X-RAY · 18 SEPTEMBER 2026
Eight AIs signed in, in conversation: which services each one looks up when you ask it something
The same eight AIs as the clean-PC x-ray, this time in my own browser, already signed in, one by one, with one short question typed into each. Signed in and in conversation, ChatGPT makes the machine talk to no third-party service at all; Gemini loads six, and all six are on advertising and tracking lists.
How to read this. GUARDIANA writes down every name the machine asks to resolve (DNS). It sees names, not content: neither the question nor the answer. Here each AI was opened alone in a fresh tab of my own browser, already signed in, and one short question was typed into it. There is no view from inside the browser, because Chrome does not allow one on a person's own profile: what is counted is the names the PC asked for during each AI's minute, with the name cache emptied before each one and Windows' own names and the browser's own account traffic left out. The signed-out landing page, measured from the inside, is in the clean-PC x-ray.
Scoreboard
How many services that are not that AI its page makes the machine ask for when it is opened with the session already signed in and answers one question, one minute per AI. Beside it, how many of those are known for advertising or tracking, according to the public lists that ad blockers use. Fewer, which is why the first figure is the one that counts.
| AI | Services besides the AI | Of those, advertising or tracking | Its own domains | Distinct names the PC asked for in its minute | The fewest |
|---|---|---|---|---|---|
| ChatGPT | 0 | 0 | 4 | 4 | yes |
| Gemini | 6 | 6 | 7 | 13 | |
| Perplexity | 7 | 1 | 8 | 15 | |
| Meta AI | 0 | 0 | 2 | 2 | yes |
| Copilot | 4 | 1 | 4 | 8 | |
| DeepSeek | 6 | 0 | 2 | 8 | |
| Grok | 15 | 7 | 2 | 17 | |
| Mistral | not measured (see its section) | — | |||
Tied on 0, the fewest third-party services: ChatGPT, Meta AI, against 15 for the one loading most. Not being on those lists does not mean it does not track: it means the lists do not know that name. This is a result of these pages on this day, not a general verdict.
What it says
Signed in and actually in conversation, ChatGPT made the machine talk to no third-party service: it asked only for its own four names. That is a big change from the signed-out landing page, where it did not even load. Gemini, on the other hand, loads six third-party services while you talk to it, and all six are on the open lists: DoubleClick three times over, Google Analytics, Google Ads and Tag Manager. Perplexity loads seven, one of them listed; the rest are product measurement and its images. Of the ones that loaded with the account but whose answer could not be confirmed from outside, Grok talks by far the most: fifteen third-party names, seven of them listed, including Twitter's advertising, DoubleClick and Google Analytics; DeepSeek loads six and none of them is on any list, among them ByteDance's measurement and a device-identification service. Meta AI asked for no third-party name in its minute. Mistral does not count: its session was not signed in and what was measured was its login screen.
ChatGPT
Signed in and answering, not one third-party name: only its own. chatgpt.com and ws.chatgpt.com (which carries the conversation), cdn.openai.com and api.oaistatsig.com, its feature measurement. On the signed-out landing page it did not even load in an externally driven browser.
Gemini
Signed in and answering, six third-party services and all six are on the open lists of advertising and tracking: DoubleClick three times over (ad, googleads and stats), Google Analytics, Google Ads and Google Tag Manager. It is the only one of the three confirmed that loads advertising and audience measurement while you talk to it. The rest of the names it asks for are Google's and count here as its own.
Third parties in this run: doubleclick.net, google-analytics.com, googleadservices.com, googletagmanager.com.
Perplexity
Signed in and answering, seven third parties, one of them listed: Datadog (browser logging). The other six are product measurement and A/B testing (Eppo, Singular), a content delivery network and three Google image names. It had to be repeated: on the first pass the address was mistyped and it never loaded.
Third parties in this run: browser-intake-datadoghq.com, cloudfront.net, eppo.cloud, singular.net.
Meta AI
It loaded with the account and asked for no third-party name in its minute: only www.meta.ai and gateway.meta.ai. It could not be confirmed from outside that it answered, because its window does not change title.
Copilot
It loaded with the account: four third parties, one listed. Sprig (in-product surveys, two names), Sentry (error reports) and browser.events.data.microsoft.com, Microsoft's telemetry endpoint, which the lists mark as telemetry. It could not be confirmed from outside that it answered.
Third parties in this run: sprig.com, microsoft.com, sentry.io.
