Scrunch AI published a study based on 3.5 million citation events and found that the half-life of an AI citation is 4-5 weeks. Profound published a study based on 883,000 pages and found that the median half-life is 11 days. Both are reputable, data-driven companies. Both used real data. Both are correct, and they disagree by a factor of 3.
This is why AI visibility research needs pre-registration and methodological transparency. Two good studies produce two different numbers because they measure two different things, and neither study makes the difference explicit.
What Scrunch measured
Scrunch AI, acquired by Sitecore in June 2026, analyzed 3.5 million citation events. Their finding: a citation has a half-life of 4-5 weeks, meaning that after 4-5 weeks, half of the citations that were present at time zero have disappeared.
The key methodological choices in Scrunch's study:
- They tracked citations across multiple AI engines (ChatGPT, Google AI Overviews, Perplexity)
- They counted a citation as "present" if the domain appeared in the source list of an AI answer
- They measured at the domain level, not the page level
- Their time window appears to be longer than Profound's, which would naturally produce a longer half-life estimate
What Profound measured
Profound analyzed 883,000 pages across ChatGPT and Gemini. Their finding: the median half-life of an AI citation is 11 days, meaning that after 11 days, half of the citations have disappeared.
The key methodological choices in Profound's study:
- They tracked citations across ChatGPT and Gemini specifically
- They measured at the page level (specific URLs), not just the domain level
- Their 11-day median is across both engines combined
- Jennifer Zou, an economist, designed the study with standard methodology
Why they disagree
The 4-5 weeks vs 11 days gap is not a contradiction. It is a difference in what they measured, driven by three factors:
1. Domain vs page level. A domain can be cited for different pages at different times. If ChatGPT cites example.com/page-a this week and example.com/page-b next week, the domain citation persists but the page citation changes. Scrunch measured domain-level persistence, which is naturally longer because a domain can rotate pages. Profound measured page-level persistence, which is shorter because individual pages get replaced faster than domains.
2. Engine mix. Scrunch included Google AI Overviews, which the Babel42 study showed are extremely volatile (only 11% of AI Overviews cite the same sites two days in a row). Profound measured ChatGPT and Gemini. Different engines have different citation stability. Google AI Overviews are more volatile than ChatGPT, so including them in the mix would shorten the half-life. But Scrunch's longer half-life suggests their engine mix or measurement window compensates for this.
3. Definition of "citation." Both studies count citations, but the definition of what counts as a citation may differ. Does a brand mention without a link count? Does a source that appears in the "additional sources" section count? The BrightEdge data showed that ChatGPT mentions 3.2x more brands than it cites. If Scrunch counts mentions and Profound counts only linked citations, the half-life would differ.
Why this matters
This disagreement is not academic. It directly affects how you plan your AI visibility work:
- If citations decay in 11 days (Profound), you need to actively maintain your visibility every 2 weeks. A month-old citation is already gone.
- If citations decay in 4-5 weeks (Scrunch), you have more time. A citation earned today is still half-present next month.
The answer is: both are correct for their definition. The practical implication is that you need to know which definition your tool uses. If your tool tracks domain-level citations, expect 4-5 week half-life. If it tracks page-level citations, expect 11-day half-life. Plan your monitoring cadence accordingly.
What a pre-registered reconciliation would look like
Elmo, the open-source competitor, writes "research reconciliation" posts that take conflicting studies and explain why they disagree. This is useful. But Elmo does not pre-register the reconciliation. He reads both studies, explains the difference, and publishes. The explanation is correct but not tested.
A pre-registered reconciliation would:
- State the hypothesis: "domain-level citation half-life is longer than page-level citation half-life"
- Define the methodology: track both domain and page citations for the same prompts over 30 days
- Pre-register the analysis plan before collecting data
- Run the test
- Publish the result, whatever it is
This is what I plan to do. The calibration study was the first pre-registered test. A citation half-life reconciliation would be the second. The hypothesis is simple, the methodology is straightforward, and the result would resolve the Scrunch vs Profound disagreement with data, not with explanation.
What this means for the industry
The Scrunch vs Profound disagreement is not unique. It is representative of the entire AI visibility research space. Different tools measure different things, use different definitions, and report different numbers. None of them pre-register. None of them publish their full methodology. None of them have been independently replicated.
The calibration study showed that on-page GEO scores do not predict citation. The Toronto study showed that on-page signals explain 2.8% of citation variation. The Scrunch vs Profound disagreement shows that even basic metrics like citation half-life are not standardized across tools.
The pattern is consistent: the AI visibility industry sells metrics that have not been validated, uses definitions that are not standardized, and produces results that are not reproducible. Pre-registration and methodological transparency are the fix. No competitor does this. It is the gap I am building on.
What to check
If you want to check whether AI crawlers can read your page, run a free audit on geoready.dev. The audit is the foundation. If your page is not readable, citation half-life is irrelevant.
The CLI is on GitHub, MIT licensed, 1,137 stars. The calibration study is open. The next study will be the citation half-life reconciliation, pre-registered, with raw data published.
4-5 weeks or 11 days. Both are correct. Neither is standardized. The industry needs pre-registration to resolve disagreements with data, not with blog posts.
