Open research on how AI search engines cite
We publish what we find while measuring, with the sample and the method up front — and with what the data does not allow us to claim.
Every finding states four things before it states anything else: what was measured, on what sample, when and with what method, and what it does NOT prove. If a number cannot come with those four, it is not published here.
The corpus
Every first-party figure comes from the same place: the answers this engine has measured and stored in full, each with its date and its method stamp.
Findings
-
01
Google AI Mode cited 67% fewer sources, and 11 brands lost a third of their visibility in the same week
Between 24 and 31 August 2026, Google AI Mode went from citing 12.22 sources per answer to 4.08 across the same 990 questions, and its answers got 32% shorter. Share of Answer for the 11 brands measured fell from 23.4% to 15.9% in that same week.
4,852 Google AI Mode answers · 990 questions per week · 11 brands across 10 sectors · 1 September 2026
-
02
Same engine, same question, three times in a row: in 12 of 27 cases it did not give the same answer
Of 269 questions asked three times in a row to the same engine, the brand appeared at least once in 27. In 12 of those 27 it did not appear all three times: 44.4%, with a 95% confidence interval between 27.6% and 62.7%.
269 questions asked three times in a row to the same engine · 27 of them with the brand present at least once · 1 September 2026
-
03
Two AI search engines answering the same question share between 4% and 13% of their sources
Across 9,124 pairs of answers to the same question, the overlap of cited domains between two engines ranges from 3.9% to 13.4%. In 6 out of 10 ChatGPT–Perplexity comparisons there is not one single source in common.
9,124 pairs of answers to the same question, compared across engines · 1 September 2026
-
04
Our first accuracy measurement returned 99.5%. It was false
Of the 107 checks producing that 99.5%, 106 could not fail by construction. Once every claim had to be contradictable by the answer, the real figure came out at 64.2% over 53 checks.
107 initial checks · 53 after requiring every claim to be falsifiable · 1 September 2026
-
05
"ChatGPT traffic converts nine times better" does not hold up
The published evidence puts conversion from ChatGPT traffic between 13% worse and 31% better than organic traffic — roughly at par. The largest peer-reviewed study — 973 sites and 164.8 million transactions — is the one finding 13% worse.
Review of the published evidence. Largest study: 973 sites and 164.8 million transactions · 1 September 2026
How this can be checked
Every answer is stored with its full text and a stamp declaring the transport, the model, whether web search was on and how many repetitions were run. That is what makes it possible to state that a series is comparable with itself — and to detect when it stops being so. The full methodology is published separately.
Read the measurement methodology →
Frequently asked questions
What is vilnux Research?
It is where we publish the findings that come out of measuring brand visibility in AI search engines. Each one states what was measured, on what sample, with what method, and what it does not allow us to claim.
Can these data be cited?
Yes. Every finding carries its date, its sample and its method, which is what you need to cite it properly. We appreciate a link to the specific finding page rather than to the section index.
How often are new findings published?
There is no schedule. We publish when the weekly measurement produces something that holds up on a sufficient sample; forcing a cadence would mean publishing weak findings, which is the opposite of what this section is for.
Measure your brand with this same method
The free audit uses the engine exactly as it is today, with its sample and its interval. No signup, no card.
Run my free audit →