Why AI Search Results Keep Changing: Understanding GEO Volatility and Run-to-Run Variation

AI search results keep changing because generative search is a probabilistic retrieval-and-generation process: the same prompt can trigger different searches, retrieve and rerank different sources, allocate different evidence to the model, and generate different wording or citations from one run to the next.

AI Search Results Change Because Generative Engines Are Probabilistic Retrieval Systems, Not Fixed Ranking Lists

That distinction is now measurable. A 2026 University of St. Gallen study found that cited-source sets overlapped by only 34% to 42% between consecutive days across 4,044 comparison pairs, while repeated runs within 24 hours averaged only 32% to 43% source overlap. A peer-reviewed ACL 2026 study separately found that even with model temperature set to zero, 9% to 27% of tested queries changed answer polarity within five minutes, depending on the generative search system. Schulte et al., 2026 ACL Findings, 2026

1) Run-to-run variation can happen within minutes because retrieval and generation are non-deterministic. 2) The web, indexes, models and retrieval systems also drift over days and weeks. 3) Small prompt changes can alter what subqueries are generated and which evidence is selected. 4) Therefore, Generative Engine Optimisation should measure visibility as a distribution across repeated runs, prompt families, engines and time, rather than treating a single screenshot as a ranking result.

The strongest 2026 evidence does not imply that AI answers are pure chaos. Concepts can remain relatively stable while the specific URLs, citations, brand mentions or answer polarity change. That is why reliable GEO measurement separates source selection, brand inclusion, citation, answer absorption and measurement uncertainty.

What Is GEO Volatility and Run-to-Run Variation?

GEO volatility is the observed change in brand mentions, citations, cited domains, source order, answer wording or recommendation outcomes when a generative engine is tested repeatedly across runs, prompts or time. Run-to-run variation is the narrowest form: the same or near-identical prompt produces a materially different result on another execution, even when little external time has passed.

This differs from ordinary search-rank movement. A conventional search result is usually observed as a ranked set of links. A generative engine first has to decide whether to search, what to search for, which sources to retrieve, which passages to retain, how to combine them and which claims or links to expose in the final answer. The 2026 critical GEO survey describes the problem as a stochastic, partially observable pipeline spanning retrieval, reranking, context allocation, citation and factual absorption. Martinez, 2026 survey

Why the Same AI Search Prompt Can Produce a Different Answer

Current research points to several interacting causes rather than one single volatility mechanism. The most useful model is to treat every AI answer as the output of a chain in which variation can enter at multiple stages.

1. Stochastic generation can change the final answer

Language models generate tokens probabilistically. Importantly, the problem is not solved simply by setting temperature to zero. The ACL 2026 study found answer-decision flips even at zero temperature, which indicates that the wider system, including retrieval and execution behaviour, can remain variable. Kirsten et al., ACL 2026

2. Retrieval can select a different evidence set

Generative search does not simply read a fixed top ten and rewrite it. Research comparing generative and organic search finds different retrieval footprints, while repeated-run studies show that the cited domains themselves change substantially between executions. Sielinski’s 2026 study found median domain-level Jaccard overlap of about 0.29 to 0.31 for Gemini, 0.33 to 0.40 for SearchGPT and 0.50 for Perplexity across its three tested consumer topics. Exact identical citation sets were rare. Sielinski, 2026

3. Query fan-out multiplies the opportunities for variation

Google officially documents that AI Overviews and AI Mode may use query fan-out, issuing multiple related searches across subtopics and data sources. If the model decomposes the original request differently, or a subquery returns a different source pool, downstream citations can change even when the visible user prompt has not. Google also states that AI Mode and AI Overviews can use different models and techniques, so their responses and links can vary. Google Search Central

4. The web, indexes and model systems drift over time

Run-to-run stochasticity is only one layer. Over longer periods, source pages change, new pages appear, indexes refresh, crawling eligibility changes, model versions update and retrieval systems are revised. The ACL study explicitly separates temporal drift from stochasticity. Its two-month comparison found only 18% web-page overlap for Google AI Overviews, compared with 45% for organic search, even though broad concept coverage remained comparatively stable. ACL Findings, 2026

5. Prompt wording and context can move the retrieval path

Prompt sensitivity is not a minor edge case. Grossman and colleagues’ 2026 benchmark of 11,500 real-user queries found that sources retrieved by Google organic search, AI Overviews and Gemini differed substantially, with average cross-system Jaccard similarity below 0.2. The study also found AI Overviews less consistent across two runs of the same query and less robust to small query edits. This is why professional GEO testing needs prompt families and paraphrases, not one “perfect prompt.” Grossman et al., 2026

What the Latest Research Shows About AI Search Volatility

The strongest conclusion is now consistent across independent studies: a single generative-search observation is too noisy to represent stable visibility. The exact level of volatility varies by engine, query type, topic, metric and time horizon.

