Recommendation–Citation Gap: research protocol

Published September 9, 2026 · HowAICite · Protocol author: Thanh Dan

Research question

Within a fixed sample of GEO tool questions, how often is HowAICite mentioned, positively recommended or cited as a linked source? How often do the recommendation and citation outcomes diverge? This initial self-study validates the collection process; it is not an unbiased league table of vendors.

Collection plan

1. Freeze this panel before collection. Start with US / English on ChatGPT Search, Gemini, Perplexity and Google AI Mode.
2. Collect three independent attempts per question and surface per wave. Use fresh sessions; document personalization and local settings.
3. Record exact prompt, full answer, URLs, timestamp, model if shown and requested/observed location. Preserve unsuccessful attempts.
4. If no AI answer is shown, record no-answer. Record provider errors as failed. Neither is a negative brand observation.
5. Keep API observations in their own panel. A web-search API result is not a consumer ChatGPT answer.
6. Repeat the same panel weekly. Any change to questions, markets or surfaces produces a new panel version.

Annotation rules

A mention identifies the target brand. A recommendation positively selects it for the user's use case and needs an exact supporting excerpt. A citation attributes a claim to a source on the target domain, including subdomains. Review whether the link supports the claim; do not count every pasted URL as a citation.

Use null for an unreviewed label. Double-review a random 20% of completed answers, document disagreement and resolve labels before publication. A reviewer must not infer recommendation from name frequency.

How the calculation works

Each rate uses fully reviewed, completed answers in the same cohort as its denominator. The tool reports failed, no-answer, unreviewed and attempted counts separately. It splits branded/unbranded prompts, collection method, model, wave and location provenance. Missing planned cells remain visible in wave coverage.

Recommendation without citation, citation without recommendation, and both are separate counts. An empty dataset produces not-collected, not a 0% market visibility claim.

node metrics.mjs prompt-panel.json observations.json

The frozen panel

Limitations and conflicts

HowAICite is both the publisher and the brand being evaluated. This is an editorial convenience sample, not the distribution of real user demand. Repeated responses may be correlated. Changing models and retrieval systems affect results. The analysis does not estimate a population recommendation probability, provide causal attribution, or prove that a website change caused an outcome.

Source guidance

Google's AI features guidance documents search eligibility without guaranteeing inclusion. Perplexity's crawler documentation describes access controls. Neither establishes that this study's proposed optimizations cause recommendations.

Cite this protocol

HowAICite. Recommendation–Citation Gap Protocol, version 1.0.0. September 9, 2026. howaicite.com/research/recommendation-citation-gap.

Panel and protocol: CC BY 4.0. Analysis code: MIT. Third-party answer content retains its own rights. No external peer review or research repository registration is claimed.