Soon You Won’t Pick the Product. You’ll Pick Which AI Picks For You.

A famous AI statistic nobody can source, and shopping agents that pick for you: how much of your attention did you really choose?

Here’s a number you’ve probably heard, and maybe even repeated yourself: by 2026, 90% of content on the internet would be written by AI. It’s been cited in keynote talks, LinkedIn posts, and no small number of “here’s what’s coming” articles, almost always traced back to a 2022 Europol report on deepfakes. Here’s the part nobody mentions when they cite it: researchers who went back to that actual report, looking specifically for the number, couldn’t find it. Not in a table. Not in a footnote. Nowhere.

That’s not really a story about a statistic. It’s a story about you — specifically, about how much of what you currently believe, watch, buy, and think about was ever actually chosen by you, versus simply delivered with enough confidence and enough repetition that choosing it stopped feeling necessary.

The Stat Nobody Can Source

The 90%-by-2026 figure has a genuinely interesting life story. It shows up, cited with total confidence, in research write-ups, trade press, and casual conversation, everyone assuming someone upstream had already checked it. When the research group OODAloop actually went hunting for it inside the Europol document it’s always attributed to, they came back empty-handed and said so in print. The number didn’t get less popular after that. It just kept circulating, because by then it wasn’t really functioning as a fact anymore. It was a vibe with a percentage attached.

Here’s what’s actually happened, as closely as anyone has measured it. Graphite, an SEO research firm, tracked 65,000 URLs published between 2020 and 2025 and found AI-written articles briefly outnumbered human-written ones around November 2024 — then leveled off to roughly even, according to Axios’s October 2025 reporting on the data. Separately, Ahrefs looked at AI involvement more broadly, not just full articles, and found AI had touched a large share of the pages it studied — but most of that was partial: a paragraph cleaned up, a summary drafted, a headline optimized, not a whole page generated from nothing.

The real picture is messier and less cinematic than 90%. It’s also the one you’d only find by going and checking — which is exactly the step most people, including plenty of smart, curious people, skip.

And it’s still happening, in real time, with new numbers. While researching this piece, I ran into a fresh crop of 2026 “AI content statistics” roundups citing a “comprehensive report” and a “Global Cybersecurity Alliance,” neither one linked, dated, or traceable to an actual publication. One listicle even contradicted its own headline stat in the very next paragraph, without seeming to notice. These pages tend to rank well, because they’re confident, specific, and written to be skimmed and cited, not checked. The pattern that sent “90% by 2026” traveling so far never went away. It just found a faster production method.

Your Feed Was Never Neutral

None of this started with generative AI, to be clear. Curation algorithms have been choosing your attention for close to two decades — just by ranking, not generating. YouTube’s recommendation engine, Instagram’s feed, TikTok’s For You page: all three optimize, at bottom, for one thing, which is how long you keep watching. Not what’s true. Not what you’d choose if someone asked you to reflect on it for thirty seconds first. What keeps the scroll going.

The economist Herbert Simon named the underlying problem back in 1971, decades before any of these products existed: a wealth of information creates a poverty of attention. He meant it as an abstract observation about an information-rich society. Engineers at a handful of the world’s largest companies later turned it into a design brief. Frances Haugen, the former Facebook product manager, made this concrete in 2021 when she brought internal company research to the U.S. Senate showing Instagram’s own researchers had found the app worsened body-image issues for a meaningful share of teenage girls who used it — a finding the company had internally before the public ever did.

None of that required AI. It required an optimization target and a few billion daily choices pointed at it.

That’s not an abstract claim, either. Journalists and researchers who tested TikTok’s recommendation system with fresh, contentless accounts found it could start homing in on a person’s specific interests and vulnerabilities from watch time and rewatches alone, well before any profile information was ever entered. The algorithm didn’t need to ask what you wanted. It just watched what you did.

Now Something Chooses What Gets Made, Too

Here’s the layer generative AI actually adds, and it’s not just “more content.” It’s a feedback loop on the input side. When ranking algorithms decide what performs, and generation tools make it nearly free to produce more of whatever performs, the stuff that gets made starts drifting toward whatever the ranking algorithm already likes — regardless of whether that was ever the same thing as worth your time.

