Run the same prompt through SearchGPT twice. You'll get different citations. Do it a third time, different again. A new IQRush preprint (corroborated by a separate University of St. Gallen study from April) puts a number on what most ops teams already suspect: single-run AI visibility rankings are statistical noise, and you need 33–94 cited answers per prompt before the order stabilizes enough to trust.
If your AI visibility tool pulls data once and hands you a rank, that rank is a coin flip dressed in a clean UI.
The measurement problem, quantified
The IQRush paper tested 30 platform-topic combinations across SearchGPT, Gemini, and Perplexity. In every case, early rankings jumped around as new answers came in. The order only settled once two conditions held simultaneously: the ranking stopped changing, and the gap between top sites exceeded the margin of error for each. That took between 33 and 94 cited answers.
Three of the 30 tests never stabilized, even after 125 questions. All three were on SearchGPT, where the top sites were too close to separate. The honest output there is "insufficient data," but most dashboards don't have that label.
This isn't one team's finding. Researchers Julius Schulte, Malte Bleeker, and Philipp Kaufmann at the University of St. Gallen ran a separate dataset and reached the same conclusion: a single reading is unreliable. You have to sample repeatedly.
Why this breaks your current reporting
Cross-engine agreement on brand recommendations sits at about 17%. A "win" in Gemini may not exist in Perplexity. And the relationship between traditional SEO rank and AI citations is collapsing: overlap between top Google ranking pages and AI-cited sources reportedly fell from around 70% to under 20% by May 2026.
So the ops team is stuck with a metric that's unstable within a single engine, inconsistent across engines, and decoupled from the organic rankings they've spent years optimizing. That's three layers of uncertainty stacked on top of each other.
Rand Fishkin's take is blunt: ranking positions in AI are "noise." He argues the only statistically valid signal is visibility frequency, the percentage of times your brand appears across many runs. SparkToro's own data shows top brands landing in the 55–77% visibility range, a figure that holds up even when the specific list composition changes run to run.
What a valid measurement setup looks like
Here's the practical translation for marketing ops:
- Metric: Visibility percentage (frequency of appearance across runs), not rank position.
- Sample size: 60–100+ runs per prompt, per engine. Count only answers that include citations. On SearchGPT, expect a 15–20% miss rate where answers return no citations at all, so budget extra queries.
- Segmentation: Report by engine separately (17% cross-engine agreement makes aggregation misleading). Segment further by query type: AI Overviews activate on 64.7% of question-form queries but only 13.7% overall.
- Confidence reporting: Show ranges, not point estimates. If two brands' confidence intervals overlap, the difference isn't real. Say so.
- Before/after testing: A three-point citation share increase after a content change means nothing from a single run. Measure before and after with multiple runs each, or you can't separate signal from drift.
One more thing worth flagging. The paper's validation is internal: it checks early rankings against final rankings from the same collection, not against an external ground truth. The St. Gallen replication is what gives the finding real weight. Treat the specific numbers (33–94) as the shape of the problem, not a universal lookup table.
Where this leaves GEO experiments
A study of 102 brands across five engines found 77.5% of brand-prompt-engine combinations are deterministic: always cited or never cited. The remaining 22.5% fall into a stochastic middle where randomness dominates. That split matters for experimentation.
For deterministic prompts, a content or distribution change should produce a clear before/after shift. For stochastic ones, you need larger sample sizes to detect real movement, and you should expect some experiments to end with "inconclusive" as the honest result.
B2B SaaS sits in an interesting position here. The vertical leads in median AI visibility scores (roughly 62–68 out of 100), yet median AI Overview citation rate is about 3%. Brands can rank for thousands of keywords organically and still be invisible in AI surfaces. The authority signals that seem to matter most for LLM citations aren't owned blog content but third-party sources: G2, Capterra, Reddit threads, analyst mentions.
The uncomfortable bottom line
AI-referred traffic grew 527% in early 2026. The channel matters. But the measurement infrastructure is where ad reporting was fifteen years ago, before confidence intervals and statistical significance became table stakes. The tools will catch up. Until they do, the job falls on ops to run the check more than once, report the range instead of the point estimate, and resist the urge to call noise a win. A dashboard that says "not enough data" is worth more than one that prints a confident number every time you ask.
