You’ve typed your own company name into ChatGPT and asked “so what do you think of these guys?”, haven’t you.
I did too. The verdict that came back was a little more lukewarm than I’d hoped, and I stewed on it for a while.
And when that happens, you want to screenshot the screen and drop it in the team channel — “here’s what the AI thinks of us.”
But that one screenshot probably tells you almost nothing.
A study that measured how brands show up in AI search across roughly 100,000 answers found that “does your name come up” and “how you get talked about” behave completely differently. So let’s look at that, and carry it all the way to what actually goes in your monthly report.
What That Screenshot Looks Like At 100,000 Answers
The thing worth starting with is a big measurement study posted to arXiv in June 2026 (the paper).
It fired prompts for 102 brands at five different AI engines — ChatGPT, Gemini, Perplexity, Claude, and Grok — over and over from March through May 2026.
That piled up 102,025 responses and 149,912 citations. It’s about as broad a foundation for AI brand visibility as anyone’s laid down in one place.
One thing to flag up front: the author is a co-founder of Ranqo, the platform used to run the measurements, and the paper says outright that he holds equity in the company. So this is research by someone with a stake in the conclusion “AI visibility is worth measuring.” Read the emphasis with that discount in mind.
But even after you dock it for that, the method is careful, and the lessons it hands you are ones you can just use.
Ask The Same AI The Same Thing, Again And Again
The clever, tedious part of this study is that it asked the same question, of the same AI, many times over.
Then, for every “brand × prompt × platform” combination, it counted whether the answer had flipped from the last run. Here’s what came out.
- (did the AI talk you up or talk you down) flipped on 45.5% of combinations
- (did your name appear in that answer at all) flipped on 6.8% of combinations
So just by asking again under identical conditions, the direction of the AI’s verdict reversed on nearly half the combinations. The paper puts it plainly: sentiment is about 6.7 times noisier than mention.
It’s the interviewer who scores you differently depending on their mood that day. All your screenshot can honestly claim is “this one time, it said this.” It’s not evidence the AI dislikes you.
But “Does Your Name Come Up” Was Weirdly Stable
Now flip that 6.8% around. It means that on more than 9 in 10 combinations, the ones that show up show up, and the ones that don’t, don’t.
That steadiness held up from a second angle too. The team sorted the 102 brands into three tiers by fame and measured how often each tier appeared in answers to prompts that never named a brand.
| Brand tier | Appeared in the AI’s answer |
|---|---|
| Household-name global brands | 72.9% |
| Established mid-size brands | 43.6% |
| Niche or small brands | 11.4% |
The drops down the staircase are about 29.3 points, then about 32.2 points. It splits cleanly.
If appearance were just the AI’s whim, you wouldn’t get a staircase that tidy. Since it tracks something real — how well-known you actually are — the spine of your report should be appearance rate, not a reputation score.
So Where Does That Reputation Get Made?
Which raises the question: where is the AI even looking when it decides to talk you up or down? Break down the 149,912 citations the same study collected, and it lands like this.
- Third-party pages that profile the brand: 75.2%
- The brand’s own domain: 2.9%
- YouTube and video: 4.2%, tech and business media: 3.8%, Reddit and community: 3.3%
Your own site coming in under 3% is a number I had to read twice.
And the single most common format among the cited content was the ranked , at 21.0% of all citations — the “10 best X” style roundup you’ve clicked a hundred times.
The point is: the reputation the AI recites about you isn’t really built from the copy you wrote. What’s written about you on third-party profile pages is doing far more of the work.
How Far Should You Trust These Numbers?
A couple of caveats, though. This study is still observational — nobody ran a randomized trial to prove that a given tactic raises visibility.
And the 102 brands lean toward SaaS, retail execution, fintech, and Indian DTC (direct-to-consumer, where a maker sells straight to shoppers with no retailer in between).
So it’s a — brands picked because they were reachable — and the tiering was still done by hand. Hold the specifics loosely.
Conclusion: Make Appearance The Spine, And Read Reputation In Bulk
So the biggest thing these 100,000-odd answers teach you is that AI-visibility metrics have different personalities. Appearance is the stable floor that flips only 6.8% of the time; sentiment is the reading that wobbles on 45.5% of runs.
Which means the metric you track at a fixed cadence should be appearance. Line up “how many times did we come up in AI answers” month over month, and your moves and your results finally connect into a line.
You don’t have to throw reputation out. Just don’t read it off one run — sweep it several times and look at the distribution. Held as “3 in 10 leaned negative this month,” it turns into a number you can actually use.
And one more. Since 75.2% of citations come from third-party profile pages, polishing your own site’s wording matters less than making sure the profile pages out there carry the right facts. That’s the shorter road.
How the AI feels about you isn’t a single screenshot — it’s a count, read at a fixed cadence. Start by counting how many times your name comes up.
Sources
- Pratyush Kumar (Ranqo), “Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines”, arXiv:2606.20065, June 18 2026, arxiv.org (5 engines, 102 brands, 102,025 responses, 149,912 citations. Sentiment flipped 45.5% vs mention 6.8%; tier appearance rates 72.9% / 43.6% / 11.4%; citations from third-party brand pages 75.2% and own domain 2.9%; listicles were 21.0% of all citations. The author is a Ranqo co-founder holding equity.)