You can watch the shape of this in any feed you already use. A format that spikes engagement gets cloned within days, at a volume no group of human creators could match, because a tool can generate dozens of variations on a winning format before a person finishes their coffee. The feed doesn’t just rank what exists anymore. It quietly commissions more of what already worked, and less of what didn’t. Nobody voted on this. It’s just what happens by default when the cheapest thing to produce and the most rewarded thing to produce turn out to be the same thing.

There’s a longer-range version of this problem too, and it’s already been documented, not just theorized. Researchers studying what happens when AI models train on data that was itself generated by earlier AI models have found a degradation effect across successive generations, sometimes called model collapse, where outputs drift toward the bland and the repetitive, losing the tail end of genuinely rare or original material. If more of the open web is written by AI, and future models learn largely from the open web, the feed doesn’t just reflect what’s rewarded today. It risks slowly forgetting whatever never got enough engagement to be noticed in the first place.

AI Is Starting to Choose Your Choices, Too

Attention is one layer. The next one is decisions themselves, and it’s moving faster than most people have clocked.

Roughly 38% of consumers already use AI while shopping, and 80% expect to lean on it more, according to industry research from the IAB reported via eMarketer. Adobe, which tracks a large share of U.S. online retail through its own analytics platform, predicted AI-assisted shopping traffic would grow 520% over the 2025 holiday season compared with the year before. And the tools doing the assisting have names now: OpenAI’s ChatGPT can complete a purchase directly through Instant Checkout with retailers including Etsy and over a million Shopify merchants; Amazon’s Rufus sits inside the Amazon app comparing options from Amazon’s own catalog; Perplexity’s shopping feature advertises its picks as unsponsored, with no ads mixed into the results. Google’s Gemini Shopping agent and Klarna’s AI assistant now do versions of the same thing, tracking price drops and financing options behind the scenes so that by the time one option finally surfaces, the timing and payment plan are already decided too.

Here’s the part worth sitting with: the job these tools are built for isn’t answering your question. It’s narrowing the field before you ever see it. Ogilvy’s head of innovation, Kaare Wesnaes, described the shift plainly in an interview with eMarketer: instead of opening ten tabs and reading twenty reviews, a shopper will increasingly just ask an agent to scan the market and bring back one recommendation. That’s a real convenience. It’s also a real handoff. You’re no longer choosing among the options. You’re choosing whether to trust whichever options got handed to you.

This Isn’t Really a “Log Off” Problem

The standard advice here is some version of “use your phone less,” and it’s not wrong exactly, it’s just aimed at the wrong layer. Screen-time limits treat this as a willpower problem, as if the issue is that you personally lack the discipline to resist a feed. But your willpower was never really competing with the feed. It was competing with a team of people, and now a growing stack of AI systems, whose entire job is to make choosing not-to-look cost more effort than choosing to look.

The more useful move isn’t resisting the object. It’s noticing the moment before belief or choice congeals — the split second where a claim either gets checked or gets absorbed, where a product either gets compared or gets quietly accepted from a shortlist. That moment is small, it’s fast, and it’s the only part of this entire chain you actually still control.

A Few Questions Worth Running Before You Nod Along

You don’t need to audit your entire information diet to make this real. A handful of questions, asked at the right moment, does most of the work:

• Where did I first hear this claim, and have I actually seen the primary source, or just someone’s summary of someone’s summary of it?
• Did I open this app right now, or did a notification pick this exact moment for me?
• When I compared my options, did I see the whole field, or a shortlist something had already narrowed on my behalf?
• Who benefits from the version of this that reached me first, and would a different version have reached someone with different incentives?
• If I asked this same question again tomorrow, or from a different account, would I get the same answer, or is this one already tailored to whatever keeps me engaged?

None of these require you to become a fact-checker or quit your favorite apps. They require pausing for about four seconds at the exact moment you’d normally just scroll, click, or repeat.

The Actual Takeaway

The 90%-by-2026 statistic was never really about AI content. It was a small, almost perfect demonstration of the exact thing this whole piece is about: a claim traveled further on confidence and repetition than it ever did on evidence, and most of the people repeating it, quite possibly including a past version of you, never once traced it back. That’s not a character flaw. It’s what happens by default, at scale, whenever checking costs more than accepting.

Attention, belief, and choice all run on the same currency: whatever reaches you first, confidently, and often enough. You can’t fully opt out of that system. You can get better at noticing when you’re standing inside it.

Soon You Won’t Pick the Product. You’ll Pick Which AI Picks For You. was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.