You are being sold the AI citation as the new link. Your brand shows up in the answer, a small link appears beside it, and there you are, visible. That is the implicit reasoning behind every GEO proposal: being cited means being seen, so it means traffic.
The reasoning rests on an equivalence nobody ever verified.
A citation is not a link. A link on a results page is an invitation to leave. A citation inside a generated answer is a guarantee placed beside an answer that is already complete. The reader has what he came for. He has no reason to click, and he does not click.
This article measures how far that goes, and draws the consequences for what you should be tracking.
The most uncomfortable number in the file
The Pew Research Center published in July 2025 an analysis covering nearly 70,000 Google searches performed by around 900 users over the month of March 2025. Provenance: an independent research institute, a real panel, published methodology. Google publicly disputes that methodology, which has to be said.
Three results.
With no generated summary displayed, a search produces an organic click in roughly 15 percent of cases. With a generated summary, that rate falls to roughly 8 percent. Half.
Links displayed inside the generated summary are clicked in roughly 1 percent of cases.
And the browsing session ends in 26 percent of cases when a summary is displayed, against 16 percent without.
One percent. That is the rate at which a user opens the source the engine judged reliable enough to cite.
A second measurement, on a completely different panel, converges. An SEO tool vendor published an analysis of 300,000 keywords, comparing click rates before and after the rollout of generated summaries in the United States. The click on the first organic result falls from 7.3 percent to 2.6 percent when a summary is present. Provenance: a vendor’s proprietary data, cross-referenced with Search Console, scope limited to informational queries.
The same vendor, measuring his own site, notes that visitors arriving from generative search click links 75 percent less often than those arriving from classic organic. That figure is self-reported on a single site: treat it as an indication, not as a norm.
What users actually do with displayed references
A paper published in March 2026 looked directly at the references themselves rather than at the click. It analyzes 1,517 references drawn from thirty question-answer pairs across nine conversational systems, scored against a standard documentary quality rubric and supplemented by a user study.
Provenance marking, and it matters here: thirty question-answer pairs is not many, and the authors themselves describe their user study as preliminary. Cite it as a converging indication, not as a demonstration.
Two results come out of it.
The quality and number of references vary enormously from one system to another. One system displays an average of 9.5 references per answer for a quality score of 15.48 out of 20; another displays 4.0 for a score of 11.65. That is not the same product, those are not the same guarantees, and the user has no way of knowing.
And above all: users rarely interact with the references. They function as a signal of seriousness, not as a point of departure.
That is exactly the role they play visually. A small numbered pill, at the end of a sentence, inside an answer that stands on its own. The citation reassures. It does not redirect.
What the reader loses, and why that concerns you
A study published in the journal PNAS Nexus on 30 September 2025 supplies the most solid element in this file, because it is a peer-reviewed journal publication, resting on seven experiments run online and in the laboratory with thousands of randomly assigned participants.
The protocol: ask participants to write advice on a topic, after giving them either a language model or classic web links.
The result: language model users develop shallower knowledge. Their advice is objectively shorter, contains fewer factual references, and looks more alike from one participant to the next.
That last point is the most interesting commercially. Homogenized answers mean that users of the same engine come away with the same representation of a market. If you are not in that representation, you do not exist in the heads of an entire buyer segment, and no missing click will ever flag it for you.
Citation against link: what you are actually buying
| Classic link | Citation in a generated answer | |
|---|---|---|
| Function for the reader | A starting point | A guarantee placed beside a complete answer |
| Measured open rate | ~15 percent organic click when no summary is shown | ~1 percent on links inside the summary |
| Effect on the session | Browsing continues | Session ends in 26 percent of cases |
| What you get | A visit | A mention, and an indirect effect that is hard to attribute |
| What your analytics sees | Everything | Almost nothing |
That last line traps everybody, in both directions.
What contradicts this, and should be read
Concluding that AI citations are worthless would be as lazy as concluding the opposite. A paper published in June 2026 supplies the most serious contradiction.
Its protocol is unusual and robust: a panel joining the browsing data of volunteer users to their real conversations with ChatGPT, Claude and Gemini. The same users on both sides. With an event study on prior trends, stance classification, conditioning on non-customers, within-answer controls, and matched backward placebos.
The result: when a brand is recommended to a user who did not previously know it, Google searches for its name rise by 4.3 points, within an interval of 3.1 to 5.5. Visits to the brand’s site rise by 2.4 points, and product pages at resellers by 1.0 point.
So the effect genuinely exists. But look at the route it takes: through a later Google search. Which is to say through a channel your analytics will attribute to branded organic, or to direct, never to AI.
Caveat for the file: observational design, preprint, and no transaction observed.
The honest conclusion is therefore not that nobody clicks so nothing happens. It is more uncomfortable than that: something is happening, and your measurement does not see it, in either direction. The people selling you measurable AI traffic are overstating. The people concluding it is useless from their dashboard are wrong too.
What you do tomorrow morning
Stop tracking referral traffic from AI platforms as your primary indicator. It is structurally tiny, it was tiny before you got here, and it will stay tiny.
Track three things instead.
Branded search, in Search Console, on queries containing your name. That is where the effect measured by the panel shows up, with a lag.
The unlinked mention, meaning answers where your brand is named without being cited as a source. It is a distinct signal, and often the most frequent one.
And the share of answers where a competitor is named in your place on your category queries. That is the only metric in this set with a direct commercial consequence.
A citation that one reader in a hundred opens is not an acquisition channel. It is a guarantee. That has value, but it is not the value you are being billed for.
Sources
- Pew Research Center (2025). Google users are less likely to click on links when an AI summary appears in the results
- Ahrefs (2025). AI Overviews Reduce Clicks by 34.5%
- Ahrefs (2025). AI search traffic and conversions
- Ouyang, J. & Narechania, A. (2026). Analyzing the Presentation, Content, and Utilization of References in LLM-powered Conversational AI Systems, arXiv:2604.15326
- Melumad, S. & Yun, J. H. (2025). Experimental evidence of the effects of large language models versus web search on depth of learning, PNAS Nexus 4(10)
- Iannelli, M. & Ai, A. (2026). From Prompt to Purchase: How AI Brand Recommendations Move Consumers on the Open Web, arXiv:2606.10907
<strong>LaFactory</strong> measures AI visibility with a published protocol: repeated measurements, paraphrases, control group. No guaranteed placement, ever. Contact us to scope an audit.