“We’re the category leader here. The AI stuff can wait a quarter.”
Ever hear that in a planning meeting and watch nobody push back?
From the other side of the same table comes the mirror image: “an AI is never going to bring up a company our size, so why spend on it.”
I sat in a room where both lines got said within about ten minutes of each other, and I couldn’t argue with either one at the time. Both are half right, which is the annoying part.
But somebody sorted 533 brands by how well known they are, ran 37,000 recommendation prompts at them, and checked where each group actually fell out. The failure points weren’t the same. They weren’t even close.
Does The Reason An AI Skips You Depend On Your Size?
It does — and the difference is the whole finding. Prominent brands almost always make the AI’s candidate list and then lose the pick; obscure ones never get onto the list at all.
Put numbers on it. Category leaders got named as the final recommendation only 25-41% of the time, despite being surfaced as candidates nearly every run. Long-tail and regional players failed at the first step in 48-52% of runs — the model never surfaced them to begin with (Unusual.ai, arXiv:2605.27439, posted 22 May 2026).
The audit covered 37,000 AI recommendations across 533 brands in 19 industries, sorted into five prominence tiers.
One thing worth putting on the table now, since it colours how you read the rest: all four authors work at Unusual.ai, which sells AI-visibility measurement. The conclusion “different tiers need different plays” happens to be the conclusion that sells a diagnostic product.
The Funnel Breaks In Three Different Places
The audit splits an AI recommendation into three stages, and that split is what makes the tier differences legible.
Stage one is discovery — can the model find you and reach information about you at all. Stage two is mention — having found you, does it actually put your name in the answer. Stage three is selection — having named you, does it pick you as the winner of the comparison.
Category leaders clear stage one almost every time and then lose at stage three, converting 25-41%. That’s not a discovery problem; that’s differentiation and positioning losing a head-to-head.
Long-tail and regional brands lose 48-52% of runs at stage one. The comparison never happens, because the model never gets far enough to hold one.
Same complaint in both cases — “we don’t show up in AI answers” — and the wall in the way is a different wall. Which means the money goes to different places.
The Best Conversion Rate Belongs To The Challengers
Here’s the number I didn’t expect. The tier the study calls “established challengers” converts at 37-52% once it makes the candidate list — the highest of any tier measured.
Higher than the category leaders at 25-41%. And without the discovery problem that eats the tiers below it.
Mid-market brands, one rung down, lose at all three stages at once and land at 34-40%. So the drop from challenger to mid-market isn’t a smooth slide down a ramp; the shape of the failure changes.
You don’t have to be the biggest name in the category to win the pick. Get onto the shortlist and the odds are genuinely good — better, in this data, than they are for the incumbent you’re up against.
Different Tier, Different Move
Since the stage you’re losing at moves with prominence, the play does too. What helps a category leader is close to useless for a regional brand.
- Category leaders (found, not chosen): differentiation content, defending the position on the comparison itself
- Established challengers (best conversion in the study): narrow the pitch by segment and push that conversion rate further
- Mid-market (losing at every stage): a hybrid, funding discovery and differentiation together
- Long-tail and regional (gone before discovery): get listed in the authoritative sources — industry comparison pages, roundups, rankings
Pick the move that belongs to a tier you’re not in and the effort doesn’t underperform, exactly. It misses.
What This Study Doesn’t Cover
Four limits, and I’d say all four out loud before this number goes into anyone’s deck.
The measurement ran in US, UK and EU markets across 19 industries. There’s no Japanese-language measurement in it, so the tier boundaries are not automatically yours.
Only two providers were tested, OpenAI and Anthropic. Gemini and Perplexity aren’t in the data at all.
It’s a single-day measurement, so nothing here tells you how much the numbers drift over time (stepping on the scale once and describing your body for the next decade).
And the authorship point from earlier. Take the funnel finding seriously; discount the “so buy a diagnostic” gradient it sits on. Maybe I’m reading the vendor angle too harshly — the three-stage split is a genuinely useful frame regardless of who published it.
Work Out Which Stage You’re Losing At Before You Spend
The move I’d make first isn’t picking a tier play. It’s one step earlier.
Ask the models the questions your buyers would actually ask, and sort what comes back into two piles: runs where your name never appears, and runs where it appears but somebody else gets recommended. Those two piles have nothing in common except the feeling of losing.
If you’re mostly in the first pile, spend on getting listed where the models look — the comparison pages, the rankings, the industry roundups. If you’re mostly in the second, the money belongs in differentiation and segment-specific pitches instead.
And if you land in the challenger tier, that’s the best conversion rate in the study. I’d think hard before copying whatever the category leader is doing, given the leader converts worse.
Sources
- Will Jack, Noah Lehman, Keller Maloney, Sarah Xu (Unusual.ai), “Prominence-Stratified Failure Modes in Retrieval-Augmented Commercial Recommendation: A 37,000-Run Audit”, arXiv:2605.27439v1, posted 22 May 2026, arxiv.org (37,000 AI recommendations, 533 brands, 19 industries, US/UK/EU markets. Recommendation modelled as a three-stage funnel — discovery, mention, selection. Category leaders converted 25-41%; established challengers 37-52%; mid-market 34-40%; long-tail and regional brands failed at discovery in 48-52% of runs. Two providers only, OpenAI and Anthropic; single-day measurement with no drift analysis; all four authors are employed by Unusual.ai, a vendor of AI-visibility measurement.)
