In short. Compare timestamped answer snapshots using the same queries and recorded conditions, with repeat measurements inside each collection window.
Definition: Comparing citation observations across dates means comparing timestamped records of answers and their linked sources under recorded search conditions. Repeated measurements within each collection window help assess how much responses vary before interpreting a difference across dates.
Key facts
- The Source Coverage and Citation Bias study describes its measurement as a snapshot of popular queries and user-facing citations.
- The Don't Measure Once study compares daily observations and a separate repeated-run collection, restricting repeated-run pairs to a timestamp difference of at most 24 hours.
- As of October 2026, Google's search variation guidance identifies newly available content and gradual ranking-system rollouts as reasons search results change.
What does the Source Coverage study actually date?
The December 2025 Source Coverage methodology collects trending seed keywords between 16 July and 10 August 2025. That interval describes query construction. A comparison record should distinguish the query collection period, the time an answer is captured and the publication date of the cited page.
The study uses incognito browser sessions and disables memory in its authenticated ChatGPT setup. It extracts citations from response payloads and counts each URL once per response. These are useful controls to carry into a repeat-measurement protocol.
What should a timestamped snapshot contain?
Google's guidance on differing results identifies location, language and device type as search context, including when personalized recommendations are disabled. Recording those conditions makes later comparisons easier to interpret.
The following is a suggested observation-sheet header, adapting that context and the Source Coverage study's response collection. It is an editorial template, with no sample observations:
observed_at_with_timezone,exact_query,engine_and_search_feature,displayed_model_label,location,language,device,login_and_memory_settings,collection_method,saved_response_path,cited_urls
Each run receives its own record. The saved response preserves the answer and visible links, while the timestamp identifies when that evidence is collected. A page's publication date belongs in a separate field if freshness is being studied.
How should repeated observations be compared?
The April 2026 Don't Measure Once methods distinguish daily observations from repeated runs over short intervals. A practical adaptation keeps a fixed query panel, meaning the same saved set of exact queries, and repeats it within each comparison window. Changed queries or collection settings remain identifiable in the report.
For source overlap, the paper uses Jaccard similarity: the number of sources shared by both sets divided by the number appearing in either set. Following the Source Coverage counting rule, duplicate links within an answer count once. The comparison should state whether a source means an exact URL or a domain.
For a particular website, a separate descriptive measure is the number of valid responses containing its link divided by the number of valid responses collected. Report both counts alongside the proportion, using the same inclusion rule across windows. Keep failed captures separate from valid answers containing no citations.
How should a change across dates be reported?
Google's explanation of search variation says newly available content can alter results within minutes or hours. A before-and-after citation difference alone cannot identify which change caused it.
A comparison report should therefore give the collection windows, matched queries, valid-run counts and recorded setting changes alongside its results. Repeat observations provide context for deciding whether an apparent increase persists across runs.
Related terms
FAQ
Can I compare an old screenshot with a new answer?
You can describe the visible difference. The Source Coverage study's snapshot approach supports keeping each captured answer as dated evidence; comparable repeat runs add information about variation.
Should I replace the earlier observation when I rerun a query?
Keep both records with their timestamps. The Don't Measure Once study uses timestamp differences to determine which repeated-run pairs enter its short-window comparison.
Sources
- arXiv, "Source Coverage and Citation Bias in LLM-based vs. Traditional Search Engines", 10 December 2025
- arXiv, "Don't Measure Once: Measuring Visibility in AI Search (GEO)", 8 April 2026
- Google Search Help, "Why your Google Search results differ from others", accessed 17 October 2026
Check how a page reads to answer engines. Open the GEO score analyzer