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 9,112 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.
9,112 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. In services and B2B, the only study with Spanish data (GA4 of 25 websites, October 2026) finds a median of ×1.87 measured on key events, not sales; among the sites that convert most, the difference can no longer be told apart from chance.
Review of the published evidence. Largest study: 973 sites and 164.8 million transactions. Since October 2026, also a Spanish study of 39 websites · 1 September 2026
-
06
We asked AI 54 times about stand building in Spain: 19 of the 20 association members never came up
Across 54 Perplexity answers to real buying questions about trade fair stand building in Spain, 24 of the 40 companies tracked (60%) never appear once. Among them, 19 of the 20 members of the sector association. The one member that does appear shows up in 87% of answers and leads the ranking, without a single block of structured data on its website.
54 Perplexity answers to 18 buying questions · 40 sector companies tracked across them · 63 websites analysed over HTTP · 7 September 2026
-
07
Google lets you opt out of its AI. On our panel, that switch would turn off 53% of mentions
Since 31 August 2026 Google offers a single-checkbox control to stop appearing in AI Overviews, AI Mode and generative AI in Discover. Across 17,710 measured answers from a closed panel of 11 brands, Google AI Mode accounts for 53% of every time an AI search engine names the brand (95% CI 50.9–55). Per brand the split ranges from 34% to 78%: for 7 of the 11, flipping that switch would turn off more than half of their AI presence.
17,710 measured answers over a closed panel of 11 brands · 9 September 2026
-
08
The same website scored 95 on our technical audit and 20 on Cloudflare's
On 10 September 2026 we ran three websites through our AI Readiness audit and through Cloudflare's agent readiness scanner. We gave them 95, 98 and 93; Cloudflare gave 20, 20 and 27. Most of the gap comes from measuring different layers: ours checks whether an AI search engine can read the page, and 8 of Cloudflare's 15 checks are about APIs, authentication and MCP servers. When we then audited the homepages of 23 client websites with the corrected engine, none had a single agent signal, 17 scored 70 or more, and 4 always rejected GPTBot, OpenAI's training crawler, while letting every search crawler through.
3 websites run through both tools · 23 client homepages audited with the corrected engine · 10 September 2026
-
09
Your Search Console has impressions no person made
Search Console counts as impressions searches no person made. On joseredondo.es, between 1 July and 20 September 2026, 28.5% of query impressions came from rows with no clicks below position 50 or shaped like a GEO tool prompt. Without them, the average position goes from 28.9 to 22.6. The most reliable signal is clicks: a robot does not click.
Two websites of our own, Search Console web search export: joseredondo.es (544 queries, 10,954 impressions) and vilnux.com (27 pages, 15 queries) · 23 September 2026
-
10
Is LinkedIn the second most cited domain in AI? In our data, it is ninth
Across 30,774 answers in Spanish measured between 26 June and 20 September 2026, linkedin.com is the 9th most cited domain, with 1,445 citations. YouTube is 1st, with 8,084. LinkedIn shows up in 7.5% of Perplexity answers and almost never in ChatGPT or Copilot. And only 20.1% of LinkedIn citations go to a post: the rest are company pages, articles and profiles.
30,774 answers from four engines about 34 brands, with 262,109 citations to websites other than the measured brand's own · 23 September 2026
-
11
An AI agent reached the contact page on 9 of 10 business websites and found the price on 1
On 2 October 2026 an AI agent carried out three tasks on 12 Spanish websites, three times each: find the price, the opening hours and the contact route. On the 10 client company websites it reached the contact route on 9 and found the price on 1. Opening hours, on 6. No website blocked it with a bot check: of the 48 failed attempts, in 40 the agent went through four pages without finding the information. And when it does not find it, it sometimes makes it up: on our own website it gave prices of €300, €225 and €790 that do not exist, until we put the prices on the homepage.
12 websites (10 client companies and 2 of our own) · 3 tasks × 3 attempts = 108 attempts · 2 October 2026
-
12
Perplexity has stopped citing Reddit: from 5.6% of its answers to 0.1% in nine weeks
Between weeks 33 and 41 of 2026 we measured the same 770 questions in Spanish about 9 brands every week on Perplexity. The share of answers citing reddit.com fell from 5.6% (43 of 770) to 0.1% (1 of 770). Sources per answer did not fall: they stayed around 18. The engine does not cite less; it has stopped citing Reddit. On Google AI Mode Reddit carried little weight and also falls, from 1.4% to 0.3%.
Perplexity: 9 brands and 770 fixed questions, 6,930 answers. Google AI Mode: 8 brands and 627 fixed questions · 13 October 2026
A weekly report generated from our own measurement: how Google AI Mode, Perplexity and ChatGPT moved on the same questions and brands, and why when we know. See this week’s report →
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 →Home · Methodology · Plans & pricing · X · Product Hunt