How we measure citation.

Five steps, run on a schedule.

1. 1

   Define the questions

   We start from the questions buyers actually type: search suggestions, keyword data, sales-call transcripts and interviews. Each question is edited to read the way a person would ask it, and grouped by the decision it informs. The list is fixed before any engine is asked, so the results cannot shape the questions.

2. 2

   Sample the engines

   Every question is asked several times on each engine, in fresh sessions with memory off, from each market we study. Location is set at the network level; the prompt never names the country. We log the engine, model and settings shown, the date and time, and the full answer text.

3. 3

   Trace the sources

   Each cited URL is reduced to its canonical form and classified by source type, by who owns it (the brand being recommended, a rival, or a third party) and by the date it was last published or updated. Two researchers code independently; a third settles disagreements.

4. 4

   Act on the gap

   For client work, the coded sources become a target list: the pages that answer engines already cite for the client’s questions, and that do not mention the client. A researcher writes to each page’s editor, by name, from our outreach address.

5. 5

   Measure again

   The same question set is re-run on a fixed schedule, so every change is measured against a baseline rather than a screenshot. Reports state the number of answers, the dates and what changed in the method, if anything.

What we count as a citation.

A citation is a link an answer presents as a source: inline, in a footnote, or in the sources panel. Links mentioned in passing, but not presented as a source, are logged and kept apart.

• Once per answer. A URL cited twice in the same answer counts once.
• Canonical form. Tracking parameters, fragments and trailing slashes are removed, and redirects are followed to the final page.
• Answers, not screenshots. We store the full answer text and the list of sources for every run, with the date, time, engine and settings shown.

How we ask.

Answer engines do not return the same answer twice. A single answer is an anecdote, so every question is asked several times, on each engine and in each market we study.

• Fresh sessions. Every run starts a new conversation, with memory and personalisation switched off where the engine allows.
• Location at the network level. The prompt never names the country; the request comes from the market being studied.
• Fixed question sets. Questions are written and frozen before the first run, so results cannot shape the questions.
• Repeats. At least five runs per question per market. Studies state the exact number.

How we classify sources.

Every cited page is coded on three things: what kind of page it is, who owns it, and when it was last published or updated. Two researchers code independently, and a third settles disagreements. We report their agreement.

Third-party lists & roundups

Pages whose job is to rank or shortlist products: “best of” lists, roundups, buyer’s guides.

Comparison & review sites

Head-to-head comparisons, review platforms and editorial reviews of a single product.

Brand’s own site

Any page owned by the brand named in the answer: pricing, docs, product pages, blog.

News & magazines

Reporting and features from news organisations and trade titles.

Forums & communities

Discussion threads, Q&A sites and community posts.

Reference & wikis

Encyclopaedias, glossaries, standards bodies and official reference pages.

Marketplaces

Listings on stores and marketplaces.

Original research

Studies, surveys, benchmarks and datasets.

Video

Video pages and their transcripts.

Owner is one of: the brand named in the answer, a rival brand, or a third party. Age is the most recent visible publish or update date; where no date is shown, the earliest archived capture. Pages that cannot be dated are reported as such, never guessed.

How we report.

• Every number carries its sample. A share is printed with the count behind it and the dates it covers.
• Limitations come with the findings. Each study lists what it cannot tell you: engines and markets not sampled, question types out of scope, changes the engines made while we measured.
• Data is released. Question lists, answer logs and coded citations are published with each study, or provided on request where an engine’s terms do not allow republication.
• Corrections are versioned. A study that changes gets a new version number and a changelog. The old version stays readable.

The same method, for one category.

A citation audit is this protocol run on a client’s own buyer questions. Citation building is step four: outreach to the editors of the pages the engines already cite. Progress is reported by re-running the same questions, not by counting emails sent. See Citation building.

Questions about the method: research@citationresearch.com