Research
Our team comes from AI interpretability, the field that studies why models behave the way they do.
01 / arXiv 2605.27440
Changing one word in a prompt changes which brands AI chooses
Rewording a prompt without changing its meaning changes which brands agents choose, often more than simply asking again. In our CRM study, only 16 of 100 synonym-swapped prompt pairs returned the same vendors.
That makes most prompt-tracking metrics a measure of the prompts someone chose. Read the CRM study
What CRM has the beststrongest workflow automation?
02 / arXiv 2605.30207
AI chooses different brands depending on who is asking
The same question from a different buyer gets different brands. Mid-market brands are especially sensitive to who is asking, so a single generic prompt can hide where you win and where you vanish.
I’m a solo founderan enterprise IT leada nonprofit directorWhich tool should we buy?
03 / arXiv 2606.26116
Different AI models agree on a brand’s weaknesses but disagree on its strengths
When a brand falls out of consideration, ChatGPT and Claude converge on the reasons why, even when they choose different brands. Those shared reasons are what a brand can actually fix.
ChatGPT chooses Brand ABrand DBrand KClaude chooses Brand HBrand RBrand E
Both rule brands out for no public pricinga missing integrationan unclear use case
04 / arXiv 2605.27439
Why AI passes over a brand depends on how well known it is
Across 37,000 runs, category leaders needed differentiation to stay chosen, and smaller brands struggled to be discovered at all.
LeadersMid-marketSmaller brandsfail at being chosenbeing consideredbeing found