DeepSeek
It loaded with the account: six third parties and none on the lists. ByteDance's Volcano Engine (apmplus.volces.com, gator.volces.com and its CDN, performance and usage measurement), portal101.cn (device identification and risk), the AWS firewall and Cloudflare's check. That they are not on the lists is exactly what a list cannot do for you. It could not be confirmed from outside that it answered.
Third parties in this run: awswaf.com, volccdn.com, volces.com, cloudflare.com.
Grok
The one that talks the most by far: fifteen third parties, seven listed. Twitter's advertising (static.ads-twitter.com, t.co) and its measurement, DoubleClick (two names), Google Analytics and Tag Manager, Cloudflare Insights, OneTrust (the cookie notice), Stripe (payments, three names), a tag manager of its own and a Google STUN server. It could not be confirmed from outside that it answered.
Third parties in this run: stripe.com, doubleclick.net, twitter.com, cookielaw.org.
Mistral
Not comparable: my session was not signed in and the page bounced to "Sign in", so what was measured is its login screen, not a conversation. What it asked for in that minute was Intercom (the support chat, three names), Sentry and Cloudflare Insights.
Third parties in this run: intercom.io, sentry.io, cloudflareinsights.com, live.com.
What can be claimed
- Signed in and answering a question, ChatGPT made the machine talk to no name that was not its own.
- Gemini, signed in and answering, loads six third-party services and all six are on the open advertising and tracking lists.
- With the page open under my account, Grok loaded the most third-party names: fifteen, seven of them listed.
- The open lists still miss most of these services: none of DeepSeek's six is on any of them.
What cannot be claimed
- That the machine went on to connect to those services, or that it sent anything: what is seen here is that the machine asked for the name. A connection almost always follows, but a browser also looks names up ahead of time and sometimes never uses them.
- What is sent when there is a connection, or what is written in the conversation: GUARDIANA sees names, not content.
- That an AI is "spying": these are documented sign-in, measurement, support and payment services, and the owner of each name is identified above.
- That the one loading least is the best, nor that zero on the lists means zero tracking: they are names, not intentions.
- That each name was asked for by that AI's page: there is no view from inside the browser here. What is counted is the names the PC asked for during each AI's minute, with the known Windows and Chrome ones left out, and another application on the machine could have asked for one.
- That the four without confirmation did not answer: it only means their window does not change title and from outside there is no proof.
The record
- Date: 18 September 2026 (Bogotá time), 13:27 to 13:38, with Perplexity repeated at 13:47.
- Machine: a Windows 11 test PC with its DNS pointing at GUARDIANA 0.2.3 on the machine itself. Only this machine's own queries are counted.
- Browser: Google Chrome 153 with my everyday profile and my sessions already signed in. No password was typed or read: the session that was already open was used.
- Each AI in its own tab: address typed, 16 seconds to load, one short question typed ("Cual es la capital de Colombia"), 40 seconds of waiting, tab closed. One minute per AI.
- Windows' name cache was emptied before each AI. Without that step the machine stops asking for what it already knows and the window comes out empty: that is how a first attempt came out, and it was discarded.
- Proof that the page loaded and answered: the window title before and after the question. In three it changed to the text of the question (ChatGPT, Gemini and Perplexity). The others do not change title and from outside it cannot be proven.
- When counting "third-party services", Windows' own names, the ones Chrome asks for by itself (updates, safety lists, account services) and the machine's search engines are left out.
- Claude is not measured here. GUARDIANA is built with its help, and measuring whoever helps you is not an independent measurement: it is left out by rule, before the result is known, not because of what it gave.
- Made by GUARDIANA in Cartagena de Indias. None of the eight companies took part.
Data and method
Full data: ocho-ia-con-tu-cuenta-al-conversar-2026-09-18.csv (name, category, whether it belongs to the maker and, per AI, queries and whether its page asked for it). Method: each AI opened in its own tab of my own Chrome (real profile, sessions already signed in) with site/src/radiografia/ia-teclas.ps1: name cache emptied before each one, address typed, 16 s to load, one short question typed, 40 s of waiting, tab closed; no view from inside (Chrome has not allowed DevTools on the default profile since version 136), with the window title before and after as proof that it loaded and answered. Names taken from GUARDIANA's ledger export for each AI's window, this machine's queries only, leaving out Windows' names, Chrome's own and the machine's search engines. My list of open tabs was saved beforehand and put back at the end. The tools are in the program's repository (site/src/radiografia/); anyone can repeat the test with Guardiana installed.
Brands named as the names of their owners. If you are the maker and something here is inaccurate, write to hola@guardianagroup.com: it is corrected, with the date.