# How Often Do AI Citations Change? What the Studies Show

Six published studies show how fast ChatGPT, Google AI Mode and AI Overviews change the sources they cite, why their numbers disagree, and how to measure turnover yourself.

A lost AI citation is usually normal churn, not a signal. Here is how fast the sources behind AI answers turn over, engine by engine, and how to measure it on your own prompts.

- Most AI citations change within days; Ahrefs saw about 60% of ChatGPT citations gone next day.
- Weekly, ChatGPT Search rotates about 74% of cited domains and Google AI Mode about 56%.
- Google AI Overviews are steadiest, with 53% of prompts unchanged over 17 weeks in SISTRIX data.
- A small core of domains and a brand's own site persist while the rest rotates.
- Judge AI visibility by presence across many runs and weeks, never one snapshot or position.

These figures cover ChatGPT, Google AI Mode, Google AI Overviews, Gemini and Copilot as the published studies measured them, and they differ by unit (URL, domain or brand) and by interval. The studies use different samples, engines and metrics, so their figures show the range of turnover, not one universal rate.

## How Often Do AI Citations Change?

Most AI citations change within days. [Ahrefs found that about 60% of the sites ChatGPT and Google AI Mode cited were gone the next day](https://ahrefs.com/blog/how-stable-are-ai-citations/), and over a full week SISTRIX measured 56% of cited domains rotating in Google AI Mode and 74% in ChatGPT Search.

Turnover at that pace is the normal state of [AI search visibility](https://growthtribune.com/search-ai-visibility/ai-search-visibility/), not a sign that something broke. Each turnover figure below is credited to the study that published it, and the studies are compared side by side rather than averaged into one rate.

The answer depends on the time scale:

- **Next day:** Ahrefs saw 50% of day-one citations gone on Google AI Overviews, about 60% on AI Mode and ChatGPT, 81% on Gemini and 83% on Copilot. With empty answers left out, the loss was 50 to 60% on all five tools.
- **Within three weeks:** About 75% of the citations that dropped out came back within Ahrefs' 20 days, and about half came back within three days. Between 20% and 27% of day-one citations disappeared and never returned, depending on the tool.
- **Week to week:** SISTRIX measured 56% of cited domains rotating each week in Google AI Mode and 74% in ChatGPT Search, while 53% of Google AI Overviews prompts kept the same sources for all 17 weeks.
- **Month to month:** Alex Birkett compared the top 500 cited URLs from about 100 prompts one month apart and found 49% in both snapshots.

**74%** (17 December 2025 to 8 April 2026, 82,619 prompts in six countries): Share of cited domains that ChatGPT Search rotated each week in the SISTRIX panel, against 56% in Google AI Mode. Source: [AI Citation drift: How stable are sources in AI search results?](https://www.sistrix.com/blog/ai-citation-drift/).

So how long does an AI citation last? For most cited pages, days rather than months. Many come back, but roughly a fifth to a quarter never return within three weeks.

## What AI Visibility Studies Measured

Six published studies track AI citation turnover, and they measure different things. SISTRIX compared 1,548,213 weekly snapshots of 82,619 prompts across six countries over 17 weeks at domain level, while Ahrefs followed individual cited websites daily for 20 days.

Each study below publishes its method and describes its sample. That is the bar for inclusion.

**Six published AI citation turnover studies compared**

| Study | Sample | Interval | Unit measured | Who ran it |
| --- | --- | --- | --- | --- |
| Ahrefs | 7,831 questions, five engines | Daily for 20 days | Cited websites | SEO vendor, own data |
| SISTRIX | 82,619 prompts, six countries | Weekly for 17 weeks | Domains per prompt | SEO tool vendor |
| Profound | About 80,000 prompts per engine | One month apart | Newly cited domains | AI visibility vendor |
| BrightEdge | Thousands of prompts, five engines | One week apart | Domain citation share | Vendor, own data |
| SparkToro and Gumshoe | Nearly 3,000 runs | Identical prompts repeated | Brand lists | Independent researchers |
| Alex Birkett | About 100 prompts, one client | One month apart | Top 500 cited URLs | Practitioner analysis |

The engines and dates differ as much as the units:

