- New experimental data show that 90% of human-written prompts fall into similarity clusters where AI brand mention rates stay consistent, directly challenging the argument that prompt tracking is unreliable.
- Mid-funnel prompts are the exception: small wording changes at the consideration stage can surface entirely different brands in AI answers, making granular tracking essential there.
People often claim that prompt tracking cannot work because every prompt is unique.
We ran the numbers and can confidently say that this concern is not justified.
This conclusion is based on two experiments:
1️⃣ We analysed two sets of prompts that were written by Rand… pic.twitter.com/3L1OSZytaC
— Malte Landwehr (@MalteLandwehr) June 16, 2026
Prompt tracking in AI search has a reputation problem. The standard objection goes like this: since every person types a slightly different question, it is impossible to monitor how your brand shows up in AI answers at any meaningful scale. A new set of experiments published by Malte Landwehr on X challenges that assumption directly, and the data is specific enough to change how marketers should be thinking about AI visibility measurement.
90% of Prompts Behave Predictably
The research covered two sets of real-world prompts. The first came from prompts written by Rand Fishkin’s followers, providing a natural sample of how actual users phrase questions. The second involved controlled variations, where a base set of prompts was altered by the smallest possible amount while keeping the intent unchanged.
The headline finding: while every single human-written prompt was worded differently, 90% of them landed in a similarity bucket where the probability of a brand appearing in the AI answer stayed essentially constant. The exact words did not matter. The intent did.
This is an important distinction for anyone running AI visibility programmes. It means you do not need to capture every possible phrasing of a query. If you have a representative prompt for a given topic and intent, that prompt is likely giving you a reliable signal for most of how real users are asking the same thing.
Where the Tracking Logic Breaks Down?
The research does identify one clear weak point: mid-funnel prompts.
At the top of the funnel and at the bottom, AI answers proved stable even when wording changed. Someone asking a broad awareness question and someone asking a highly specific purchase-intent question both produced consistent brand appearances regardless of how the question was phrased.
The mid-funnel is different. These are the consideration-stage queries where a user is comparing options, looking for recommendations, or weighing one solution against another. At this stage, minor variations in how a prompt is written can quickly bring different brands into the answer. The margin for error in tracking is much narrower.
For marketers, this is where prompt-to-AI visibility workflows need to get more granular. A single representative prompt per mid-funnel topic is not enough. Multiple variations need to be tracked because each one is likely to surface different competitors, different sources, and different positioning cues.
Prompt Style Changes What AI Returns
Beyond wording variations, the research tested how prompt structure affects brand mentions. The findings here are practical and immediately applicable.
Asking for “the best” option or requesting a list significantly increased the number of brands that appeared in AI answers. Giving the AI a role, such as “you are an SEO expert,” consistently reduced the number of brand mentions.
This means the same underlying intent can produce very different competitive landscapes in AI answers depending purely on how a question is framed. For brands trying to appear in AI recommendations, list-style and superlative queries are where visibility is most accessible. For marketers trying to understand the competitive picture, role-framed prompts may actually undercount how many brands the AI knows about.
Platform Behaviour Splits in an Unexpected Way
One of the more counterintuitive findings involves how constraints, such as limiting answers to a specific country, budget range, or category, affect brand mentions across different AI platforms.
In ChatGPT and Perplexity, adding constraints narrowed the number of brands shown, which is the behaviour most marketers would expect. In Gemini and Google AI Overviews, constraints had the opposite effect, increasing brand mentions. The research suggests this may be because constraints trigger additional fanout queries within those systems, effectively broadening the search scope rather than narrowing it.
This matters for anyone using AI visibility measurement tools across multiple platforms. A constrained prompt on Gemini is not comparable to the same constrained prompt on ChatGPT. The tracking methodology needs to account for platform-specific behaviour, not just query intent.
What This Means for How You Track AI Visibility?
The practical takeaway is cleaner than expected. Stop optimising tracking programmes around exact phrasing and start organising them around intent, funnel stage, and context. At the top and bottom of the funnel, a small set of well-chosen representative prompts will give you reliable data. At the mid-funnel stage, you need more variation, not less.
Prompt length also turned out not to matter. Conversational filler, added detail, and casual phrasing had no significant effect on AI answers as long as the core intent remained unchanged. This removes one variable that many tracking programmes have been over-engineering around.
The broader implication is that AI answer tracking is more viable as a discipline than the “every prompt is unique” argument suggests. The uniqueness is real but largely irrelevant for measurement purposes because AI systems appear to converge on consistent brand sets for consistent intents. The exception is the mid-funnel, and that is exactly where brand battles in AI search are most likely to be won or lost.
