How Many Queries Do You Need to Measure AI Visibility?

It's one of the first questions anyone asks about AI visibility measurement, and it seems like it should have a simple answer. Pick a number, run that many prompts, and read the results.

The industry has certainly picked numbers. Look across published studies and commercial tools, and you'll find collection budgets that differ by nearly an order of magnitude: 25 prompts here, 50 there, a few hundred somewhere else. Rarely does anyone explain where their number came from.

That's because there isn't a number that works everywhere. The amount of data a measurement needs depends on the platform, the topic, and how the citations happen to be distributed, none of which we know in advance. What we can know is when we've collected enough.

That turns out to be the more useful question.

One size fits none

In an earlier study, we measured how quickly confidence intervals on Citation Share narrowed as prompts accumulated. We set a target width of five percentage points and asked how many prompts each platform needed to reach it.

Gemini needed about 40 to 50. Perplexity needed about 100. SearchGPT needed 150 or more, and for one topic, it never reached the target within 200.

A budget of 50 prompts would be adequate for Gemini and badly inadequate for SearchGPT. A budget of 200 would be wasteful for Gemini and still not enough for some SearchGPT topics.

Nor is the problem just a matter of platforms. In a follow-up study of ten topics on three platforms, no platform was consistently the fastest or the slowest. Each one converged earliest on at least one topic and latest on another. Speed depended on the combination of topic and platform.

What makes a prompt worth more

Two properties drive most of the variation.

The first is citation density. Gemini's answers carry roughly 20 to 50 citations each, depending on the topic, while SearchGPT's carry 4 to 6. More citations per answer means more information per prompt, all else being equal.

The second is clustering, and it's why all else isn't equal. Citations within a single answer aren't independent: An engine that cites one page from a site often cites another from the same site. Statisticians measure this with the design effect, the factor by which clustering inflates variance. In our data, Gemini's median design effect was about 3.9, so each Gemini response contributed roughly a quarter of an independent observation. SearchGPT's was about 2.1, so each response contributed roughly half.

The result is counterintuitive. Despite producing about one sixth as many citations, SearchGPT (when it converged at all) did so at response counts comparable to Gemini's. The same nominal budget simply doesn't buy the same amount of information on different platforms.

There's also a practical wrinkle with SearchGPT: It returns some answers with no citations. For one topic in our data, about 17% of prompts came back without a single source. Those answers still count (a prompt that produced no citations is a prompt where no one was cited), but they add nothing to the citation data. So 120 prompts on SearchGPT might yield only 100 or so answers that actually say anything about which sources it prefers.

The law of diminishing returns

Confidence intervals narrow with the square root of the sample size. Doubling the number of prompts shrinks an interval by about 29%. Cutting it in half takes four times the data.

That relationship sets the economics of measurement. Early prompts are enormously valuable, because each one reduces uncertainty noticeably. Later prompts add less and less. At some point, additional collection costs more than the precision it buys. The goal is to find that point for the measurement in front of us, rather than picking a number in advance and hoping it lands in the right place.

Two tests for "enough"

In our research, we defined enough as the point where two conditions hold at the same time.

The first is rank stability: Has the ordering stopped changing? Early in a run, the ranking of cited domains shifts with every new response, because the eventual leaders haven't yet pulled away from the field. So we track the correlation between the current ranking and the ranking a few responses earlier, weighted so that swaps near the top count more than swaps in the tail. As data accumulates, that correlation rises and then levels off. We declare stability once the curve has plateaued.

Note that we detect the plateau itself rather than waiting for the correlation to cross a fixed threshold, such as 0.95. A fixed threshold fails because sparse platforms often can't reach it, even after their rankings have settled. With only a few citations per answer, the correlation between successive snapshots levels off below 0.95 for purely mechanical reasons. Across our 30 platform and topic combinations, only 19 sustained a correlation above 0.95 within 125 responses. The plateau test declared stability in all 30, somewhere between response 33 and response 75.

The second condition is sufficiency: Is the signal larger than the noise? A stable ranking can still be too imprecise to act on. If the top two domains have nearly identical shares, a stable order between them isn't a meaningful one. So we compare two quantities across the domains that are cited reliably: how spread out their shares are (the signal) and how wide their confidence intervals are on average (the noise). When the spread exceeds the typical interval, the measurement has enough resolution to support comparisons.

Neither test asks us to specify a target prompt count, a target correlation, or a target interval width. Both read their answers from the data as it's collected.

What the tests found

Of the 30 combinations, 27 met both conditions within 125 responses. The median stopping point was 42 responses for Gemini, 51 for Perplexity, and 37 for SearchGPT (among the topics where it converged). Three SearchGPT combinations didn't converge at all: Their shares were packed too tightly for the intervals to separate them within the budget.

A single topic illustrates the range. For "unusual travel destinations," Gemini was ready at 41 responses. Perplexity stabilized at 49 but didn't reach sufficiency until 94. SearchGPT stabilized at 46 and never reached sufficiency within 125.

Both conditions matter. To check the stopping rule, we compared each stopped ranking to the ranking at the end of the full collection. Stopping at rank stability alone produced one ranking that agreed with the final result at a correlation of just 0.36. Requiring both conditions raised the worst case to 0.72. And in every combination where the sufficiency test delayed the stop, the ranking at the stopping point agreed more closely with the full-collection ranking.

Finally, no budget below 94 responses would have certified every combination that converged, yet several of them were ready by the mid-30s. A fixed budget would have been too small for some and wasteful for others.

So, how many?

If you need a planning figure, decision-grade measurement typically takes more than 100 prompts per platform. Treat that as a starting point for budgeting, though. The number to rely on is the one your data gives you.

In practice, that means checking stability and sufficiency rather than counting prompts, and assessing each platform separately, since an all-up result is ready only when every platform in it is ready. Because every platform receives the same prompts (which is what makes cross-platform comparisons paired and fair), it also means sizing the shared prompt set for whichever platform needs the most data on that topic. (That's often, but not always, the one that cites the fewest sources or returns the most citation-free answers.) It means keeping the prompt set balanced and evaluating readiness after complete cycles of prompt types and intents, so that every comparison rests on the same mix of questions. (With 14 prompt types and 4 intents, for example, a balanced cycle is 28 prompts, and 112 or 140 prompts give every option equal representation.) It means treating every slice of the data, such as a single prompt type or intent, as a smaller measurement that has to earn its own readiness. And it means accepting "not yet" as a valid answer. Some combinations need more data than a standard budget provides, and a measurement that says so is more useful than one that reports false precision.

That's why IQRush builds balanced prompt sets and reports both conditions alongside the results: whether rank stability has converged, and whether sufficiency has been reached. They appear on the Insights / Citations tab (and through our agent and MCP server), and they update as you slice the data, so a narrow filter shows its own readiness. How much weight to give them is your call. Our job is to make sure you never have to guess.

Asking how many queries we need assumes the answer can be known before we start. It can't. But we can know, while we're collecting, whether we have enough.

References

In this series

• AI Visibility Isn't a Number. It's a Range.

• Does Rank Still Matter?

Research papers

Sielinski, R. (2026). Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement. arXiv:2603.08924.

Sielinski, R. (2026). From Stochastic to Stable: Rank Stability and Structural Sufficiency in AI Visibility Measurement. arXiv:2607.10341.