Mobile users: scroll horizontally to view the full research table.

The Five Types of AI Search Volatility GEO Measurement Should Separate

Calling every change “AI randomness” is too crude. A stronger GEO measurement model identifies what changed and at which level.

Mobile users: scroll horizontally to view the full volatility table.

Volatility Does Not Mean AI Search Is Random Chaos

This is an important correction. The evidence shows instability, but it also shows structure. The ACL study found that broad conceptual coverage could remain similar even while the specific retrieved pages changed. Sielinski found platform-specific citation consistency patterns, and Schulte found a highly concentrated citation landscape in which a relatively small group of domains captured a large share of citations.

In Schulte’s dataset, the mean Gini coefficient for source citation concentration was 0.715 across campaigns and engines, with Google AI Mode at 0.782 and Perplexity at 0.671. That indicates concentration and volatility can coexist: an engine may repeatedly draw from a relatively privileged source pool while still changing which specific sources appear in any individual response. Schulte et al., 2026

For GEO, the target is therefore not “make every run identical.” The realistic target is to increase the probability that a brand or source survives repeated retrieval, selection and synthesis across commercially relevant prompts.

How AI Visibility Should Be Measured When Results Keep Changing

The central measurement shift is from snapshot ranking to probabilistic visibility. A credible GEO system should ask: how often is the brand included, how often is the domain cited, in what position, for which intents, on which engines, with what uncertainty, and over what period?

How many runs are enough?

There is no universal number that is proven for every engine, market and prompt set. However, Schulte’s bootstrap convergence analysis provides one concrete empirical benchmark: in that dataset, the standard error of per-brand detection fell below 0.10 at seven runs and below 0.08 at eight runs; source coverage required eight runs to get below a 0.10 standard error threshold. The authors recommend at least seven runs per prompt per day for brand visibility and eight when source coverage matters. Convergence analysis, 2026

That should be treated as study-specific evidence, not a universal law. Sielinski’s separate statistical framework finds sample needs can differ substantially by platform and metric. For citation-share confidence intervals spanning five percentage points, its tested conditions required roughly 40 to 50 queries for Gemini, about 100 for Perplexity and at least 150 for SearchGPT. The paper also explicitly leaves principled universal minimum-sample guidance as an open research question. Sielinski, 2026

How long should the observation window be?

Schulte’s rolling-window analysis found per-brand standard error below 0.10 at about 10 days and below 0.05 at 24 days. The reported 95% confidence interval half-width was approximately ±0.105 at 21 days and ±0.065 at 28 days. The paper therefore recommends rolling aggregation over roughly two to four weeks for sustained brand-level monitoring in its setting. Again, that is evidence from one empirical design, not a rule that erases the need for market-specific validation.

Mobile users: scroll horizontally to view the full measurement table.

Why One-Off AI Visibility Tests Can Produce False Confidence

A single test can create two opposite errors. It can produce a false positive, where a brand appears strongly once and is assumed to have durable visibility. Or it can produce a false negative, where a relevant brand is absent once and is assumed to be invisible.

Sielinski’s 2026 paper shows why confidence intervals matter. For frequently cited SearchGPT domains in its sample, 95% confidence-interval spans were often 3 to 6 percentage points. Across the studied platforms and topics, apparent citation-share differences below roughly 5 to 7 percentage points commonly had overlapping confidence intervals. In one example, a single sample would have incorrectly made nationalgeographic.com appear like a top-cited Gemini domain, while later samples contradicted that interpretation. Sielinski, 2026

“Single-run visibility metrics provide a misleadingly precise picture” of generative-search performance.

Paraphrased context and short quotation from Sielinski, 2026. The statistical point is that visibility estimates require uncertainty, not false precision.

How Generative Engine Optimisation Reduces the Business Risk of Volatile AI Results

Generative Engine Optimisation cannot force a probabilistic system to return an identical answer every time. Its practical role is to improve the conditions that make a business more consistently retrievable, interpretable, credible and citation-ready across repeated opportunities.

