How to Attribute Value When AI Citations Don't Get Clicked

The uncomfortable arithmetic of AI citations: almost nobody clicks them, and they still may matter. This page is the attribution model we actually use — 3 measured surfaces and a dated ledger — built to survive the fact that the citation click is nearly extinct as a behavior, and built without the invented multipliers this niche loves.

How rare are citation clicks, really?

Rare enough to stop modeling them as traffic. Pew Research's analysis of March 2025 browsing data found users clicked a link inside an AI summary on roughly 1% of visits where a summary appeared — and clicked any traditional result on just 8% of those visits, versus 15% when no summary was present (Pew Research Center, July 2025).

Yet the same evidence base refuses to write citations off. Seer Interactive found being cited in an AI Overview delivered 120% more organic clicks per impression than not being cited, yet still 38% below queries with no Overview (53 brands, January 2025–February 2026; Seer Interactive, 2026). That advantage sits inside a shrunken click pool: Seer's earlier study saw organic CTR on Overview queries fall from 1.41% to 0.64% between January 2024 and January 2025 (Seer Interactive, 2025). Pew's 1% and Seer's 120% measure different events: Pew counts user clicks on the summary's own links; Seer compares a brand's organic clicks per impression when cited versus not cited.

A citation is a weak pipe and a real marker — the full reconciliation is in are AI citations worth anything. Attribution's job is booking the marker's value without pretending the pipe exists.

What are the 3 surfaces you can actually measure?

Direct clicks, branded search, and the answers themselves — each with a free instrument and a monthly reading:

The first surface is the smallest and the easiest: a channel group that isolates assistant referrers, set up once in GA4. Its numbers will be humble — ours are [our data] — but they are measured, and a measured floor beats an estimated ceiling in any model you have to defend.

The third surface is where unclicked value shows up most directly: a frozen battery of buyer questions run monthly across the engines, scored cited / mentioned / absent / wrong, per the free share-of-voice method. Share of voice trending up while sessions stay flat is not a failed channel — it is the expected shape of a channel where 99% of AI-summary visits produce no click on the summary's links (Pew, March 2025).

For Microsoft Copilot and Bing's AI summaries, the engine now reports part of that third surface itself. The AI Performance report in Bing Webmaster Tools shows how often your pages were cited in those answers and in select partner AI integrations, along with the grounding queries tied to those citations (Bing Webmaster Tools). Microsoft says grounding queries are grouped phrases, not full user questions or prompts. Microsoft describes the data as a sample and says a citation does not represent clicks or traffic; the report also shows no individual answers or prompts, so log its citation total monthly beside the battery scores rather than in place of them.

How does branded search capture unclicked citations?

It catches the deferred arrival. A buyer who reads an answer citing your brand rarely clicks the little link — but the exposure happened, and when that buyer acts later, the common path is typing your name into a search box. That behavior lands in surfaces you already measure: branded-query impressions and clicks in Search Console, and direct traffic.

The honest protocol is baseline-and-trend, not formula. Record your branded-query impressions and clicks now, dated. Watch the series monthly alongside the share-of-voice trend. If citations are doing brand work, branded demand is where it should eventually show; if a quarter of rising share of voice produces no branded movement, that is evidence the exposures are not converting to memory — a real finding either way.

What we refuse to do is convert the correlation into a coefficient. Assigning, say, 30% of branded-search growth to AI citations would be an invented number wearing a methodology, and this niche has enough of those. Branded search enters our model as a monitored surface with a dated baseline — direction and timing, not decimals.

Why not just use a conversion multiplier?

Because the famous one describes the winners, and the borrowed version describes nothing. Semrush reports AI-referred visitors converting 4.4x better than average (Semrush) — and that figure never travels without its caveat: sites measuring meaningful AI referral volume are, by definition, sites already visible in AI answers. The multiplier is a portrait of survivors, not a property of the channel.

Our own position is stricter than the caveat. We run a dedicated AI-referral channel in production, and we publish no conversion-quality claims from it, because the channel is weeks old and days of separated data cannot carry a finding [our data]. When it has months of history, the number ships whichever direction it points. Until you have measured your own, your model should contain no conversion multiplier at all — an estimated figure is an invented figure.

How do you assemble the model for a quarter?

As a dated ledger read quarterly, not a formula computed monthly. Month 0: baseline all 3 surfaces — channel sessions and conversions, branded impressions and clicks, battery scores and, for Microsoft's surfaces, the Bing AI Performance citation total — with dates. Each month after: append the readings, note ships and changes (a new page, a citation observed) in the same ledger, and resist conclusions until 3 readings exist.

At day 90, the ledger supports honest sentences of the form: share of voice rose from X to Y prompts; branded impressions moved Z% over the same period; the AI channel produced N sessions and M conversions [your data, your dates]. Those sentences — not a blended "AI ROI" figure — are what plug into budget decisions, and the cost side they weigh against is worked through in our GEO ROI model.

Against interest: if the quarter shows near-zero on all 3 surfaces, believe it. Many businesses' buyers do not ask assistants for recommendations yet, and the correct attribution finding is "not material — recheck in 2 quarters," not a softer multiplier. The measurement costs almost nothing precisely so you can afford that answer. What you would be measuring — pages structured to be worth citing at all — is the subject of our generative engine optimization guide.