Measured citation rate across ChatGPT, Perplexity and Claude

Ask ChatGPT the same buying question twice and you will often get two different source lists. Across our runs, a brand that appears in the first answer reappears in roughly half of the repeats of the identical prompt. A screenshot therefore proves nothing, and a board slide built on one is worse than no slide at all. We measure appearance as a rate: every prompt is run thirty times per assistant, and every number we hand you carries the run count it came from. What gets measured: Citation rate, meaning the share of runs in which your brand appears at all. Position inside the answer, because third in a list of five is not a recommendation. The character of the mention, sorted into recommended, listed without comment, or named with a caveat, since the third category is usually the finding that changes what a company does next. Share against the competitors you name at kickoff, run on the same prompts on the same days so the comparison holds. And the source ledger: the actual URLs each assistant retrieved. Retrieval against memory: This distinction decides whether your problem is fixable this quarter. Perplexity retrieves and cites for nearly every answer. ChatGPT sometimes retrieves and sometimes answers from what the model already holds. Claude with search behaves differently again. When an assistant names your competitor without retrieving anything, that competitor sits in training data and no page you publish this month will dislodge it; that work is slower and mostly not a content problem. When the assistant does retrieve and still picks somebody else, the fix is usually one specific document on one specific third-party property. We label every prompt as one or the other, because the recommendations that follow have nothing in common. Where the prompts come from: Not from keyword tools. We sit with two of your salespeople and take the questions buyers actually ask, including the hostile ones: the pricing objection, the security review question, the is-this-category-dead question. Assistants field those far more often than they field your head term. We also run the prompts a competitor's champion would ask, phrased in their favour, because that is the conversation you are absent from and the one that reaches your buyer's internal channel. The finding you will not enjoy: Frequently the reason assistants do not cite you is that a single review site or comparison directory owns the retrieval slot for your category, and the only near-term fix is a presence on that property. We do not sell that placement, we take no fee from those sites, and we will name the property and what it charges publicly. Sometimes the honest conclusion is that your category is answered from training data by three incumbents and nothing you ship in six months moves it. We have written that conclusion twice this year. Both clients moved the budget to paid and partner channels, which was the correct outcome and not a good one for us. What lands on your desk: A baseline with citation rate, confidence interval and run count for every prompt and assistant. The full run log as timestamped raw text, so your team can rerun any prompt and check our arithmetic. A source ledger ranked by how often each domain was retrieved across the whole set. A written recommendation split three ways: what your own pages can fix, what needs a third party, and what is not addressable on this time horizon. A two-hour session with your growth and product marketing leads. Out of scope: No content production, no outreach to publishers on your behalf, no directory submissions, and no promise about assistant behaviour, which changes with model releases nobody here controls. We will not report a citation rate from fewer than thirty runs, even if you ask for a faster read, because below that the number moves more than the effect you are trying to see. Who should not buy this: Companies where assistant visibility is not yet the growth constraint, which in practice means most businesses under roughly two million in annual revenue. Anyone who needs the baseline to come out well; it lands where it lands and we report it unedited. And anyone expecting a lasting answer from a single read. Measure once and you have described one fortnight of a system that shifts with every model release. If that is genuinely all you want, take the baseline and stop, but price it internally as a snapshot rather than a strategy.