- **Ahrefs** asked the same 7,831 questions every day in Google AI Mode, ChatGPT, Google AI Overviews, Gemini and Copilot, following about 37,500 websites cited on day one, using its own Brand Radar data.
- **SISTRIX** took 17 weekly snapshots from 17 December 2025 to 8 April 2026 of Google AI Mode, Google AI Overviews and ChatGPT Search in Germany, the USA, the UK, Italy, Spain and France. It measured churn-in (the share of new domains each week) and retention (the share of domains that survive).
- **Profound**, a vendor that sells AI visibility tracking, published its own analysis comparing 11 to 13 June 2025 with 11 to 13 July 2025 across four engines.
- **BrightEdge** compared the week of 1 February 2026 with the week before, across five engines and nine industries, from its own AI Catalyst data.
- **SparkToro and Gumshoe**, as reported by Search Engine Land, had 600 volunteers run 12 identical prompts through ChatGPT, Claude and Google's AI.
- **Alex Birkett** compared two snapshots of one client's cited URLs one month apart.

Because the units, engines and dates differ, these figures can be compared but never averaged. A domain-share number from one week and a URL-set number from one month answer different questions.

## Which AI Engines Change Citations The Most?

Of the three engines SISTRIX tracked, Google AI Overviews changed least week to week and ChatGPT Search most. SISTRIX found no source change over 17 weeks for 53% of AI Overviews prompts, against 74% weekly domain rotation in ChatGPT Search. Gemini and Copilot lose even more citations from one day to the next, but they often answer with no citations at all.

**AI citation turnover by engine in the Ahrefs and SISTRIX studies**

| Engine | Next-day loss (Ahrefs) | Longer-run finding, metric named |
| --- | --- | --- |
| Google AI Overviews | 50% | 53% of prompts never changed |
| Google AI Mode | About 60% | 56% of domains rotate weekly |
| ChatGPT | About 60% | 74% of domains rotate weekly |
| Gemini | 81% | 64% of answers uncited (Ahrefs) |
| Copilot | 83% | 53% of answers uncited (Ahrefs) |

The engines also differ in how many sources they cite, which shapes how a change feels:

- **Google AI Overviews** cite about 11 domains per answer in SISTRIX data, with about 8 of them consistently present.
- **Google AI Mode** cites 14 to 16 domains per answer and rotates over half of them each week.
- **ChatGPT Search** cites about 3 to 4 domains per answer. For the median ChatGPT prompt, no domain was present in all 17 weeks.
- **Gemini and Copilot** left 64% and 53% of answers without citations in the Ahrefs study, while ChatGPT and Google's tools almost always cited sources.

So does ChatGPT change its sources? Yes, more than Google's engines from week to week. With about 3 to 4 domains per answer, one swapped domain is a quarter to a third of the answer's sources.

One study points the other way. Profound's own analysis reported domain-level drift of 59.3% on Google AI Overviews, 54.1% on ChatGPT, 53.4% on Microsoft Copilot and 40.5% on Perplexity between June and July 2025. Profound's metric counts only domains newly appearing a month later, not those that dropped out, so it measures something different from SISTRIX's weekly rotation.

## Why Do AI Citation Studies Disagree?

AI citation studies disagree mainly because they count different units. [BrightEdge reported that 96.8% of cited domains had zero week-over-week change in citation share](https://brightedge.com/resources/weekly-ai-search-insights/ai-search-citations-week-to-week-changes), a measure of share across many prompts, while SISTRIX counted which domains left each prompt's citation set and found over half rotating weekly.

Both numbers can be true at once. Seven differences explain most of the gap:

- **Share versus set membership:** A domain can keep the same overall citation share while appearing on different prompts each week. BrightEdge measured share; SISTRIX measured which domains stayed on each prompt.
- **Domain versus URL:** SISTRIX calls its domain-level rates conservative and puts URL-level drift about 15% higher. In Alex Birkett's data, 49% of URLs survived a month while 65% of publishers (133 of 204) did.
- **New domains only:** Profound counts only domains newly appearing a month later. Its January to July comparison raised drift to 70 to 90%, and the underlying data for that figure is not published.
- **Empty answers:** Ahrefs' next-day losses of 81% on Gemini and 83% on Copilot fall to the 50 to 60% range once answers with no citations are left out.
- **Brand-name variants:** In a manual check of 55 brands, Ahrefs found that 30 to 36% of apparent losses were the same brand under a different name.
- **Time span:** BrightEdge reports one week-over-week comparison, aggregated across engines. SISTRIX ran 17 weeks and saw no sign of drift levelling off.
- **Who measured:** BrightEdge, Ahrefs and Profound each measured with their own tools and data, and BrightEdge describes its sample as "thousands of prompts" without an exact count.

