# How Often Do AI Answers Change?

> Ask an AI assistant the same buyer question twice and the brand list often differs. What studies say about run-to-run change and which metrics hold steady.

Page: https://trylectern.com/blog/how-often-do-ai-answers-change

By Edgar Li, October 30, 2026.

AI answers change constantly, and how much depends on what you count. The exact list of brands in an answer almost never repeats. The brands that appear most often across many runs are far steadier. Sources behave the other way: a page cited today is often not cited tomorrow. Published volatility numbers disagree because they measure different things, so before you read any of them, check the unit.

## Key Takeaways

- Two runs of the same prompt return the identical brand list under 1 time in 100 ([SparkToro and Gumshoe](https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/), January 2026).
- Brand mentions are steadier than citations. In [one daily panel](https://otterly.ai/blog/ai-search-visibility-stability/), 12 to 19 percent of brands appeared on only one day in 30, against 26 to 67 percent of sources.
- Stability depends on the engine. Google AI Overviews were far steadier week to week than ChatGPT in [one 2026 study](https://www.sistrix.com/blog/ai-citation-drift-how-stable-are-sources-in-ai-search-results/).
- The intent of the question matters. The first-named brand matched across 91 percent of runs on comparison prompts and 30 percent on education prompts in [one test](https://www.conductor.com/academy/ai-recommendation-consistency-analysis/).
- Report rates over many runs, with the run count, and never a single answer.

## Why Answers Move

Several things change between two runs of an identical prompt.

- **The model samples.** Output is generated with randomness, so wording and the set of brands named differ.
- **Retrieval varies.** When an assistant searches the web, it may run different sub-queries each time. In [one Profound study](https://www.tryprofound.com/blog/what-ai-engines-actually-search-for) from April 2026, 91 percent of ChatGPT's search strings for a prompt were unique across runs, against 14 percent for Perplexity.
- **The index changes.** New pages appear, old pages drop out, and ranking shifts.
- **The platform changes.** Providers update models and search behaviour, sometimes overnight. [Our guide to timing](https://trylectern.com/blog/how-long-does-aeo-take) lists examples.

A preprint by [Sielinski (arXiv 2603.08924, 2026)](https://arxiv.org/abs/2603.08924) checked whether page changes explained the movement. It fetched and hashed the text of every cited page at each collection. Citation shares crossed ranks while the pages were unchanged, and two samples taken ten minutes apart showed the same swaps. The variation comes from the engine.

## What the Studies Measured

Published figures are not interchangeable. Each row below counts something different.

| Source | What it counts | Result |
|---|---|---|
| [SparkToro and Gumshoe](https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/), Jan 2026 | Identical brand list across two runs | Under 1 in 100 |
| [SISTRIX](https://www.sistrix.com/blog/ai-citation-drift-how-stable-are-sources-in-ai-search-results/), May 2026 | Cited domains replaced from one week to the next | About 5% in AI Overviews, 56% in AI Mode, 74% in ChatGPT (Germany) |
| [Otterly](https://otterly.ai/blog/ai-search-visibility-stability/), Sep 2026 | Sources kept from one day to the next, per prompt | 22 to 36% (AI Overviews), 65 to 75% (Perplexity) |
| [BrightEdge](https://www.brightedge.com/resources/weekly-ai-search-insights/ai-search-citations-week-to-week-changes), Feb 2026 | Cited domains with no change in share over a week | 96.8% unchanged, five engines pooled |

The BrightEdge and SISTRIX figures look contradictory and are not. One counts domains whose share did not move, the other counts domains that were replaced, and one is pooled across engines while the other is split. The AI Overviews average in [the SISTRIX study](https://www.sistrix.com/blog/ai-citation-drift-how-stable-are-sources-in-ai-search-results/) also hides a split: 53 percent of prompts never changed a source, while 19 percent churned at 46 percent a week.

## Which Metrics Hold Steady

**Brand presence over many runs.** The share of answers that name you, across a fixed set of questions, is reportable if you state how many answers it rests on. [Otterly found](https://otterly.ai/blog/ai-search-visibility-stability/) that coverage readings on 10 prompts fell anywhere from 39 to 71 percent, on 50 prompts from 49 to 62 percent, and on 100 prompts from 51 to 60 percent. More questions narrow the range faster than more repeats of the same question.

**The leading brand on comparison prompts.** [Conductor (June 2026)](https://www.conductor.com/academy/ai-recommendation-consistency-analysis/) found the first-named brand was the same in 91 percent of 50 runs for comparison prompts, 59 to 72 percent for purchase and recommendation prompts, and 30 percent for education prompts. Position within a single answer is noise, but a leader across many runs can be stable.

**Citations of one page for one question.** Noisy. One run finds only about a tenth of the URLs an engine cites for a query ([Evertune](https://www.evertune.ai/resources/insights-on-ai/the-geo-opportunities-that-small-prompt-samples-miss)), and in one test just 2.3 percent of page and query pairs were cited in all three of three runs ([AirOps](https://www.airops.com/report/the-fan-out-effect-what-happens-between-a-query-and-a-citation)).

## Prompts and Personas Move Results Too

Wording matters. [One vendor report](https://www.unusual.ai/blog/how-stable-is-share-of-voice) found that changing a single synonym changed the vendor set in 84 of 100 prompt pairs, though it had no same-wording baseline. [Profound found](https://www.tryprofound.com/blog/do-answer-engines-customize-responses-to-different-personas) that different stated personas shared about 25 percent of brands, against 40 percent for the same persona. A reworded question is a new question, so keep your panel fixed if you want to see change over time.

## How to Read Your Own Results

1. **Fix the questions and the assistants.** If either changes between runs, the change may be the panel.
2. **State the number of answers.** A rate with no count is not a measurement. [Tracking brand mentions in ChatGPT](https://trylectern.com/blog/track-brand-mentions-chatgpt) shows the manual method.
3. **Split by assistant.** A flat total can hide a rise in one and a fall in another.
4. **Compare against a null.** With nothing changed, 57 percent of a top five turned over between two adjacent three-day windows in [one Searchable analysis](https://www.searchable.com/blog/chatgpt-research-depth-study). A rank change smaller than that is not news.
5. **Mark platform change dates.** Do not read a move that starts on one as the effect of your work.

## What This Means in Practice

Do not react to a single answer, in either direction. If an assistant names you once, you have an example. If it omits you once, you have an example. The decision-grade view is a rate across enough questions and runs, tracked over weeks, split by assistant.

Lectern runs a fixed set of buyer questions on a schedule and keeps the history, so movement is read against the panel and not against one lucky answer. The free Hobby plan runs 10 questions weekly, and Lead, at $990 a month per brand, refreshes daily. Details are on the [pricing page](https://trylectern.com/pricing). For how to turn the readings into a comparison with competitors, see [AI share of voice](https://trylectern.com/blog/ai-share-of-voice).