Recommendations Are Not Citations: What 19,588 AI Answers Show

Why the Gap Runs One Way

Citation counts became the standard proxy for AI visibility for a practical reason: they are the countable half. Sources arrive as a list, a list can be counted, and the number goes in a slide. Whether the answer actually recommended you is messier to establish, so the industry reached for the easier measurement and treated it as a stand-in for the harder one.

The stand-in is not dishonest. It is just narrow. Everything it reports is true, and it reports roughly a third of what happened. A team optimising against it is working on a real signal while most of their actual exposure sits outside the frame, invisible to the dashboard and unattributed to any of the work that produced it.

There is a louder failure people worry about instead, where the citation number climbs while the answer recommends someone else. It is real, and it is the rarest thing in this dataset at 1.0% of responses. Worth watching if it lands on your commercial prompts, but it is not the reason your citation count and your visibility disagree. The undercount is.

What We Measured

We took every tracked response that arrived with at least one source attached, which is the only population where the question is even answerable, and sorted each one four ways.

Most responses sit in the first bar, because most tracked prompts are unbranded questions where a given brand simply does not come up. The interesting comparison is between the next two. Being named with no supporting citation happened 4,232 times. The tidy case, named and cited together, happened 1,963 times. The reverse case, where a cited page mentioned the brand but the answer never did, is rare at 1.0%.

The cleanest way to state it is as a test. Treat "was cited" as a test for "was recommended" and it scores 90.6% precision against 31.7% recall. When the test fires it is nearly always right. It just fails to fire on two out of every three recommendations.

The precision is not the surprise once you look at the mechanism. A cited page that mentions the brand has already put the name into the model's context, a few tokens from where the answer is being written, so of course it tends to come out in the answer. The recall is the finding.

It Holds Across Every Engine We Could Measure

A single-engine finding is usually an artefact of one product's interface. This one is not.

Four engines with genuinely different architectures land between 60% and 68% named-without-citation. The detail worth sitting with is Perplexity at 66.2%. It is the engine built around showing its sources, the one people reach for when they want receipts, and it still names brands without a supporting citation about two thirds of the time it names one at all.

ChatGPT is the outlier in the other column, with effectively zero responses where a cited page mentioned the brand but the answer did not. When it pulls a brand into the citation set, it names it. Claude sits a little lower on the gap, though it runs on a smaller and higher-stature cohort, so treat that one as directional.

Run the precision-and-recall framing per engine and the shape holds everywhere. Precision stays above 85% on all four, reaching 100% on ChatGPT. Recall never climbs out of the low thirties. Different retrieval stacks, different citation interfaces, same answer: the source list is a reliable signal that catches very little.

What This Actually Means

The simplest reading is that recommendation is largely untethered from the source list. When an engine names a brand, it is usually drawing on what it already knows rather than on the specific page it happens to be showing you. The citations are doing a different job: they support the surrounding explanation, and they satisfy an interface convention that answers should look sourced.

This is the brand-level version of a pattern that has already turned up at the answer level. An analysis of AI Overviews found accuracy improving while the share of correct answers actually supported by their own citations fell, which we covered in our piece on AI attribution. A citation sitting under a sentence is not proof the sentence came from it. Our data says the same thing about brand names.

Read Your Own Split

The reason to separate the two measures is that the shape of the gap tells you which problem you have. Run the same four buckets for your own brand and the diagnosis mostly falls out.

Mostly named, rarely cited

The common case, and on the surface a good one: engines recommend you from what they already know. The exposure is that none of it rests on anything you control. If the model's impression of your category shifts, or a competitor starts showing up in the pages engines read, you have no position in the evidence base to defend. This is the profile where earning third-party coverage does the most work, because you are trying to convert reputation into something referenceable.

Cited but rarely named

The most uncomfortable pattern, and the one that produces the screenshot everyone circulates. Your page is good enough to be used as evidence, and the answer it supports recommends somebody else. You are functioning as a reference work in a market where a competitor is the answer. Rare overall at 1.0% of responses, but if it is concentrated on your commercial prompts, that is a positioning problem rather than a content problem: the content is landing, the recommendation is not.

Neither

Most responses, and for most prompts that is simply correct. It only becomes a signal when it clusters on the questions your buyers actually ask. Absence there is a visibility problem, and no amount of citation work touches it until the engine has a reason to name you at all.

The point of the exercise is that these four states call for different work, and a single blended visibility score collapses them into one number that cannot tell you which one you are in.

Track Both, Separately

The fix is not to abandon citation tracking. It is to stop reporting one number as though it covers both jobs, and to be explicit about which question each one answers.

Mention rate answers "is the engine recommending us?" It is the number that maps to commercial outcomes, and it is the one to put in front of leadership. Citation rate answers "are our pages part of the evidence base?" It is a supply-side measure, useful for deciding what to publish and where to earn coverage, and it moves on a different clock. Put them side by side and the first thing you should see is the gap itself: the mention rate running close to three times the citation rate, which is the part of your visibility a citation-only report was never going to show you. Our guide to measuring share of voice covers how to define the denominator so the mention side holds up.

One practical warning. Because both measures move run to run, a single check will not tell you which way either is heading. We have written before about how much these systems vary between identical runs. Compare rates over repeated runs, not screenshots.

What This Does Not Mean

It does not mean citations are worthless. Roughly a third of recommendations did come with a supporting citation, the pages engines cite are how they learn what is true about a market, and only 2.9% of citations point at a brand's own domain, which is why earning third-party coverage remains the highest-leverage work available. Our off-site playbook covers that in order of leverage. The argument here is narrower. A citation count is a floor on your recommendation count, not a substitute for it.

It does not mean the citation number lies to you. This is the reading we expected going in and the data went the other way. At 90.6% precision, a citation that mentions your brand is close to a guarantee the answer named you. The problem is not that the signal misleads. It is that it is silent for two thirds of the events you care about.

It also is not a measurement of whether cited domains get named. Other people have looked at that question, which uses a different denominator and answers something else. We measured whether a page cited beside an answer mentions the brand being tracked. The two sets of numbers are not interchangeable and should not be stacked against each other.

The remaining limits, beyond the model-judged source flag noted at the top: the brands are Ranqo customers rather than a random sample of the web. Grok and Google AI Overviews produced too few qualifying responses to report, so the finding covers four engines rather than six. And this is one window in an ongoing tracking corpus, not a controlled experiment.

How Ranqo Helps

The reason we could run this analysis is that Ranqo stores both halves of every answer: whether the brand was named, where it placed, and separately the full list of sources the engine returned with a flag for whether each one mentions the brand. That is what lets you see the two rates side by side for your own brand rather than inferring one from the other.

Which is the actual recommendation of this post, and it costs nothing to adopt: put mention rate and citation rate next to each other in whatever you report, label which question each answers, and watch what happens when they diverge. That gap is usually the most informative thing on the page.