Industry matters too. BrightEdge found finance had the most churn, with 51.1% of domains changing, and government and institutional sites the least, at 3.6%.

> **Warning:** Never read a single headline percentage on its own. Compare a turnover figure only with one measured the same way, on the same unit, interval and engine.

Citation turnover is only one reason [AI visibility scores differ between platforms](https://growthtribune.com/why-ai-visibility-scores-differ-between-platforms/), and the other causes of those score gaps sit outside this guide.

## Which AI Citations Stay Stable?

A small core of sources stays while the rest rotates. SISTRIX found a stable core of domains for 86% of Google AI Mode prompts, with the remaining domains rotating at 89% per week, and a brand's own domain held its place for all 17 weeks on 43% of brand queries.

The studies show four patterns in what stays and what rotates:

- **The brand's own domain on brand queries:** In SISTRIX data, the brand's own domain was present in all 17 weeks on 43% of brand queries, while co-cited domains rotated at 70% per week. Ahrefs found that when a question named a brand, that brand's site was cited on all 20 days in 28 to 33% of cases on Google's tools and ChatGPT, against 4.0 to 6.8% for any cited site.
- **Pages that rank in Google, on AI Overviews:** Ahrefs found 29% of pages that ranked in Google for the same question were still cited in AI Overviews for 7 or more days, against 4% of pages that did not rank. That covers 230 questions and shows an association, not a cause.
- **Brand names more than cited URLs:** About one in five brand names appeared on only one day in the Ahrefs data, compared with 50 to 70% of cited websites. That is why it pays to record [brand mentions and citations as two separate signals](https://growthtribune.com/how-to-separate-ai-brand-mentions-from-citations/).
- **Few news articles persist:** SISTRIX found that only 1.4% of cited news articles stay in the citation set permanently.

So are brand mentions more stable than citations? In the Ahrefs data, yes. A brand name tends to recur across days even when the page cited for it changes.

## Why Do AI Citations Change?

Run-to-run variation in the answers themselves is one documented reason AI citations change. [In a SparkToro and Gumshoe study of nearly 3,000 runs, the odds of an AI tool returning the same brand list twice were under 1 in 100](https://searchengineland.com/ai-recommendation-lists-rarely-repeat-study-468076), and the odds of the same list in the same order were closer to 1 in 1,000.

Three more drivers show up in the studies:

- **Each engine draws on different sources.** SISTRIX found Google AI Overviews and Google AI Mode cite different domains 83% of the time, a Jaccard index of 0.17. AI Mode and ChatGPT overlap less still, at 0.125.
- **A stable core sits inside a rotating ring.** For 86% of AI Mode prompts, a few domains stay while the rest rotate at 89% per week. Most measured turnover happens in that outer ring.
- **Freshness may play a part.** An AirOps study, as reported by Alex Birkett, covered 4,000+ ChatGPT-cited pages across 900 high-intent queries in 15 industries. For commercial queries, 83% of citations came from pages updated within the past year, and pages not updated for over a year were more than twice as likely to lose citations. That is one study, reported second-hand.

None of these studies isolates a single cause. What they show together is that variation between runs, differences between engines and a churning outer ring of sources all add up to the turnover rates above.

## How To Measure AI Citation Turnover Yourself

To measure AI citation turnover, rerun a fixed prompt panel on a fixed schedule and compare the cited URLs between runs. [Alex Birkett compared the top 500 cited URLs from about 100 prompts one month apart and found only 49% in both snapshots](https://alexbirkett.com/citation-drift/), a URL-level Jaccard similarity of 0.32.

> **Jaccard similarity:** The number of sources two runs have in common divided by the number of distinct sources across both runs. A score of 1 means the sets are identical and 0 means they share nothing.

The protocol, step by step:

- [ ] Fix the prompt panel, the engines and the regions before the first run, and change none of them mid-series
- [ ] Run every prompt on a fixed schedule, daily to catch next-day loss or weekly to track drift
- [ ] Log every cited URL with its run date, one row per citation
- [ ] Record answers with no citations as their own count, never as lost citations
- [ ] Normalise brand names and domain variants before comparing runs
- [ ] Compute churn-in, retention and Jaccard similarity at both URL and domain level
- [ ] Collect several weeks of runs before reporting a rate
- [ ] Publish the method, dates, denominators and limitations next to every figure

Each step answers a problem the studies ran into:

- **Fixed schedule:** Ahrefs ran daily for 20 days and SISTRIX weekly for 17 weeks. Mixing intervals makes the rates incomparable.
- **Empty answers:** Gemini and Copilot left 64% and 53% of answers uncited in the Ahrefs study, which inflates apparent losses if counted as drops.
- **Brand normalisation:** Ahrefs traced 30 to 36% of apparent losses in a 55-brand check to the same brand under another name.
- **Two levels:** SISTRIX puts URL-level drift about 15% above domain level, so report both and label each.
- **Several weeks:** SISTRIX saw no levelling off over 17 weeks, and SparkToro and Gumshoe found lists rarely repeat, so a single comparison says little.

