Citation-Domain Persistence After One Week: A 177-Series Baseline
A first-to-last comparison of visible source domains across repeated AI answers—and why the result is not yet a citation half-life.
Citation “half-life” is an appealing idea: measure how quickly a source disappears from AI answers and reduce volatility to one number. Our data says the first useful step is more modest—measure domain retention over a fixed interval and do not call it a half-life until enough intervals exist to fit a decay curve.
We compared the visible citation-domain set in the first and last observation of repeated prompt-model series. The average gap ranged from four to eight days, depending on the model.
Perplexity retained 82.1% of first-run citation domains on average after four days in this cohort. Claude retained 37.2% after 4.9 days; Google AI retained 34.2% after eight days.
The measure
For every project-prompt-model series with at least two observations, we selected the first and last completed run. We normalized citations to publisher domains, deduplicated domains within each answer, and calculated:
retention = domains in both first and last answer / domains in first answer
Only series with at least one visible citation domain in the first run entered the retention calculation.
| Model | Series | Avg. gap | Mean retention | Median retention |
|---|---|---|---|---|
| Perplexity | 48 | 4.0 days | 82.1% | 82.4% |
| Copilot | 26 | 4.2 days | 68.5% | 75.0% |
| ChatGPT | 48 | 4.0 days | 62.4% | 66.7% |
| Claude | 25 | 4.9 days | 37.2% | 20.0% |
| Google AI | 30 | 8.0 days | 34.2% | 33.3% |
Gemini was excluded. Its connector exposed the same Google proxy hostname for 1,879 citation records, making domain retention appear to be 100%. That is a connector artifact, not evidence of perfect source stability.
This is persistence, not half-life
A half-life requires repeated observations over enough time points to estimate a decay process. Two endpoints cannot tell us whether change was gradual, immediate, or reversed between runs.
This paper therefore reports first-to-last domain retention. It is descriptive and reproducible. The pending half-life experiment remains unpublished until it has a longer panel with scheduled intervals and model-version controls.
That distinction matters. Research becomes less useful when a convenient label outruns the design.
Finding one: source stability is model-specific
Perplexity preserved more than four-fifths of first-run domains on average across 48 series. ChatGPT retained roughly three-fifths. Claude and Google AI retained about one-third in their measured windows.
The values should not be interpreted as a permanent ranking. Google AI's average interval was twice as long as ChatGPT's and Perplexity's, and prompt/project coverage differs. But the spread shows why a single global “citation freshness” assumption is unsafe.
A source can be durable in one answer system and transient in another.
Finding two: stable brand presence can hide source churn
A brand can remain in an answer while the supporting sources change. Conversely, a familiar source can remain while the brand drops out of the shortlist.
This creates two monitoring tracks:
- Outcome persistence: whether the brand remains present, recommended, and described consistently.
- Evidence persistence: whether the same domains and pages continue to appear.
The run-to-run variance study found brand-presence flips in 6.3% to 23.3% of repeated series by model. Citation retention adds the upstream evidence layer. Monitoring only one of the two cannot explain the other.
Why citation domains change
The data records change, not cause. Several mechanisms are plausible and should be tested rather than asserted:
- the engine generated different search queries;
- the retrieval index or page inventory changed;
- a reranker promoted a different passage;
- freshness or location signals shifted;
- the source page changed or became unavailable;
- the answer interface presented a different subset of internally used evidence.
Official APIs make parts of this observable. Google returns grounding search queries and chunks for grounded answers. Perplexity exposes structured search results. OpenAI web-search outputs can contain URL citations. Retaining those raw fields makes later attribution analysis possible.
A better longitudinal design
To estimate a true citation half-life, run a fixed prompt panel at predetermined intervals—daily for two weeks, then weekly for at least two months. Pin model versions where the provider allows it and preserve the effective model otherwise.
For each page and domain, record first seen, last seen, reappearance, and censoring. Use a survival curve rather than treating every disappearance as permanent. A citation that returns after two missed runs did not “die” in the ordinary sense.
Stratify by model, prompt family, source type, and whether the page is owned or third-party. Report connector changes as breaks in the series.
That is the study we are continuing. The current four-to-eight-day sample is the baseline, not the final decay model.
Operational use
Even a short retention window can improve decisions.
A domain that appears in nearly every run across several prompts is a structural source. It deserves page-level analysis and relationship work. A domain that appears once is evidence of eligibility, not durable influence. A newly recurring source is more important than a one-run novelty.
Track three thresholds:
- Recurring: appears in at least half of recent runs for a prompt.
- Cross-prompt: appears across multiple questions in the same buying job.
- Cross-model: appears in more than one answer engine.
Sources that clear all three are the strongest candidates for the co-citation map.
Limitations
The comparison uses first and last observations only. Intervals differ by model, and the cohort is operational rather than randomized. Domain-level retention can hide page-level turnover inside the same site. Visible citations are not a complete retrieval log. Small first-run source sets make percentage retention coarse. Connector behavior can dominate results, as the excluded Gemini proxy demonstrates.
Most importantly, these estimates are not half-lives and should not be extrapolated beyond the measured intervals.
Reproducibility and data
The series counts, intervals, means, medians, and exclusion note are in the aggregate JSON.
Primary sources and evidence
- Google: grounding queries and chunks
- Perplexity: structured search results
- OpenAI: web search tool
- Anthropic: citation structures
Continue through the system
- Run-to-Run AI Visibility Variance
- The Third-Party Citation Tax
- Build a Brand-Mention Monitor Across Four LLMs
Build your baseline with the free AI visibility scan.