How the CiteCurrent benchmark works
The questions, services, schedule, and classification rules behind CiteCurrent's monthly measurement of the sources AI search engines cite.
Published
Draft. The final version comes with the first report.
The short version
Each month we put a fixed, published set of questions to six AI search services, three times each. We record every source the answers show. Each cited site is classified by who publishes it. Each report gives every kind of source’s share by service, with the monthly change and the raw data. We repeat each question because a service can cite different sources for the same question from one run to the next.
The questions
The core panel has 80 questions, 10 in each of eight topics, from personal finance to food. They are phrased the way people type into a search box and name a brand only when comparing two products. The full list gives each one’s ID, topic, and intent.
| Intent | Questions | Example |
|---|---|---|
| Information | 25 | How much sleep do adults need? |
| Buying | 30 | What is the best air fryer? |
| How-to | 18 | How do I fix a running toilet? |
| Comparison | 7 | Is Airbnb or a hotel better for a family trip? |
The wording is frozen. An ambiguous question is retired and replaced under a new ID, never edited. Candidate questions run in a separate exploratory set, reported apart. The first panel runs in US English.
The services
- ChatGPT (OpenAI), with web search
- Perplexity
- Microsoft Copilot
- Gemini (Google)
- Google AI Mode
- Google AI Overviews, the summary at the top of some Google results pages
Answers are collected through a third-party data service that records the answer each service’s website shows, from the United States. We keep its copy of every page as displayed, so any figure can be checked against what the service showed. Google doesn’t show an AI Overview for every search, so a missing Overview is recorded as a result.
What we record
For each run:
- the question ID, the service, and the date and time (UTC);
- the full answer, and every source it shows, in order;
- whether each source was cited in the answer text or only listed beside it;
- the copy of the page as displayed.
A failed run is recorded with its reason, not quietly retried.
Citations, mentions, and recommendations
A citation is a source the answer shows, either linked in its text or listed beside it. A mention is a brand or product named in the text. A recommendation tells the reader to choose something. We count each separately, because a site can be cited in an answer that recommends a competitor.
How sources are classified
Each cited site gets one class, by who publishes it.
| Class | Covers |
|---|---|
| Earned media | News, trade press, and review publishers |
| User-generated | Forums, Q&A sites, and user reviews |
| Brand-owned | Vendors and retailers |
| Reference | Encyclopedias, dictionaries, and medical references |
| Social and video | Social networks and video platforms |
| Government and academic | Government bodies and universities |
| Other | Reviewed sites that fit none of the above |
Classes come from a reviewed, versioned list of domains. A domain that isn’t on the list goes to a person, never a guess. Each report says how many citations needed review.
Headline statistics
Every figure gives its numerator and its denominator.
| Statistic | Counts | Out of |
|---|---|---|
| Share of citations | Citations to sites in one class | All citations on that service that month |
| Answers citing a class | Answers showing at least one source in the class | All answers on that service that month |
| AI Overview rate | Searches where Google showed an AI Overview | All Google searches in the panel that month |
Answers to the same question are related, so uncertainty comes from resampling whole questions. Each share comes with a 95% interval.
What the numbers can’t show
- We chose the questions. They are not a sample of what people ask, so the shares describe this panel only.
- Answers change between sessions, and signed-in users may see different ones. Ours are recorded from the United States.
- Services change their models and retrieval without notice. A monthly change may come from a service update.
- A citation shows that a service linked to a page. It doesn’t show that the page shaped the answer, or that anyone clicked.
- Every chart states its sample size, dates, service, and country.
Data and reuse
Each report links to its observations as CSV and JSON: every run, with its sources, their positions, and their classes. We publish counts and cited URLs, not the text of the answers. A data dictionary comes with them, and each download names the versions of the question panel and domain list it used. A new method version means regenerating whole editions, and we never re-run questions for a brand on request. Corrections to published data are versioned and logged on the corrections page.
Changes to this method
- : First draft, before the first collection run.