Churn-in is the share of domains or URLs in a run that were not in the previous run. Retention is the share from the previous run that are still there. Keep [one row per answer and one denominator per metric](https://growthtribune.com/how-to-measure-ai-visibility-without-double-counting/) so runs with different answer counts stay comparable.

Copy this log template into a sheet or table, one row per cited URL:

**AI citation turnover log**

```
prompt_id,engine,region,run_date,answer_status,cited_url,cited_domain,brand_normalised
{{prompt-id}},{{engine}},{{region}},{{run-date}},{{answer-status}},{{cited-url}},{{cited-domain}},{{brand-normalised}}
```

Store the log somewhere that keeps history. A free BigQuery sandbox deletes tables after 60 days, which is shorter than the series a turnover report needs.

## Which Tools Record AI Citations Over Time?

Four AI visibility trackers log cited URLs on a schedule you can export, and a warehouse such as BigQuery turns those exports into a turnover series. Profound runs prompts daily by default, OtterlyAI lists daily tracking on its self-serve plans, Rankscale schedules from hourly to monthly and LLMrefs updates every keyword at least once a week.

**Tools for logging AI citations over time**

| Tool | Run frequency | Cited-URL export | Panel limitation |
| --- | --- | --- | --- |
| Profound | Daily by default | CSV or JSON | 24 to 48 hours to start |
| OtterlyAI | Daily | CSV with URL and date | Some engines are add-ons |
| Rankscale | Hourly to monthly | CSV, Google Sheets | No export on Essentials |
| LLMrefs | At least weekly | CSV | Weekly minimum update |
| BigQuery | Scheduled queries | Stores tracker exports | Sandbox tables expire |

This is a view of how each tool supports a turnover panel, not a ranking. For a full comparison, see the guide to [AI citation tracking tools](https://growthtribune.com/best-ai-citation-tracking-tools/).

### Profound

- **Records:** The Answer Engine Insights dashboard shows which websites AI answers cite.
- **Runs:** Prompts run daily by default across every selected AI platform and region, and teams with high prompt volumes can customise frequency and scheduling.
- **Export:** Data exports as a CSV or JSON file from the Answer Engine Insights tab.
- **Panel limitation:** New prompts take 24 to 48 hours to produce first data, so a panel's first day is not day one of the series.

### OtterlyAI

- **Records:** Citations exports, full and summary, include a URL field, and the full export adds a Date field. Raw AI responses export as CSV or JSON, and a Looker Studio connector is available on Standard, Premium and Enterprise.
- **Runs:** The Lite, Standard and Premium plans list daily tracking.
- **Export:** CSV citations exports carry the cited URL and run date a turnover log needs.
- **Cost driver:** Extra blocks of 100 search prompts are sold monthly on Standard and Premium. The Google AI Mode and Gemini add-on runs from $9 a month on Lite to $149 on Premium, and the Claude add-on reaches $439 a month on Premium.
- **Panel limitation:** ChatGPT, Google AI Overviews and Microsoft Copilot come on every plan, but Google AI Mode, Gemini and Claude are paid add-ons, so a panel covering every engine costs more.

### Rankscale

- **Records:** Citations analysis and sources box analysis come on all plans, with domains and URLs in AI source boxes open to inspection.
- **Runs:** Scheduling is listed as hourly to monthly on every plan, across 17+ AI engines including ChatGPT, Google AI Mode, AI Overviews, Gemini, Claude and Microsoft Copilot.
- **Export:** CSV and Google Sheets exports are included on Pro and above.
- **Cost driver:** Plans are metered in monthly credits, and each answer collected uses credits, so a daily panel burns credits faster than a weekly one.
- **Panel limitation:** Essentials has no CSV or Google Sheets export.

