribution model reports and what is actually driving conversions will widen.
The practical implication is to treat AI visibility investment as a demand generation channel that feeds branded search, not just as an organic traffic play. If your brand starts appearing more consistently in LLM responses to soft-brand queries, you should expect to see branded query volume increase - and your PPC team needs to be resourced to capture it. AI visibility and paid search are not separate workstreams. They are increasingly the same funnel.
Running Your Own Brand Bias Audit
You do not need a 300-prompt experiment to get started. Pick ten queries that represent genuine soft-brand searches in your category - the kind of questions your customers ask before they start evaluating specific vendors. Run them across ChatGPT, Gemini, and Perplexity. Record whether your brand appears, where in the response it appears, what context surrounds it, and which competitors are mentioned alongside it.
Do the same with ten non-brand category queries and five branded queries. What you will almost certainly find is that your brand's presence is inconsistent across platforms, and that the reasons for its presence or absence vary. Gemini may cite you where ChatGPT does not. Perplexity may source from a third-party review site you had not considered. These differences are actionable. They tell you where to concentrate entity-building and content authority work.
Query type is a variable that shapes AI citation behaviour in measurable ways, and early data from prompt-based experiments reinforces what practitioners have been observing in client work for some time. Brands that treat all LLM optimisation as a single undifferentiated task will continue to see inconsistent results. Those that segment their visibility goals by query intent - branded, soft-brand, non-brand - and build accordingly will find it considerably easier to measure progress and justify investment.