57% of AI-Cited Domains Appeared in One Month Only
57% of AI-Cited Domains Appeared in One Month Only
Quick answer: A Somantra analysis of 2,437,107 citation records across 28,725 domains in Australian insurance found that 57.2% of domains appeared in exactly one month of a seven-month observation window, while only 2.7% appeared in all seven months [1]. The study does not show that those domains will never return, nor does it establish why individual citations disappeared. It does show why a one-time citation is a weak measure of durable AI visibility. Brands need longitudinal metrics that separate new, retained, lost, and recovered citations by prompt and engine.
One citation proves that a source entered an answer once. It does not prove that the source has become part of an engine's dependable citation set.
---
What did the “content graveyard” study measure?
The study examined citation persistence across seven observed months in one market. Somantra tracked ChatGPT and Google AI Overviews citations in Australian insurance between November 2025 and July 2026 [1] [2].
| Study element | Reported scope |
|---|---|
| Citation records | 2,437,107 |
| Unique domains | 28,725 |
| Observation window | Seven observed months between November 2025 and July 2026 |
| Market | Australian insurance |
| AI surfaces | ChatGPT search and Google AI Overviews |
| Domains appearing in exactly one month | 57.2% |
| Domains appearing in all seven months | 2.7% |
The unit behind the headline is the domain, not the individual page. A domain that appeared in one month may have received one citation or many citations during that month. The public summary does not provide enough detail to calculate how many individual URLs returned, how consistently the same prompts were run, or how persistence differed between ChatGPT and Google.
The phrase “never cited again” therefore needs a time boundary. The defensible statement is that 57.2% of the domains appeared in one month and did not reappear during the seven-month observation window. Anything beyond July 2026 falls outside the study.
---
Why is a one-time AI citation a weak success metric?
A single citation combines at least three events: the engine searched, retrieved a page, and selected it as a source for one answer. Each event can change on the next run.
An April 2026 information-retrieval preprint makes the measurement problem explicit: AI answers vary across runs, prompts, and time, so one-off observations are unreliable. The authors recommend repeated measurement and treating visibility as a distribution rather than a single-point outcome [5].
That distinction changes what a dashboard should celebrate. “We were cited” is an occurrence. “We remained eligible and were selected repeatedly for the same question set” is evidence of retention. Neither guarantees future placement, but the second signal carries more information.
A useful reporting hierarchy is:
- Occurrence: Was the domain or URL cited at least once?
- Repeatability: Did it appear again when the same prompt set was measured repeatedly?
- Retention: Did it remain present over multiple reporting periods?
- Recovery: After disappearing, did it return?
- Contribution: When cited, did the source actually support the answer or merely appear in the source list?
The Somantra finding concerns the first three levels. It does not measure whether the cited material shaped the answer, produced a brand mention, or influenced a recommendation.
---
How can AI citations look stable and volatile at the same time?
They can be stable at the core and volatile in the long tail.
A BrightEdge week-over-week analysis across five AI engines and nine industries found that 96.8% of cited domains recorded no change during the measured week. Yet among the roughly 3% that changed, 87% declined, and most changes were binary: a domain moved from cited to absent for a prompt rather than fading gradually [4].
Those numbers do not invalidate Somantra's seven-month result. The studies use different prompt sets, periods, industries, and definitions. Read together, they suggest a two-zone pattern:
| Citation zone | What the studies suggest | Reporting implication |
|---|---|---|
| Core | A smaller set of sources can remain highly stable over short intervals | Measure how much of your visibility comes from repeat citations |
| Fringe | One-off and borderline sources can enter briefly and then disappear | Do not count every new citation as durable growth |
BrightEdge also found that the highest-volume domains had a larger fringe, although only about 5% of their citation share was typically in play during the measured week. This matters because a large citation footprint can contain both a dependable core and a noisy edge [4].
The practical mistake is averaging both zones into one visibility score. A flat total can hide a shrinking core offset by one-time arrivals, while a rising total can be driven by citations that disappear in the next cycle.
---
Do comparison pages and FAQs survive longer?
Somantra observed an association, but the public evidence does not establish causation.
In the Australian insurance dataset, pages featuring comparison, FAQ, and discount or pricing structures appeared among persistent sources at roughly twice the rate seen among one-month sources. Content labeled as a “complete guide” was 3.5 times more common among domains that appeared in one month than among those present throughout all seven [1] [2].
A separate DeltaV Digital study offers partial support and an important limit. It analyzed 25,337 citations from 21,075 responses across five engines, eight industries, and a 90-day window. Comparison pages generated the highest citation rate at 1.87 citations per retrieval, 45% above the portfolio average of 1.29, but accounted for only 4.1% of total citations [3].
