Why citation counts aren't enough

Spencer Thursfield · Published 2026-08 · Evidence: emerging research + our protocol

The seductive metric in AI SEO is the citation count: how many times did an engine list your site? You can gather it in an afternoon and chart it in a minute, and it collapses five different outcomes into one number that answers none of the questions that matter.

The pipeline a count conceals

1. Retrieved. Your page was fetched as a candidate. The preconditions live here.
2. Cited. Your page was listed as a source. The count sees this stage and no other.
3. Used. The answer drew on your material. Research increasingly finds citation and use diverge: engines list pages decoratively, and pages shape answers without prominence. Emerging
4. Represented fairly. The answer said something true about you. A citation that misstates your prices does more damage than absence, and large-scale audits have found a meaningful share of AI-answer claims unsupported by their cited pages. Emerging
5. Worth something. A human enquired, visited, remembered, bought. This is the stage with commercial value.

Two further reasons counts mislead

Volatility. Citation sets change between identical runs. A count from one collection day reflects that day's runs, and the next day's can differ. Repetition and variance belong inside the metric, so our protocol runs every prompt repeatedly and reports stability alongside rates.

Survivor bias. A count says nothing about the pages that went uncited, and the uncited pages carry the information. Measuring cited and uncited pages side by side turns tallies into evidence; that comparison group forms the spine of our study design.

What to report instead

Report rates with denominators, stability across repeats, use and fidelity separately from listing, and outcomes in visits and enquiries. Several small honest numbers carry more information than one large hollow one. Any tool selling the large hollow one sits in the ledger, entry 8.