Blog · Measurement

Citation share vs mention share

Published 14 September 2026 · Updated 8 October 2026 · 9 min read

Citation share vs mention share
Photograph: Myles Bloomfield / unsplash

The short answer: Citation share is your slice of the links an engine lists as sources. Mention share is how often your brand name turns up in the written answer. One is a retrieval event, the other is a generation event, and a brand can win either without the other. Tracked together they tell you which half of the machine is failing you. Tracked as a single blended score they tell you nothing you can act on.

Most reporting collapses the two because both feel like presence. They are produced at different moments by different parts of the system, and they have different denominators, which is where the arithmetic quietly goes wrong.

Why does citation share matter

Citation share matters because it is the half of the pair you can act on directly, and because acting on it moves the other half. Retrieval decides whether a document of yours is in front of the model at all. What the answer then says about you is written from what was retrieved, plus whatever the weights already held.

The September 2026 preprint set out below puts a figure on that link. Brands with no retrieval presence were named in roughly 2.8 percent of GPT answers and 3.8 percent of Gemini answers. When the brand's own domain appeared in the retrieved sources, those rates rose to about 49.0 percent and 58.4 percent. So citation share is the leading indicator and mention share is the outcome you sell on, which is why it earns a column of its own rather than being blended into a single visibility score.

It matters for a second, duller reason. A citation is a URL in a list, so it leaves a trace you can re-read and argue with months later. A mention is a sentence somebody had to capture at the time, and if nobody captured it, it is gone.

Two different events, two different denominators

A citation happens when the retrieval layer pulls a document into the context the model writes from, and the interface attributes it. Your URL was fetched, chunked, ranked and kept. Citation share is usually computed over source slots: of all the URLs cited across your prompt set, what fraction point at your domain.

A mention happens later, inside the generated text. The model names you. It can do that from a retrieved page, from somebody else's retrieved page that happens to list you, or from what it absorbed in training with no live document involved at all. Mention share is normally computed over answers: of the answers in your prompt set, in how many does your name appear.

Those denominators behave differently when you change the prompt set. Add fifteen prompts where nobody gets cited much and your mention share barely moves while your citation share shifts, because the source pool grew. This is the same denominator problem that makes a single blended visibility number so slippery, and it is covered in more detail in Share of voice in AI answers, and its limits.

There is a second split inside citation share that most teams skip. Citations to your own domain and citations to pages about you are not the same asset. The first you control. The second is corroboration, and it is the thing that keeps working when your own page is outranked for a given passage.

The direction of travel, with a number attached

The useful finding is that the two are not independent. A September 2026 preprint fitted a stage model of brand visibility across 34,960 unbranded prompt and engine observations, covering 2,854 distinct prompts on GPT and Gemini between June and September 2026. With no retrieval presence at all, brands were named in roughly 2.8 percent of GPT answers and 3.8 percent of Gemini answers. When the brand's own domain appeared in the retrieved sources, those mention rates rose to about 49.0 percent and 58.4 percent. When own-domain retrieval and branded follow-up queries were both present, the figures reached 91.4 percent and 100 percent in that sample.

Read that as a conditional probability, not a promise. Getting cited makes being mentioned far more likely in the same answer. It does not make it certain, and the reverse path is weak: plenty of brands get named with no document of theirs anywhere in the source list.

That asymmetry is the practical reason to treat citation share as the leading indicator and mention share as the outcome you actually sell on. It also explains why original research pays twice. A page with a number nobody else has gets cited for the number, and the citation drags the name into the prose with it, which is the argument in Original data gives an engine something only you have.

Four cells, and what each one means

Record two booleans per run: was my name in the answer text, and was my domain in the source list. That gives a two by two, and each cell has a different fix.

Named and cited. Working as intended. Check the sentence is true and move on.

Cited but not named. Your page was retrieved and used, and the model wrote the claim without attaching you to it. This is almost always a passage problem. The useful sentence on your page says "the median turnaround is seven working days" with the brand name three paragraphs north, so the liftable chunk carries no identity. Put the subject in the sentence that answers the question.

Named but not cited. The model knows you from training or from third-party pages. This is real visibility and it is fragile, because it survives on weights and other people's documents rather than anything you can edit. Model updates move it without warning.

Neither. Baseline. Expect single-digit percentages until something changes upstream.