### LLMrefs

- **Records:** The dashboard shows the full set of source URLs AI engines cite, with filters.
- **Runs:** Every keyword updates at least once per week.
- **Export:** Data exports as CSV.
- **Cost driver:** The All in One plan tracks 500 prompts for brand mentions, sources and fan-out queries, with geo-targeting in 50+ countries and 20+ languages and unlimited seats and projects.
- **Panel limitation:** A weekly minimum update suits tracking drift over weeks.

### BigQuery

- **Records:** BigQuery stores whatever the trackers export, so the turnover log sits in tables the team controls.
- **Runs:** Scheduled queries automate recurring GoogleSQL work at intervals as short as 5 minutes, and the @run\_date parameter stamps each run with its calendar date.
- **Export:** Churn-in, retention and Jaccard similarity can be computed in GoogleSQL over the stored log.
- **Cost driver:** The first 1 TiB of query data processed and the first 10 GiB of storage each month carry no charge, and on-demand query pricing applies beyond that.
- **Panel limitation:** The sandbox needs no billing account, but its tables expire after 60 days, storage is capped at 10 GiB for life, and it lacks the BigQuery Data Transfer Service that scheduled queries rely on.

## What Citation Turnover Means For AI Visibility Reporting

Because most AI citations rotate within weeks, a citation report should show how often a source appears across many runs, not whether it appeared once. SparkToro and Gumshoe recommend tracking how often a brand appears across many prompts run many times, and found some brands in 60 to 90% of responses for an intent.

Four reporting rules follow from the evidence:

- **Report presence rate, never position.** SparkToro and Gumshoe call AI rank tracking meaningless, since the same list in the same order comes up about 1 time in 1,000.
- **Separate the brand's own domain from co-cited pages.** SISTRIX found a brand's domain stayed all 17 weeks on 43% of brand queries while co-cited domains rotated at 70% per week. Blending them hides both trends.
- **Report each engine on its own.** Google AI Overviews and Google AI Mode share only a fraction of their cited domains, so a combined figure describes neither.
- **Expect monthly drift and alert above it.** Alex Birkett saw 49% of URLs and 65% of publishers survive a month. An alert threshold set below normal churn fires every week.

Brand names hold steadier than cited URLs, so a report that tracks both shows whether a brand is losing ground or only its cited pages are rotating.

> **Tip:** Report each source's presence rate over at least several weeks of runs, by engine, with the brand's own domain shown apart from co-cited pages. Treat a single lost citation as noise until the rate moves.

**Verdict.** AI citations turn over fast, with half or more of the sources gone the next day and most weekly sets rotating, yet a small core and a brand's own domain persist. Teams that measure presence across a fixed panel over weeks, at URL and domain level, can tell real loss from normal churn. Single snapshots and AI rank positions cannot.

## FAQ

**Do AI citations change every time you ask?**

Yes, AI citations can differ between two runs of the same prompt, because AI answers vary from run to run. SparkToro and Gumshoe put the odds of the same brand list appearing twice at under 1 in 100. Ahrefs found 50 to 60% of cited websites gone the next day once empty answers were left out.

**Is a lost AI citation gone for good?**

Usually not. In the Ahrefs study, about 75% of citations that dropped out came back within 20 days, and about half returned within three days. Between 20% and 27% of day-one citations never returned, depending on the tool, so a loss that lasts several weeks deserves a closer look.

**Do AI citations differ by country?**

The rate of change looks similar across countries. SISTRIX measured weekly drift between 54% and 59% across Germany, the USA, the UK, Italy, Spain and France, with no sign of levelling off over 17 weeks. A team reporting several markets should still run each country as its own panel and compare like with like.

## Sources

- [How Stable Are AI Citations? 10 Things We Learned Tracking 37,500 Citations](https://ahrefs.com/blog/how-stable-are-ai-citations/), ahrefs.com
- [AI Citation drift: How stable are sources in AI search results?](https://www.sistrix.com/blog/ai-citation-drift/), sistrix.com
- [AI Search Citations: How Much Do They Really Change Week to Week?](https://brightedge.com/resources/weekly-ai-search-insights/ai-search-citations-week-to-week-changes), brightedge.com
- [AI recommendation lists repeat less than 1% of the time: Study](https://searchengineland.com/ai-recommendation-lists-rarely-repeat-study-468076), searchengineland.com
- [Citation Drift: How Stable Are AEO Citation Sources (Research)?](https://alexbirkett.com/citation-drift/), alexbirkett.com

---
Source: https://growthtribune.com/original-study-how-often-do-ai-citations-change/