DeltaV's larger lesson was that no page type won everywhere:
| Industry example | Leading page type | Share reported by DeltaV |
|---|---|---|
| B2B technology services | Listicles | 61% |
| Local medical aesthetics | Homepages | 55% |
| Healthcare information | Articles | 54% |
| Higher education | Program pages | 53% |
The safe conclusion is narrower than “replace guides with comparison pages.” Structured comparisons can convert retrieval into citations efficiently in the observed datasets. Every category still has its own citation pattern, and Somantra did not run a controlled experiment that changed page format while holding authority, topic, freshness, and distribution constant.
A long guide can also contain strong comparison tables and self-contained answers. Length alone is not the tested variable. The operational question is whether a page contains passages that answer the recurring questions in its category clearly enough to remain useful.
---
Why might a citation disappear?
The available studies identify several failure points, not a complete causal model.
The answer itself is probabilistic
Repeated runs can produce different answers and source sets even when the visible prompt does not change. That is why the academic measurement literature recommends distributions and repeated samples [5]. A citation lost on one run may reflect output variance rather than a permanent source-level decline.
The retrieval process changes with the mode
A Semrush and Growth Memo study tested 100 prompts across 20 buyer journeys and reported only 25.6% overlap in cited domains between minimal and high-reasoning ChatGPT responses. High reasoning also performed more searches [6]. This is vendor research, but it shows why “ChatGPT visibility” is too broad a unit when modes assemble different source sets.
The page may sit in the fringe
BrightEdge's data suggests that changes cluster outside the stable core. A page can be relevant enough to enter one answer while remaining replaceable when the engine narrows its citation set or retrieves a more specific source [4].
The surrounding evidence changes
Competitors publish new material, official sources update their guidance, pages change, and technical access can improve or break. These are plausible inputs to citation turnover. None of the cited studies isolates their individual causal effect on a particular lost citation, so a diagnosis must inspect the affected prompt, engine, source set, page, and date rather than assigning every loss to “stale content.”
---
What should an AI citation-retention dashboard measure?
Retention needs fixed comparison units. Without the same prompt set, engine, mode, location, and sampling cadence, a month-to-month difference can reflect a changed test rather than changed visibility.
The following metrics are a practical reporting framework, not published ranking factors:
| Metric | Working definition | Question it answers |
|---|---|---|
| New citations | Domain–URL–prompt combinations present now but absent in the prior period | Where did visibility first appear? |
| Retained citations | Combinations present in both the current and prior period | What continued to be selected? |
| Lost citations | Combinations present previously but absent now | Where did selection stop? |
| Recovered citations | Previously lost combinations that reappear | Which losses were temporary? |
| One-period retention rate | Retained baseline combinations ÷ all baseline combinations | How much of the previous footprint survived one interval? |
| Three-period survival rate | Baseline combinations present in each of the next three periods ÷ all baseline combinations | How large is the durable core? |
Track each metric separately at domain and URL level. Domain retention can remain healthy while a specific product page loses its place; URL retention can fall while another page on the same domain replaces it.
The prompt is part of the key as well. A URL cited for “best payroll software for startups” has not retained the same position if it later appears only for a broad definition query. Topic-level visibility and prompt-level retention answer different questions.
---
How do you investigate a lost citation?
Use the loss as a diagnosis trigger, then work from the answer back to the page.
- Confirm the loss across repeated runs. One missing answer is a signal to recheck, not proof of a durable decline.
- Hold the test conditions constant. Use the same prompt, engine, mode, market, and account state where possible.
- Record the replacement sources. A lost citation is easier to explain when you know which pages entered the answer.
- Compare passage-level usefulness. Check whether the replacement answers the exact question more directly, with clearer evidence or a more suitable format.
- Check technical access. Confirm that the page remains crawlable, returns the expected status, exposes key facts in rendered HTML, and has not changed canonical or robots directives.
- Check factual currency. Update expired prices, dates, product availability, regulations, and source links. Do not change evergreen facts merely to create a new timestamp.
- Run a controlled repair. Change one meaningful variable, preserve the URL where appropriate, and measure the same prompt cohort over multiple periods.
This workflow avoids two common errors: rewriting a page after one noisy result, and changing several variables at once so the team cannot tell what influenced the next measurement.
---
What should content teams change?
The first change belongs in reporting. Separate citation acquisition from citation retention.
The second change belongs in planning. Build content around recurring buyer questions and the page types that your category repeatedly uses. DeltaV's data shows why a generic “AI prefers format X” playbook fails across industries [3].
The third change belongs in maintenance. Review pages when their evidence, products, rules, or cited sources change. A calendar-driven rewrite with no factual improvement can alter a useful passage without solving the reason it disappeared.
