X-RAY · 20 SEPTEMBER 2026
The same eight AIs, asked something they cannot know: the tracking was already there
In the morning they were asked the capital of Colombia, which they all know by heart. In the afternoon, the weather in Barranquilla right now, which none of them knows: every one has to go and look it up. The surprise is that almost nothing changes. Gemini loads the same six advertising services both times; Copilot the same four. What they load, they load on opening, not on searching. The one real change: ChatGPT goes from one to three, and the two new ones belong to Google.
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. Each AI was opened alone in a fresh tab of the everyday Chrome, already signed in, the same question was typed into it, and the tab was closed. The name cache was emptied before each one. From what the machine asked for in that minute, Windows' own names, the signed-in browser's own traffic and the machine's search engines are subtracted. "Third party" means the name does not belong to that AI's company.
Scoreboard
How many services that are not that AI its page makes the machine ask for when it is opened signed in and asked something it cannot answer from memory. Beside it, how many of those are on the public advertising and tracking lists.
| AI | Services besides the AI | Of those, advertising or tracking | Its own domains | Distinct names the PC asked for in its minute | The fewest |
|---|---|---|---|---|---|
| Claude | 0 | 0 | 4 | 4 | yes |
| Meta AI | 0 | 0 | 2 | 2 | yes |
| ChatGPT | 3 | 0 | 3 | 6 | |
| Perplexity | 3 | 1 | 7 | 10 | |
| Copilot | 4 | 1 | 4 | 8 | |
| DeepSeek | 5 | 0 | 2 | 7 | |
| Gemini | 6 | 6 | 10 | 16 | |
| Grok | 14 | 6 | 2 | 16 |
Tied on 0, the fewest third-party services: Claude, Meta AI, against 14 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
The idea was that forcing them to search would open a new world: the search engine behind them, the pages they fetch, the measurement of each query. It did not happen. Gemini loads its six advertising services with the easy question and with the hard one, the same six name by name; Copilot, the same four. What they load does not depend on what you ask: <b>it is there from the moment the page opens</b>. The exception is ChatGPT, which goes from one third party to three, and the two new ones are t1.gstatic.com and t3.gstatic.com, the thumbnails of the search results: you ask ChatGPT, and your machine ends up talking to Google. Claude and Meta AI repeat zero. And Grok repeats the other end: fourteen.
Claude
Zero third-party services, the same as with the easy question. Four names, all four its own. It answered: the window title became "Clima actual en Barranquilla". GUARDIANA is built with Claude Code, by Anthropic. It is said here, in its own row, because hiding the partner is worse than showing it: the figure can be checked like any other.
Meta AI
Zero third parties. Two names, both its own. Its window never changes title, so from outside there is no proof it answered; what can be said is who it made the machine talk to while it was open, which is nobody.
ChatGPT
Goes from one to three third parties, and the two new ones are Google's: t1.gstatic.com and t3.gstatic.com, the thumbnails of the search results. You ask ChatGPT, and your machine ends up talking to Google. The third is a.nel.cloudflare.com, network error reporting. None of the three is on advertising or tracking lists. A SharePoint name also fell inside its window and is not counted: it turns up five separate times across the day, it is OneDrive in the background and has nothing to do with the page.
Third parties in this run: cloudflare.com, gstatic.com.
Perplexity
Seven names of its own and three third parties, one listed: browser-intake-datadoghq.com (tracker, application monitoring). The other two are product and campaign measurement: fscdn.eppo.cloud and sdk-api-v1.singular.net. With the easy question it was five; with this one, three.
Third parties in this run: browser-intake-datadoghq.com, eppo.cloud, singular.net.
Copilot
Four third parties, one listed -- exactly the same four as with the easy question: Microsoft's telemetry (browser.events.data.microsoft.com), Sentry for errors, and Sprig, two names, for in-product surveys. Two different questions, the same result.
Third parties in this run: sprig.com, microsoft.com, sentry.io.
DeepSeek
Two names of its own and five third parties, none on advertising lists: ByteDance's Volcano Engine (apm.volccdn.com, apmplus.volces.com, gator.volces.com), Cloudflare's bot check and fp-it-acc.portal101.cn. chatgpt.com fell into its window again in the first seconds: that is the browser guessing the previous address as it was typed, and it is not counted.
Third parties in this run: volccdn.com, volces.com, cloudflare.com, portal101.cn.
Gemini
Six third parties and all six listed, the same six as with the easy question, one by one: googletagmanager, google-analytics, googleadservices and DoubleClick by three routes. They are Google's, as Gemini is, but they are not Gemini. Two questions, the same thing twice.
Third parties in this run: doubleclick.net, google-analytics.com, googleadservices.com, googletagmanager.com.
Grok
Fourteen third parties, six listed: DoubleClick by two routes, Google Analytics, Tag Manager, X's advertising (static.ads-twitter.com) and t.co. The other eight are payments and measurement: Stripe with three names, analytics.twitter.com, OneTrust's cookie notice, a Google STUN server, stapecdn.com and a tag container of its own. Still, by a distance, the one that talks the most. gemini.google.com is discounted from its window: the browser guessing the previous address.
Third parties in this run: stripe.com, doubleclick.net, twitter.com, stripe.network.
What can be claimed
- With a question that forces a search, Gemini made the machine ask for the same six advertising and tracking names as with one that does not.
- ChatGPT made it ask for two Google names that did not appear with the easy question: the thumbnails of the search results.
- Claude and Meta AI made the machine talk to no third-party service, with either question.
- Four of the eight left proof that they answered: their window title became the answer.
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.
- That the four that never change title really searched: they never change, answer or not.
- What is sent when there is a connection, or what the answer says: GUARDIANA sees names, not content.
- That an AI with zero third-party names collects nothing: it can do so through its own domains.
- That these numbers are anyone's numbers: they depend on the account, the country and the minute.
- That Mistral is better or worse: it was not measured signed in.
The record
- Date: 20 September 2026 (Bogotá time), 11:23 to 11:33. About a 65-second window per AI.
- Question typed into all eight: "Que tiempo hace ahora mismo en Barranquilla y que temperatura hay". No accents, because the script types key by key.
- Machine: a Windows 11 test PC, the everyday Chrome with the real profile and the sessions already signed in, GUARDIANA 0.2.3 as its DNS.
- Nine AIs were measured; eight are published. Mistral is out again: its window ended on a sign-in screen.
- Four of the eight changed their window title to the answer -- ChatGPT, Gemini, Claude and Perplexity -- which is the proof that they answered. The other four never change title.
- An earlier attempt, at 10:39, was thrown away whole: with the window stretched to 60 seconds the run took thirteen minutes, the machine locked the session at ten, and the keystrokes stopped arriving. Eight of the nine measured zero, which is what a run that types nowhere measures. The script now aborts if it cannot bring Chrome to the front.
- 8,000 events in the ledger extract these counts come from.
Data and method
Full data: ocho-ia-buscando-en-internet-2026-09-20.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 the everyday Chrome (real profile, sessions already signed in) with site/src/radiografia/ia-teclas.ps1, launched as a scheduled task in the interactive session: name cache emptied before each one, the address typed into the omnibox, 16 seconds of waiting, the question typed, 40 more seconds and the tab closed. From each window are subtracted Windows' names, the signed-in browser's own, the machine's search engines and -- written one by one in each AI's note -- whatever the browser guessed while the previous address was being typed. 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.