Yes. Foglift's frozen 75-question benchmark found 77.0% to 86.7% mean same-question citation-set turnover across five AI engines from Q2 to Q3 2026. A source map captured in one quarter did not describe the next.
Methodology
Controlled layer: the same 75 brand-neutral buyer-intent prompts across 25 verticals and five engine lanes in Q2 and Q3 2026, for 375 responses per quarter. Domains were normalized and deduplicated within each response. For each prompt and engine, turnover is one minus the mean Jaccard similarity between the Q2 and Q3 cited-domain sets. Empty-to-empty pairs were excluded. ChatGPT had 73 comparable pairs; every other lane had 75. Provider models and search indexes changed, so this measures observed drift rather than a single causal mechanism.
Yes. Across identical buyer questions, mean same-question citation-set turnover was 77.0% for ChatGPT, 77.8% for Claude, 85.3% for Gemini, 86.7% for the Google AI Overview lane, and 80.3% for Perplexity. The prompt set was frozen, so a change in the questions cannot explain the difference. Provider models and search indexes did change, which is why this report describes observed drift rather than a controlled causal effect.
75
Identical buyer-intent questions per quarter
5
Engine lanes measured separately
77.0%–86.7%
Observed same-question turnover
Same-prompt turnover answers the narrow stability question: when the same buyer asks the same engine the same question in a later quarter, how much of the cited-domain set changes? Top-25 and top-100 turnover instead compare each engine's aggregate ranked source sets.
| Engine | Same-prompt turnover | Top-25 turnover | Top-100 turnover | Q2 top-10 concentration | Q3 top-10 concentration |
|---|---|---|---|---|---|
| ChatGPT | 77.0% | 78.0% | 69.3% | 11.0% | 11.3% |
| Claude | 77.8% | 88.9% | 81.0% | 9.0% | 12.2% |
| Gemini | 85.3% | 81.0% | 81.7% | 10.4% | 15.3% |
| Google AI Overview lane | 86.7% | 75.0% | 80.2% | 9.0% | 15.8% |
| Perplexity | 80.3% | 88.9% | 80.2% | 13.4% | 12.9% |
ChatGPT had 73 comparable prompt pairs. The other four lanes had 75 each. Empty-to-empty pairs were excluded from the same-prompt mean.
ChatGPT's 77.0% mean turnover was the lowest of the five lanes. Nine Q2 top-25 domains remained in its Q3 top 25, while top-10 concentration stayed nearly flat at 11.0% and 11.3%.
Claude moved from 332 distinct cited domains in Q2 to 273 in Q3. Its top-10 concentration rose from 9.0% to 12.2%. In the controlled panel, business-press coverage rose 6.7 percentage points while listicle-site coverage fell 10.7 points.
Gemini's same-question turnover reached 85.3%. YouTube coverage rose from 14.7% to 49.3%, and Reddit coverage rose from 10.7% to 24.0%. The Google AI Overview lane had the highest turnover at 86.7%; YouTube rose from 13.3% to 44.0%, while Reddit rose from 5.3% to 22.7%.
Implementation boundary: the benchmark's Google AI Overview lane uses Foglift’s grounded Google model path. It is not a captured consumer search-results panel.
Perplexity grew from 393 distinct cited domains in Q2 to 880 in Q3. Its top-10 concentration moved from 13.4% to 12.9%, while same-question turnover reached 80.3%.
The privacy-safe production window contains 37,518 responses from April 1 through August 18, including 28,345 responses with at least one surfaced citation. April citation capture was incomplete across several engines, so comparisons require at least 500 citation-bearing responses per engine-month. Every engine clears that gate for May and August.
| Engine | May top-100 domains retained in August | Retained share |
|---|---|---|
| ChatGPT | 16 of 100 | 16% |
| Claude | 31 of 100 | 31% |
| Gemini | 19 of 100 | 19% |
| Google AI Overview lane | 19 of 100 | 19% |
| Perplexity | 24 of 100 | 24% |
These figures are a directional cross-check only. Production prompt mix, customer mix, provider models, and search indexes all changed over the window. No workspace, customer, prompt, answer text, or full source URL is published. Observational riser and faller lists are intentionally excluded because changing customer and prompt mix can dominate individual domain movement.
Published results are not interchangeable. Interval, product surface, prompt set, counting rule, and stability unit all change the percentage. The studies below support recurring measurement while answering related, but different, questions.
Repeated executions and time comparisons across Google organic search and five generative search systems.
Comparison boundary: Establishes stability as an evaluation dimension, but does not use Foglift's quarterly same-question turnover metric.
Week-over-week portfolio comparison across thousands of prompts, five engines, and nine industries.
Comparison boundary: Its stable portfolio-share measure is not the same as source-set turnover for an identical question.
Weekly comparisons over 17 weeks and 82,619 prompts across consumer AI search surfaces.
Comparison boundary: Its weekly churn-in definition and product surfaces differ from Foglift's quarterly panel.
Adjacent repeated checks across more than 20,000 answers to fixed buyer questions.
Comparison boundary: Sequential checks measure a shorter interval than Foglift's Q2-to-Q3 comparison.
AI citation visibility is a moving, engine-specific measurement problem. A quarterly source audit can reveal structural shifts, but operating teams need recurring engine-level monitoring to see when an old source map stops describing the answers buyers receive. Foglift connects that monitoring to Technical Audits, recommendations, crawler evidence, referral tracking, and developer workflows through its API, CLI, and MCP server.
For the underlying quarterly panel, read the Q3 2026 AI Search Citation Benchmark.
Yes. Foglift's frozen benchmark found 77.0% to 86.7% mean same-question citation-set turnover across five AI engines from Q2 to Q3 2026. The comparison used the same 75 buyer-intent questions in both quarters.
For each prompt and engine, Foglift compared the Q2 and Q3 sets of cited domains with Jaccard similarity. Turnover is one minus the mean Jaccard score. A higher percentage means fewer of the same cited domains remained attached to the same questions.
The Google AI Overview lane had the highest mean same-question turnover at 86.7%, followed by Gemini at 85.3%. Foglift's Google lane uses a grounded Google model path, not a captured consumer search-results panel.
ChatGPT was the least volatile lane in the controlled comparison, but its 77.0% same-question turnover still meant that most of the observed source sets changed between quarters.
The controlled benchmark repeats the same 75 questions, so prompt drift cannot explain the result. The production cross-check is larger but directional because customer mix, prompt mix, provider models, and search indexes all changed from May to August.
Interval, product surface, prompt set, counting rule, and unit of analysis all affect a stability percentage. Portfolio-level citation share can remain stable while the sources attached to an identical question change substantially.
Foglift will extend the frozen panel quarterly. The same prompt set and engine-specific counting rule will be retained so future comparisons can compound without changing the measurement contract.
Yes. The CSV contains controlled engine-level turnover and privacy-safe monthly aggregate rows. It contains no workspace, customer, prompt, answer text, or full source URL.
Free tool
Run a Technical Audit on your own site
Check technical SEO, performance, security, accessibility, and structural AI Readiness. No signup required.
Run Free Technical Audit