Recommendation–Citation Gap
A predeclared panel, collection rules, annotation definitions and analysis tools for independent review.
Read the protocol →HOWAICITE RESEARCH
Research that starts with a checkable method.
Study questions, source data and explicit limitations. Our first protocol separates being mentioned, being recommended and being cited.
Publication status: protocol and tools available; AI benchmark results pending.
We have published a frozen prompt panel and an analysis tool. They contain no measured chatbot results. The public website audits below are actual observations of our own pages.
Readiness baseline
Three timestamped HowAICite page audits, with every finding and a raw JSON download.
Inspect the sample reports →Metric dictionary
Understand what a score counts, what its denominator means and where human review is needed.
Read the definitions →Our research agenda
| Study | Question and design | Status |
|---|---|---|
| Recommendation–Citation Gap | How often is a brand recommended without a source link, or cited without being recommended? | Protocol published |
| AI Visibility Index | Repeat a fixed commercial prompt panel and publish mentions, recommendations and citations separately. | Planned |
| Citation Source Concentration | Measure how strongly answers depend on a small set of source domains. | Planned |
| Answer Volatility | Repeat identical questions to distinguish day-to-day variation from a persistent change. | Planned |
| Country and Language Parity | Compare equivalent, native-reviewed questions across country and language cohorts. | Planned |
| Search Rank and AI Citations | Compare observed organic result URLs with cited URLs for the same question panel. | Planned |
| Citation Freshness Lag | Observe how long controlled, timestamped page updates take to appear in answers. | Planned |
| Third-Party Source Graph | Map which review, media and community sources co-occur with recommendations. | Planned |
| Citation Persistence | Track repeated source URLs over time while preserving model and prompt changes. | Planned |
| GEO Intervention Replication | Pre-register content improvements and compare control pages; publish null findings too. | Planned |
Publication standard
Each empirical release must include its panel version, collection dates, surfaces, requested and observed location, model where available, failures, missing observations, annotation process, denominators and limitations. Raw customer evidence is private by default. Public research has a separate review and publication boundary.
HowAICite sells software in the category it studies. That commercial interest is disclosed in every study. We do not select only the answers in which our brand appears or use our own readiness score as evidence that an AI engine recommends us.
Start with your own evidence.
Inspect a website, save an observed answer, and decide what to improve.