AI Answer Volatility Statistics: 2026 Report
Our research team measured how much AI search answers change when the same question is asked again, back to back, day to day and week to week, across ChatGPT, Perplexity, Gemini, Claude, Google AI Mode and Google AI Overviews. The results show anyone tracking AI visibility how much of an apparent change is real movement and how much is noise, and how to sample prompts for a reliable read.
The answers moved. We ran 1,800 prompts across six industries on every engine daily for 13 weeks, from June 1 to August 30, 2026, with two extra back-to-back runs each Monday, for just under 1.2 million responses. Retention in this report means the share of brands, domains or URLs in one answer that appear again in a later one, and churn is the share that does not.
AI Answer Volatility by Unit and Time Window
In the table below, we show how much of an AI answer survives from one observation to the next, pooled across all six engines, for each unit we tracked and each measurement window. Brand and domain figures are averaged per prompt, so a prompt with a long list of sources carries the same weight as a prompt with only two.
| Unit Measured | Back-to-Back Runs | Day over Day | Week over Week | Over 13 Weeks |
|---|---|---|---|---|
| Brand mentions retained | 81% | 76% | 69% | 51% |
| Cited domains retained | 58% | 52% | 45% | 24% |
| Cited URLs retained | 41% | 33% | 26% | 10% |
| Prompts with an identical brand set | 6.8% | 4.1% | 2.6% | 0.7% |
| Prompts with an identical brand order | 1.9% | 1.1% | 0.6% | 0.1% |
Two gradients run through the table. Volatility rises as the unit becomes more specific: brands are the most durable element of an answer, domains sit in the middle, and exact URLs are the least stable, with only 10% of cited pages still present after 13 weeks. Volatility also rises with the length of the window, though less steeply than many readers expect. The largest single drop for every unit occurs between the first run and the second, before any time has passed at all. Roughly three quarters of the domain turnover we observed over a full week was already present between runs submitted a few seconds apart.
The whole-answer rows show why ranking-style reporting breaks down. Even when 81% of individual brands carry over between back-to-back runs, the full set matches across all three Monday runs only 6.8% of the time and the order matches 1.9% of the time, because a single substitution anywhere in a list of six to ten brands is enough to break the match. SparkToro and Gumshoe reached a similar conclusion from a much smaller sample, finding the same brand list fewer than 1 time in 100 across 2,961 runs. Mention frequency is the more stable quantity, and it is the one we recommend tracking. ChatGPT gave an identical set of recommended brands in fewer than 5 of every 100 prompts asked three times in a row, and about two thirds of the websites it cited changed from one week to the next.
Week-over-Week Cited Domain Churn by Engine, June to August 2026
The chart and table below track the share of cited domains that changed from one week to the next for four engines, sampled every other week across the study period. Google AI Overviews and Claude are shown in the engine table that follows.
| Week Of | ChatGPT | Google AI Mode | Perplexity | Gemini |
|---|---|---|---|---|
| Jun 8 | 63% | 52% | 36% | 66% |
| Jun 22 | 66% | 55% | 34% | 69% |
| Jul 6 | 61% | 50% | 37% | 65% |
| Jul 20 | 68% | 54% | 33% | 71% |
| Aug 3 | 64% | 57% | 35% | 68% |
| Aug 17 | 63% | 51% | 37% | 71% |
- We found no meaningful trend in any engine over the 13 weeks. Each series moved within a band of 4 to 7 points, and the ordering of the four engines never changed, which suggests that churn is a stable property of each engine's retrieval design. Holiday weeks, including the week of July 4, showed no visible disruption.
- Our data showed Perplexity churning at roughly half the rate of Gemini in every sampled week, with averages of 35.3% and 68.3%. The gap held in all six industries, which points to engine architecture as the main driver.
- We observed lower churn than SISTRIX reported for ChatGPT (74%) and Google AI Mode (56%) in its 17-week study. Our prompt set leans toward commercial and comparison queries with established source pools, which likely explains part of the difference, and the direction of the engine gap matches theirs.
AI Answer Volatility by Engine
In the table below, we compare the six engines on four measures, averaged across the full study period and all industries. Engines are ordered from most to least stable by day-over-day domain retention.
| Engine | Day-over-Day Domain Retention | Week-over-Week Domain Churn | 13-Week Domain Retention | Identical Brand Set Across 3 Runs |
|---|---|---|---|---|
| Perplexity | 69% | 35.3% | 38% | 11.2% |
| Google AI Overviews | 63% | 45.0% | 34% | 9.6% |
| Google AI Mode | 50% | 53.2% | 24% | 6.1% |
| Claude | 50% | 64.0% | 17% | 5.6% |
| ChatGPT | 44% | 64.2% | 17% | 4.9% |
| Gemini | 36% | 68.3% | 14% | 3.4% |
Perplexity and Google AI Overviews form a clear stable tier. Both lean heavily on a consistent pool of high-authority pages, and both returned an identical brand set across three back-to-back runs roughly twice as often as ChatGPT. Google AI Mode is noticeably less stable than AI Overviews despite drawing on the same index, a gap that likely reflects AI Mode's longer, multi-step answers, which pull in more sources per response and give each individual source more competition. Profound's month-over-month data also ranked Perplexity as the most stable engine it tested, with 40.5% domain drift.
