Citation share measures how often a domain appears as an attributed source in AI-generated answers. It is a proxy for influence over those answers, not a measurement of it.
The gap between the two is structural rather than a matter of precision. A model can reproduce a community's framing, vocabulary, and consensus judgment about a product without linking to the place it learned them, and most answers are produced this way. Citation trackers typically count only responses that cite at least one source, which means answers drawn from a model's priors are excluded from the denominator before measurement begins. An instrument built on attribution therefore cannot observe unattributed influence — not because it measures it poorly, but because that influence never enters the sample.
Cross-vendor comparison compounds the problem, because published figures rarely share a denominator. A study reporting that 📝Reddit appeared in 40.1 percent of AI responses is measuring the share of responses that reference a domain. A tracker reporting Reddit at under four percent is measuring that domain's share of all citations across responses that cite anything. Both can be accurate and neither can be plotted against the other; arranging them into a single declining trendline produces a dramatic figure that describes nothing real. 📝Semrush and independent citation trackers have both been quoted this way in the same paragraph by writers who did not check.
📝Reddit's ChatGPT Citation Collapse (August 2026) is a working example of the discipline being necessary: a dramatic reported decline that the publishing tracker itself flagged as provisional pending verification.
None of this makes citation share worthless. It is the best available public signal for visibility, it is genuinely actionable for anyone whose goal is attribution, and a sharp move in it is worth investigating. The discipline is to state the basis before quoting the number, to refuse comparisons across incompatible bases, and to treat a decline as evidence about visibility until something measuring the other layers of 📝Learning Source vs. Trust Source vs. Culture Source says otherwise.