The Take — Google's Gmail ad promise ended at the word "directly"

Google did not break its promise about your inbox. It made the promise impossible to break, which is worse.

Read the October 16 privacy policy next to the help page that explains it and the same sentence does two jobs. The policy keeps the five-year-old line — "We don't show you personalized ads based on your content from Drive, Gmail, or Google Photos" — and attaches a condition: "unless you choose to make data from these services available to other Google services that support personalized ads." The help page then reassures you that ads "will not be based directly on your Gmail inbox or your Google Photos library." Directly is the whole firewall. Everything a reader would want protected now hangs on a distinction between raw content and derived content, and Google's own documentation says that distinction can disappear.

"In some cases, for example, where the email or file is brief or particularly relevant to respond to the inquiry, the summary may just be the email or file." That is Google's Search help page on connected content apps, as PPC Land's line-by-line read of the change documents. The same page describes what Search draws from your mail — "summaries, excerpts, generated media, and inferences" — and then says it plainly: "That same info may be used to show you ads, so they are even more helpful for you."

So the reassurance is not "we can't see your inbox." It is "the ad system sees a summary" — except when the summary is the email.

What actually got easier to do

I don't think this is a word game, and I don't think it's a secret. It's a scope change, published, effective, with a redline anyone can read. Personal Intelligence, the opt-in that lets Search and Gemini read your Gmail, Calendar, Photos and Drive, is now a legitimate input to the ad system. The policy adds "personalized content" alongside "your interests and activity" as a basis for choosing an ad. That is a different class of signal — not what you clicked, but the model's reading of documents you never wrote for an ad system. Our analysis walked through how the four edited passages fit together; the detail worth repeating is that the definition of "personalize content" lives in a linked note about hotel bookings and never mentions advertising.

Then there's the switch itself. One opt-in covers all Search services, which Google defines as Search including AI Mode and Discover, Maps including Ask Maps, and Shopping. It is available to signed-in personal accounts, 18 and over. And disconnecting doesn't undo it: Google's help page says removing a connected app "doesn't delete any data" Search already holds, and that data "can still be used for product improvement, including training generative AI models." A consent you can't reverse at the data level is not the kind of consent the word implies.

Why now is not mysterious either. AI Mode answers the question, which means fewer result pages to sell. Google had more than 75 million daily users on that surface when it started putting Shopping ads into it in February, and it began a small US test letting healthcare advertisers into AI Mode and AI Overviews in June. An untargeted sponsored slot inside a conversational answer is worth very little. One placed next to a restaurant suggestion, informed by the hotel confirmation sitting in your mail, is worth a great deal. OpenAI solved the same problem by keeping the ad system architecturally separate from the answer — an orchestrator that nominates candidates into its own auction. Google's version doesn't need that separation, because its ad system and its assistant were always going to share a corpus.

There is a historical joke here that nobody at Google will enjoy. In June 2017 the company stopped scanning consumer Gmail content for ad personalization — a change Diane Greene announced for more than 1.2 billion users, as NPR covered at the time. It wasn't a legal retreat. It was a trust retreat, made to close enterprise deals. Nine years later the capability is back through an assistant, and the sentence that retired it is still in the policy, with an "unless" bolted on the end.

The strongest case for Google

Take the other side seriously, because it is not stupid. Connecting your apps is genuinely opt-in, and the feature is genuinely useful — asking about dinner near a hotel you booked and getting the confirmation is a real thing a real assistant should do. Advertisers never get your inbox or your photos; Google's statement to Ars Technica is that ads "will not be based directly on your Gmail inbox or Photos library, and as always, sensitive information will not be used to tailor ads," as Ars reported. Targeting on race, religion, sexual orientation and health stays prohibited. My Ad Center gives users controls and had more than 60 million daily users by early 2024. And search history has informed ads for two decades without the sky falling — a model summary is, in one reading, a further step away from the raw document, not a step toward it. Strip most of an email out and you get a signal that is coarser, not finer, than the text.

If that were the whole picture I would call this a reasonable trade and move on. It isn't.

Why the take holds anyway

Three reasons. First, a promise whose boundary you cannot see is not a promise, it's a posture. The policy says Search may process data "to personalize content, including ads"; the definition of "personalize content" is in a note that talks about hotels; the reassurance about advertising is on a help page, and the operative word there is "directly." A reader trying to work out whether their medical appointment email can shape an ad has to assemble the answer from three documents, one of which never says the word ads. PPC Land notes that neither the policy nor the help page explains how Google separates a connected email about a medical appointment from one about a flight. That is not a detail. That is the entire question, and it is unanswered.

Second, every safeguard that survives governs what advertisers see, not what Google's models may infer. No advertiser access is real and worth having — but it protects against the wrong failure. The harm people are worried about isn't a marketer reading their correspondence; it's an inference drawn from correspondence and used to choose what they're shown. Sensitive-category rules are written for labelled categories. A model's summary of you does not arrive carrying a label. It arrives carrying the fact that you are moving flats, or that someone in your family is ill, and the rulebook has no vocabulary for that.

Third, consent. The toggle was designed to make answers better, at the moment a user wants an answer, several taps deep, framed around utility rather than advertising. GDPR wants consent that is specific and informed as to purpose. Google has now written an advertising purpose into a consent flow built for convenience, which is precisely the argument European regulators have been having with Meta about paying for privacy. If that consent were later accepted as valid, my complaint would shrink to a UX grievance. Until then, "you chose this" is doing work the choice cannot support.

What would change my mind

Publish the separation, not the sentence. If Google states, technically and auditably, that Personal Intelligence output cannot enter the ad-ranking pipeline — no summaries, no inferences, a hard boundary like the one OpenAI built between its answer model and its auction — then most of this collapses into a documentation problem, and I'll say so. Second, a data protection authority ruling that this exact flow meets the bar for specific consent to ad personalization. Third, numbers: how many accounts have connected Workspace or Photos, and how many disconnect after being told what it feeds. Google has published none. A year with no ad shaped by connected content would also be evidence, though I'd want to see how anyone outside Google could tell.

The uncomfortable part is that none of this requires bad intent. It requires a company that needs its answer surface to earn, a policy team that writes careful sentences, and a help page that answers the question in a footnote. The promise didn't break. It just stopped meaning anything a reader can check.

If you had Personal Intelligence on before this policy change, would you still have opted in? Write to us via our contact page.