Can you trust a single AI citation snapshot?
Not by itself. A 2026 paper by Ronald Sielinski found that the same query sent to an AI answer engine at different times can produce different answers and cite different sources (Sielinski, 2026). So a one-run "you were cited 12% of the time" figure is a sample, not a fixed fact.
At a glance
- Finding: Citation share behaves like an estimate drawn from a distribution, not a stable score.
- Why it matters: A competitor "beating" you in one report may be noise.
- Limit: The study covered three consumer product topics and three platforms. It is based on the paper's abstract, so I cannot say how far the numbers extend to your market.
What the research tested
The author sampled Perplexity Search, OpenAI SearchGPT and Google Gemini on three consumer product topics. Collection ran two ways: daily for nine days, and every ten minutes in a high-frequency run.
The abstract reports that citations followed a power-law shape. A few domains appear often, and a long tail appears rarely. It also reports substantial variation between repeated samples.
Why differences between domains can be misleading
The paper used bootstrap confidence intervals. In plain terms, this resamples the collected answers to show a plausible range around each number. Many apparent gaps between domains fell inside that noise range.
Rankings were also unstable. The abstract says this applied across the frequently cited set of domains, not just the top few. A domain ranked fourth in one run could reasonably be seventh in the next.
Illustrative example: if your report shows you at 8% and a rival at 10% from one run of 50 prompts, the paper's logic says the real difference may be indistinguishable from zero. The numbers are invented, only to show the reasoning.
What this means for reporting
A useful citation report should state how many times each prompt was run, over what period, and on which platform. It should show a range, not only a point figure. The paper also offers guidance on sample sizes needed for interpretable intervals, but the abstract does not give the figures, so check the full paper before setting your own.
This is a measurement problem separate from page quality. A report can be precisely wrong about whether a content change helped. Single-benchmark correlations about what gets cited should be read with the same caution.
What remains uncertain
The study does not say why citations vary, or how much variation your topic would show. Behavior may differ by platform, topic and time, and products change. The findings describe the sampled period, not a permanent property of every answer engine.
Next steps
- Ask your reporting tool or agency how many repeated runs back each figure, and whether it gives confidence intervals.
- Run a small test: repeat the same 10–20 prompts on one platform several times across days. Record how much your citation share moves.
- Set a decision rule before changing content: act on a shift only if it exceeds the variation you observed.