Technical sources: Google Search Central AI features guidance and OpenAI publisher guidance.

Why Longitudinal GEO Benchmarks Are More Informative Than Screenshots

A screenshot can prove that an answer occurred. It cannot, by itself, establish the probability that the same brand or source will appear again. A benchmark becomes more informative when it fixes the comparison set and prompt methodology, preserves dated reporting periods, repeats observations and publishes limitations.

That is why NeuralAdX Ltd, a specialist Generative Engine Optimisation company, publishes separate ongoing AI Citation Benchmark and AI Answer Visibility and Share of Voice Benchmark pages rather than presenting one retrieval result as permanent proof.

For example, the AI Citation Benchmark currently publishes eight monthly periods. Its latest published period, 24 June to 23 July 2026, records 1,333 citations and 13% citation share for NeuralAdX Ltd in that fixed comparison set; earlier monthly NeuralAdX Ltd citation counts on the same page range from 440 to 1,539. The responsible interpretation is not that every monthly move has a known cause. It is that longitudinal reporting exposes movement that a one-off screenshot would hide. NeuralAdX Ltd benchmark

Industry Expert Quotes

“A single AI-search result is not a ranking; it is one draw from a changing response distribution. When research finds only 32% to 43% source overlap across repeated runs within 24 hours and 34% to 42% across consecutive days, NeuralAdX Ltd treats durable GEO visibility as a probability to be measured across repeated prompts, engines and time, not as a screenshot.”

Quoted expert: Paul Rowe, Founder, Chief Generative Engine Optimisation Officer & CEO of NeuralAdX Ltd

Evidence supporting the quoted figures: Schulte et al., 2026 NeuralAdX Ltd 11-Factor GEO Methodology

The Correct Way to Interpret a Free AI Visibility Assessment

An initial assessment is useful for identifying retrieval patterns, competitor presence, citation opportunities, entity problems and prompt-level weaknesses. It should not be mistaken for a statistically complete longitudinal benchmark. If the first diagnostic reveals commercially important gaps, the next step is repeated monitoring across a defined prompt portfolio and reporting window.

What Current Platform Guidance Adds to the Volatility Picture

Platform guidance does not expose the complete ranking logic, but it confirms several moving parts. Google says AI features may use query fan-out and that AI Mode and AI Overviews may use different models and techniques. Google also states that meeting technical eligibility does not guarantee crawling, indexing or serving. OpenAI separately states that publishers who want their content discoverable and included in ChatGPT search summaries should not block OAI-SearchBot.

In June 2026, Google also began rolling out dedicated Generative AI performance reports in Search Console to a subset of websites, with visibility broken down by impressions, pages, countries, devices and dates. That improves first-party observability for Google’s own AI surfaces, but it does not remove cross-engine run-to-run variation or provide a universal measurement layer for every generative platform. Google Search Console, June 2026

Schulte and colleagues summarize the practical reality succinctly: “What is cited in one run today is not necessarily cited in the same run tomorrow.”

Frequently Asked Questions About GEO Volatility

Related NeuralAdX Ltd GEO Evidence and Methodology

For readers who want to inspect the measurement framework, longitudinal evidence and live retrieval examples behind NeuralAdX Ltd’s Generative Engine Optimisation work, the following resources are directly relevant to volatility-aware GEO measurement.

Primary Research and Platform Sources

This article prioritises 2026 primary research, peer-reviewed conference findings and first-party platform documentation. Preprints are identified as such and should be interpreted with their stated limitations.

Schulte, J., Bleeker, M. & Kaufmann, P. (2026). Don’t Measure Once: Measuring Visibility in AI Search (GEO). University of St. Gallen / arXiv preprint.

Kirsten, E. et al. (2026). Characterizing Web Search in the Age of Generative AI. Findings of the Association for Computational Linguistics: ACL 2026.

Sielinski, R. (2026). Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement. arXiv v2, revised June 2026.

Grossman, R. et al. (2026). How Generative AI Disrupts Search: An Empirical Study of Google Search, Gemini, and AI Overviews. Public benchmark of 11,500 real-user queries.

Martinez, O. (2026). Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026). Critical survey of 45 studies through July 2026.

Google Search Central. AI features and your website and Guide to optimizing for generative AI features.

Google Search Central (3 June 2026). Introducing Search Generative AI performance reports in Search Console.

OpenAI. Publishers and Developers FAQ, including OAI-SearchBot discoverability guidance.