Claude is the engine with the least published volatility research, so its numbers deserve a closer read. Its day-over-day retention matched Google AI Mode at 50%, yet its weekly churn was close to ChatGPT's. In practice, Claude tended to reuse the same sources for a few days and then replace a large share of them at once, a stepwise pattern that averages out to similar figures but produces sharper visible swings for anyone checking results weekly. Gemini was the least stable engine on every measure, keeping just 14% of its cited domains across the full 13 weeks.
AI Answer Volatility by Industry
The next table breaks the results down by industry. Each industry accounts for 300 prompts, and the All Industries row is the average across all six engines. Industries are ordered from most to least stable.
| Industry | Day-over-Day Domain Retention | Week-over-Week Domain Churn | 13-Week URL Retention |
|---|---|---|---|
| Insurance | 59% | 48% | 13.1% |
| Financial Services | 57% | 50% | 12.4% |
| Travel | 51% | 55% | 10.2% |
| Consumer Electronics | 50% | 57% | 9.0% |
| Healthcare | 49% | 58% | 8.2% |
| B2B Software | 46% | 62% | 7.1% |
| All Industries | 52% | 55% | 10.0% |
- We found that insurance and financial services were the most stable industries. Their prompts drew on a relatively small set of government sites, established publishers and large carriers or banks, which kept both domain and URL retention above the average.
- Our data showed B2B software to be the most volatile industry, with 62% weekly domain churn and just 7.1% of cited URLs surviving 13 weeks. The category produces a steady stream of new comparison pages, review roundups and vendor blog posts, and engines rotate among them freely.
- We observed that healthcare combined high churn with concentrated brand visibility. Its weekly domain churn of 58% was second highest, yet the leading hospital systems and health publishers appeared in most answers to their prompts, which fits a pattern where the brands stay put while the specific pages cited about them turn over.
AI Answer Volatility by Prompt Type
In the table below, we group the 1,800 prompts by intent and show each group's share of the prompt set along with its day-over-day domain retention and brand set consistency.
| Prompt Type | Share of Prompt Set | Day-over-Day Domain Retention | Identical Brand Set Across 3 Runs |
|---|---|---|---|
| Definitional prompts | 20% | 66% | 10.0% |
| How-to prompts | 18% | 60% | 8.8% |
| Comparison prompts | 16% | 49% | 9.4% |
| Pricing prompts | 14% | 47% | 5.6% |
| Current events prompts | 10% | 44% | 4.9% |
| Best-of prompts | 22% | 42% | 2.1% |
Prompt intent turned out to be a stronger predictor of volatility than industry. Definitional and how-to prompts retained the most sources, since the pool of authoritative explanations for a concept changes slowly. Best-of prompts, such as requests for the best project management software for a small team, retained only 42% of their cited domains from one day to the next and returned an identical brand set just 2.1% of the time. These are also the prompts brands care most about, which means the most commercially important answers are the least stable ones to measure.
Comparison prompts show an interesting split. Their domain retention was below average at 49%, but their brand set consistency was the second highest in the study at 9.4%, because a prompt that names two products fixes most of the brands in the answer in advance. For measurement design, the implication is to weight sampling toward open-ended best-of and pricing prompts. In our data, a single daily run of a best-of prompt disagreed with that prompt's weekly majority result on whether a given brand appeared about one time in four, while pooling seven daily runs brought that disagreement below one in ten. For teams building a tracking program, the results point to a simple structure. Track brand mention rates in preference to rank positions, sample a broad set of prompts in preference to many runs of a few, and report results as rolling averages over at least two weeks. Mark any known model release in the reporting window, since a sudden shift in citations often traces back to the engine itself. SISTRIX's analysis of the GPT-5.5 update, reported by Search Engine Journal, found day-to-day variation in ChatGPT's citation patterns jumping from its typical 1% to 2% to 47% during the change, enough to swamp months of ordinary movement.
Requesting a Copy of This Report
If you would like a PDF copy of this report, or the full prompt set and response logs behind it, you can reach out here.
Sources
- AI Answer Volatility Study, Independent Research Team, September 2026, New York, New York.
- NEW Research: AIs Are Highly Inconsistent When Recommending Brands or Products. Rand Fishkin, SparkToro, January 2026, Seattle, Washington.
- GPT-5.5 Update Changes How ChatGPT Cites Sources. Matt G. Southern, Search Engine Journal, June 2026.
- AI Citation Drift: How Stable Are Sources in AI Search Results?. Johannes Beus, SISTRIX, May 2026, Bonn, Germany.
- AI Search Volatility: Why AI Search Results Keep Changing. Josh Blyskal and Sartaj Rajpal, Profound, 2025, New York, New York.
- AI Overviews Change Every 2 Days (But Never Change Their Mind). Louise Linehan and Xibeijia Guan, Ahrefs, November 2025, Singapore.
- AI Mode Research: Sources, Volatility, and Differences Between AIO and Organic Search. Yevheniia Khromova, SE Ranking, June 2025.
- AI Citation Volatility by Industry. Dimitry Apollonsky, Parse, 2026, New York, New York.
- AI Citation Volatility: A 530,875-Citation Study. Anirudh Agarwal, GetMentions AI, July 2026, Dover, Delaware.
- AI Visibility Sample Size: How Many Prompts and Runs to Trust. Chris Han, MaxAEO, July 2026.