Facebook Knows My GPU: Inside My Data Export

Disclosure: This post was mostly written by AI, based on my Facebook data export.

My first Facebook status was “at work”, on November 2nd, 2006. My last feed post was a shared link in December 2019. In between, I apparently had quite a lot to say. Then I mostly stopped.

I still checked cycling groups and used Messenger, but I’d come to think of Facebook as somewhere my old photos lived. Downloading my account gave me a different picture. There was a history of my graphics cards, thousands of advertiser names and records of visits to my Cloudflare dashboard.

I’d stopped writing updates. Other companies were still sending them.

Eight years of graphics cards

The hardware file was the first thing that caught my attention. It contains nine records, six of which look like this:

I recognise these machines. The dates are when Facebook last fully detected the hardware, so this isn’t a purchase history. But the October 2018 record is still there alongside the August 2026 one. A new computer didn’t replace the old entry.

I build software. I can see why hardware details would be useful when debugging a web application. Finding eight years of them in my social-media archive still made me pause. I had never thought of Facebook as a place that kept track of my graphics cards.

Why is my Cloudflare dashboard in here?

The more interesting discovery was in the separate data logs download. A file covering activity on other businesses’ websites, apps and shops contains ten detailed events from July and August 2026. Eight name Cloudflare, Inc.

Here’s what those eight records say:

The file’s description says contact information is supplied in hashed form. It lists the type, email, without exposing the address or hash. The records also include a browser user-agent string.

The dashboard URL is the part I care about. I use Cloudflare to run things. Seeing activity from that environment in a Facebook download is a very different experience from seeing a brand I once followed.

The export attributes these events to Cloudflare and labels the connection as server-side. It doesn’t show which dashboard pages I used, what I did there, or whether the events changed any ads I saw. It does show that Meta received more than a count of anonymous visits: the records include types of identifiers supplied with them.

There is a Notion event too, with email, first-name and last-name fields listed as hashed contact information. My account was receiving information from tools I used elsewhere, years after my last feed post.

The updates kept arriving

The account download also contains a broader summary of activity that businesses reported to Meta. Each record describes an event on another website or app: a page view, a purchase, or some other action the business chose to report. Facebook calls these records “events”; in the charts, I call them activity reports. A report isn’t necessarily a click or a purchase.

This summary contains 432 reports from 28 named businesses, received between January 2024 and August 2026. It is a separate file from the ten detailed log entries above. All 432 are dated years after my last feed post.

In September 2024, four businesses accounted for 76 reports. In January 2025, six businesses accounted for 36. The number of reports tells me how many records reached Meta; the number of businesses tells me how many sources contributed them.

Most of the reports—378 out of 432—are labelled simply CUSTOM. That gives me a count, but little explanation of what happened. The smaller totals later in the chart don’t establish that tracking declined: this is what the download contains, and August 2026 is only a partial month.

396 friends. 4,230 advertiser names.

The download also contains two advertiser lists. These count names on lists, while the chart above counts dated reports from 28 businesses. The figures describe different parts of the archive.

In the full export, one list contains 2,196 names under “A list uploaded or used by the advertiser”. Another contains 2,142 under interactions I may have had with a website, app or shop. Only 108 names appear in both, leaving 4,230 distinct names. That doesn’t mean all 4,230 sent the activity reports shown above.

The export lists 396 current friends. I can place most of them. I couldn’t tell you how I ended up on most of these advertiser lists.

I’ve omitted some names from the public list below for privacy. The figures above describe the full export; the explorer counts only the names I’ve published.

LiveRamp, Acxiom, Experian Marketing Services - Audiences and Epsilon Audience Data Provider are in there. So are plenty of familiar brands. In the full export, about 95% of the names in the uploaded-list category don’t appear in the interaction category, comparing exact names. These lists describe much more than brands I remember clicking.

They don’t give me a chain of custody for my details. Different names can refer to the same business, and an entry doesn’t prove a purchase or a direct relationship. That’s the frustrating limit of the download: I can search for a company, but I can’t follow the route that put us in the same table.

I had opted out. Of some things.

The settings files helped explain why “I’ve opted out” was too vague a description of my choices.

My advertising preferences say I’m opted out of using third-party interactions and interests for targeting. A different file, covering off-Meta activity settings, contains “Account association enabled” and records zero uses of the clear-history and disable-profile-association controls. That settings record carries a 2023 timestamp, so it shouldn’t be mistaken for a view of today’s controls.

The files distinguish advertising use from account association. The controls have since changed: in June 2026, Meta announced the retirement of the old off-Meta disconnect setting and an expanded “Activity from other businesses” setting covering ads and other personalised content.