The fourth change belongs in experimentation. Comparison tables, FAQs, pricing explanations, official product pages, articles, and listicles should be tested against category-specific prompt cohorts. Somantra's format findings provide hypotheses worth testing; they do not provide a universal replacement schedule.
Finally, keep the page's identity stable when the topic remains the same. Replacing URLs, changing canonicals, or moving key facts behind client-side rendering can introduce technical uncertainty at the same time the team is trying to measure a content change.
---
Where Innflows fits
Citation retention requires repeated, comparable observations. A screenshot from one answer cannot provide them.
Innflows supports periodic monitoring built around simulated user questions across platforms including ChatGPT, Gemini, Google AI Overviews, Google AI Mode, DeepSeek, Qwen, and Grok. Its monitoring, source-verification, and website-structure layers can be used to establish a per-engine baseline, identify where brand visibility changes, and inspect whether the supporting source and page remain accessible and consistent.
For a citation-retention program, the practical workflow is to freeze a question cohort, preserve the cited domains and URLs for each run, and classify each observation as new, retained, lost, or recovered. Innflows can provide the repeated cross-platform measurement layer; the resulting changes still require human diagnosis. A lost citation may reflect answer variance, a retrieval shift, a stronger replacement source, or a page-level issue. No platform can infer causation from the disappearance alone.
That boundary is important. GEO can improve the probability that useful, accessible, well-supported content remains eligible for retrieval and citation. It cannot guarantee that a particular engine will keep citing a page on a fixed schedule.
---
Where does the evidence stop?
The headline comes from one vendor study in one vertical. Somantra sells AEO and GEO services, and the dataset covers Australian insurance. Its scale is large, but its market and commercial context limit generalization.
The public headline is domain-level. It does not show that 57.2% of individual pages disappeared, nor that every one-month domain received only one citation.
The observation window ends in July 2026. “Did not reappear in seven months” does not mean “will never return.”
Format findings are correlational. Persistent sources used comparison, FAQ, and pricing structures more frequently, but the study did not isolate format as the cause. Authority, query fit, distribution, technical access, and other variables may differ at the same time.
Different studies measure different kinds of stability. Somantra studied presence across seven months in insurance; BrightEdge compared one week across multiple industries; DeltaV examined retrieval-to-citation behavior across 90 days. Their numbers should inform different decisions rather than be merged into one benchmark.
---
FAQ
What is AI citation retention?
AI citation retention is the share of previously observed citation combinations that continue appearing in later measurements under comparable conditions. The combination should include at least the engine, prompt, cited domain or URL, and measurement period. It is an operational metric, not a ranking signal published by an AI platform.
Did 57.2% of AI-cited pages disappear after one month?
No. Somantra reported that 57.2% of domains appeared in exactly one month of a seven-month Australian insurance dataset [1]. The public headline does not establish the same percentage for individual pages, and it does not show what happened after July 2026.
Does one lost citation mean a page has failed?
No. AI answers vary across repeated runs, prompts, and time, so a single missing citation can be noise [5]. Recheck the same prompt and engine over multiple runs and periods before treating the change as a persistent loss.
Should brands stop publishing long-form guides?
The evidence does not support a universal ban. Somantra found “complete guide” content more frequently among one-month domains, while DeltaV found no single page type dominated across eight industries [2] [3]. A guide can remain useful when its sections answer specific questions clearly and its evidence stays current.
Are comparison pages better for durable AI citations?
They are a strong test candidate. Somantra associated comparison structures with persistent citation, and DeltaV found comparison pages had the highest citation rate in its dataset at 1.87 citations per retrieval [1] [3]. Neither study proves that changing a page into a comparison will cause longer retention in every category.
How often should AI citations be measured?
No universal cadence has been established. Choose a cadence that matches the speed of change in your market and use repeated runs within each period. Weekly measurement may suit fast-moving categories; a slower category may use monthly cohorts. Consistency of prompts, engines, modes, and markets is more important than claiming one schedule fits everyone.
---
Bottom line
Somantra's 2.4-million-record dataset found a large one-month citation tail: 57.2% of 28,725 domains appeared in exactly one month, while 2.7% appeared throughout all seven [1]. That is a reason to improve measurement, not a reason to declare most content permanently dead.
The useful distinction is between a stable core and a noisy fringe. Track citations as new, retained, lost, and recovered for fixed prompt cohorts. Test formats against your own category rather than copying a universal template. Investigate losses across answer variance, retrieval behavior, replacement sources, technical access, and factual currency before rewriting a page.
One citation is an event. Durable AI visibility is a pattern measured over time.
---
References
[4] - AI Search Citations: How Much Do They Really Change Week to Week? — BrightEdge, February 6, 2026