The cell that most reports hide is the second one. A blended score counts it as a partial win. In practice it is the cheapest fix on the list and it is usually sitting in plain sight, which is why the self-check starts with reading your own page the way a chunker would.

Why your analytics cannot see mention share

Search Console now reports AI Overviews and AI Mode inside the performance report. A click on an external link inside an AI Overview counts as a click, and every link in that overview is assigned the same single position, since the overview occupies one slot rather than several. Those are citation metrics. An answer that names your brand in prose and links to nobody produces no impression you will ever see.

So the instrument is blind in exactly the half that buyers read. Anyone quoting Search Console as evidence of AI visibility is quoting citation share and calling it presence. The only way to get mention share is to run the prompts and read the text, which is the method argued for in How to measure AI visibility without fooling yourself.

The traffic side is worth knowing too. In a Pew Research Center analysis of browsing data from 900 United States adults covering 68,879 Google searches in March 2025, users who saw an AI summary clicked a result in 8 percent of visits against 15 percent without one, and clicked a link inside the summary in about 1 percent of visits. Of the summaries examined, 88 percent cited three or more sources. Citation share is therefore computed across a crowded slate, and the citation itself buys very little direct traffic. Its value is mostly the mention it tends to bring with it, plus the brand effect of appearing beside the answer.

That economics is the whole argument publishers are having right now, and it is unpacked in News sites, publishers and the citation economy.

How to count without fooling yourself

Three rules keep the two metrics honest.

Count coverage, not raw occurrences. An answer that says your name four times is one mention, not four. Count-based metrics reward verbose answers and punish concise ones, and they are the first thing to drift when an engine changes its writing style.

Freeze the denominator and version it. Write down the prompt list, the competitor list and the engine list, give that bundle a version number, and when you change it, say so on the chart. Any comparison across versions is not a comparison.

Run enough times to have an interval. At twenty runs per prompt, a measured rate of 40 percent carries a confidence interval wide enough to swallow most of the movement people report as a win. Repeat runs are the only defence, since the same prompt asked twice can retrieve different documents.

If you want to see the two reported side by side with evidence attached to each finding, the sample audit report shows the format we use: the prompt, the run, the answer text, and the source list, so a number can be argued with rather than believed.

What this framing does not settle

No engine publishes mention data, so every mention-share figure in the category is built from sampled runs by somebody outside the system. Different samplers get different numbers for the same brand, and none of them is the ground truth.

We also cannot cleanly separate a mention sourced from training weights from one sourced from a retrieved third-party page, unless the page happens to be cited. Running the same prompt with retrieval disabled gets closer, and it is not the same question.

And the link from either metric to revenue stays unproven at our scale. We can show that citation moves mention. We cannot yet show, with evidence we would stand behind, how much either moves a pipeline.

FAQ

Why does citation share matter? Because it is the half of the pair you can act on directly, and because it raises the odds of being named in the same answer. In the September 2026 preprint cited below, brands with no retrieval presence were named in roughly 2.8 percent of GPT answers and 3.8 percent of Gemini answers; with their own domain in the retrieved sources, about 49.0 percent and 58.4 percent. It also leaves an auditable trace, since a citation is a URL you can re-read later.

Which should I optimise first? Citation share, in most cases, because it is the one you can act on directly and because it raises mention probability in the same answer. Mention share without citation is real but it is harder to move and harder to hold.

Can mention share be high while citation share is near zero? Yes, and it is common for well-known consumer brands. The model names them from training and from third-party coverage. It is pleasant until a model update changes what the weights favour, at which point there is nothing of yours in the source list to fall back on.

Should third-party citations count in my citation share? Track them in a separate column. A citation to a review site that names you is doing work for your mention share, but it is not an asset you can edit, so mixing it with own-domain citations hides which lever you are pulling.

How many prompts do I need before the numbers mean anything? Enough prompts to cover the buying questions, run enough times each to put an interval around the rate. A short prompt list run many times beats a long list run once, because the variance between runs is larger than most people expect.

Sources

Related reading

How to measure AI visibility without fooling yourselfShare of voice in AI answers, and its limitsOriginal data gives an engine something only you haveNews sites, publishers and the citation economyHow many runs per prompt before a number means anythingReporting AI search to the people who fund it