%20-%20kopie.png)
Cart
0
your cart
No items found.
Product is not available in this quantity.
By AI Marketing Entrepreneurship · Published October 13, 2026
Prepared with AI assistance using linked primary sources. Facts checked October 13, 2026.
An AI citation study starts with the questions your audience asks, then records the sources each assistant actually displays. The useful result is a dated map of sources and content gaps, not a permanent ranking of websites. This guide provides a complete 50-question starter set and a repeatable method; it does not present findings from a study already conducted.
For a marketing publisher, the commercial question is straightforward: where could your own experience add something that existing answers lack? A citation log can help you choose the next detailed guide, comparison, or worked example. It cannot reveal an assistant's private ranking system.
Use three separate labels. A brand mention names an organisation. A citation explicitly links a claim or passage to a source. A related link appears as further reading without a clear claim-level connection. Do not combine them into one visibility number.
ChatGPT's search documentation says responses may include citations and other relevant links. Gemini's help page also describes sources and related content, and explains that some responses provide neither. A response without visible sources is therefore an observation to record, not proof that no external information influenced the answer. [S2, S3]
Run each question in a fresh conversation. Record the product, visible model or mode label, account plan, date, language, location settings, and whether web search was enabled or observed. Keep these conditions as consistent as the products permit. Also record memory, custom instructions, connected sources, and other available personalisation settings. A fresh conversation alone is not a complete control: OpenAI says saved memories can influence search-query rewriting. For a public-web baseline, disable avoidable personalisation and private-source connections where possible, and record anything you cannot standardise. Avoid uploading your own site or asking the assistant to include it during the baseline. [S2]
Use the same 50 questions on both platforms and repeat them on three separate dates. That produces 300 planned responses: 50 questions multiplied by two platforms and three rounds. Mark technical failures and unavailable features separately, using exclusion rules set before collection. Keep completed answers with no citations in the denominator; report refusals separately and state how they are counted. Do not quietly replace an inconvenient response with a more favourable retry.
Choose the cadence before beginning. Three weekly rounds are a workable exploratory design, not a statistically representative sample of every user or query. If a model changes mid-study, label the change and analyse those observations separately.
These prompts target the readership of an AI marketing publication. Adapt the market and audience once, then freeze the wording for the study. The questions intentionally mix learning, implementation, evaluation, and purchasing decisions.
Give every run a unique identifier. Save the exact prompt, complete answer, displayed source URLs, observation time, and source category. Open each cited page and check whether it supports the associated statement. A plausible link is not automatically a valid citation.
Classify sources as platform documentation, independent editorial, vendor content, research, community discussion, or another explicit category. Keep the full URL while also recording the domain. Deduplicate a domain within a response when calculating response-level coverage; keep multiple page citations in a separate detail table.
| Measure | Suggested definition |
|---|---|
| Cited-domain coverage | Valid responses containing at least one citation to a domain, divided by all valid responses |
| Citation-bearing response rate | Valid responses with an explicit citation, divided by all valid responses |
| Verified support rate | Checked citations supporting their associated statement, divided by checked citations |
Inspect the cited page, the question it answers, and the evidence it offers. If answers repeatedly lack a worked implementation example, create one that readers can reproduce. If an existing source already answers the question well, add a distinct contribution rather than rewriting it.
Report counts alongside percentages and separate the platforms and rounds. A hypothetical result of 12 citation-bearing responses for one domain out of 150 valid runs is 8% coverage in that sample; it is not 8% of the platform's users or a measure of market share. Publish the question set, collection method, exclusions, and limitations with any eventual findings.
No. The study observes outputs. It can suggest useful follow-up questions, but it does not isolate the effects of backlinks, wording, brand familiarity, or other possible influences.
Pilot five questions, verify the logging method, and then run the full set unchanged. The first useful outcome is a reliable record of what happened, followed by one evidence-based improvement to your content.
%20-%20kopie.png)