The export doesn’t establish that Meta ignored a setting. It does show why I need to read what a control actually governs—and check the current version before relying on it.

Why Muse is a harder sell—and a good business for Meta

This is the context in which I’m looking at Muse, Meta’s personal AI agent. Its pitch includes connecting email and calendars, shopping and getting things done across other services. That’s a lot more access to consider after finding a few dashboard visits in an archive.

Meta explicitly says Muse conversations and data in its virtual machine aren’t shared with its advertising systems. But its own safety explanation also says that websites Muse visits on your behalf can use that activity to show you ads. Its example is asking Muse to buy a shirt and then seeing an Instagram advert from the seller.

That is the connection to this export. Delegating a visit doesn’t make the website’s reporting disappear. Keeping a conversation out of ad systems doesn’t, by itself, keep the resulting shopping or browsing activity out of them.

There is a separate training choice too. Meta says it uses sanitised conversations and tool interactions to train Muse unless you opt out. Its launch documentation also says operational policies restrict staff access to VM data, while access remains possible to support, secure or operate the service. These are different promises from an agent whose provider cannot access its contents. Meta describes those boundaries here.

For me, connecting an inbox or a collection of work accounts would need a much better reason than curiosity. The convenience could be real; so is the amount of personal context I’d be concentrating in one service. This archive makes me less willing to treat a broad privacy promise as a substitute for understanding the individual controls.

The commercial appeal is easier to see. Muse for Small Business connects business tools with Facebook, Instagram and Meta ads, and helps create and improve advertising. It also offers paid subscriptions. My read is that Meta can gain a place in everyday business operations, make its advertising tools easier to use and sell subscriptions—all without feeding private chats into ad targeting. An assistant that helps a business make better ads could be valuable to both the business and the company selling the ads.

That’s why Muse can be a sensible product for Meta while still being a poor trade for someone trying to give it less access to their life.

If you still want to use Facebook

There are still groups and people worth keeping an account for. I’d start with a few practical changes.

1. Use an ad and tracker blocker. On desktop, I’d choose Firefox with uBlock Origin. For Chrome, uBlock Origin Lite is the project’s current option. These can block known advertising and tracking requests, as well as hide ads. Keep the extension updated and its tracking filters enabled. This protection applies in the browser; installing an extension doesn’t extend it to the native Facebook app.

2. Keep Facebook separate from other browsing. Firefox’s Facebook Container isolates Meta sites and blocks embedded Facebook trackers elsewhere. I’d avoid allowing unrelated sites into that container, and use a separate login instead of “Continue with Facebook” where possible. Some embedded comments and login buttons will stop working. Facebook can still observe what I do on Facebook itself.

3. Review the current account controls. Find “Activity from other businesses” in Meta’s account settings and disable its use for personalisation if that’s your preference. Meta describes this as a control over how it uses activity supplied by businesses; it isn’t a promise that those businesses stop sending it. The rollout varies by country, so check Meta’s current wording rather than following an old “disconnect future activity” walkthrough.

4. Check the companies sending the data. Review their advertising and data-sharing choices too. For an event like the Cloudflare one, I’d ask the sender what triggers it, which identifiers are included and how to opt out. A browser extension can’t make that decision on the company’s behalf.

The limit of ad blocking is right there in the export: “Server connection”. A blocker can stop a matching request from my browser, and that may reduce the information a business gathers. It cannot directly intercept a later request from that business’s server to Meta, or stop an advertiser uploading a customer list it already holds. I haven’t tested which events in this archive a blocker would have prevented.

So yes, I’d use one. I’d also keep Facebook contained and review the account and business settings. An empty ad slot tells me an advert was blocked; it doesn’t prove the reporting stopped.

What I’d check next

I’m glad I requested the downloads. The GPU table was a good curiosity; the Cloudflare records gave me a specific question worth asking. Why were those dashboard events sent to Meta, and what controls that reporting?

I also want to save the current settings and compare later exports after making changes. New event dates and the types of identifiers reported would be useful things to examine. A missing row alone wouldn’t prove that collection had stopped: the exports cover different periods and levels of detail.

For years, I judged my Facebook footprint by how little I posted. Next time I’ll also look at who’s contributing to the account on my behalf.

Source notes: the personal data comes from my August 18 and September 3, 2026 exports; the post’s date is its publication date. Advertiser counts use exact names, not verified corporate identities. These exports predate Muse’s public launch and provide no evidence of my using it. Product and privacy guidance draws on the linked documentation, checked on October 25, 2026.