GEO statistics: what we actually know about AI visibility
AI usage numbers, a visibility-lift study, and one widely quoted prediction, sorted by source and date, with an honest question mark where one belongs.
Short answer: the hard numbers on AI visibility are scarcer than the confidence with which they get repeated. A few things we know fairly well (roughly how many people use ChatGPT), a few things are a prediction dressed up as a fact (that 25% drop), and at least one widely quoted figure turns out to be a lot shakier than the thousand repetitions suggest. Below is the breakdown, with source and date attached.
What this article is and isn't
This isn't a market report, and it isn't ammunition for a hockey-stick slide in an investor deck. It's an honest inventory: which AI usage and AI visibility numbers trace back to an actual source, which are a prediction being passed off as a measurement, and which ones just circulate without anyone still knowing where they came from. That last category is exactly what GEO-Crafter has a problem with: we measure real AI answers for brand mentions, we don't invent a potential score. This article applies that same standard to the topic of statistics itself.
How many people actually use AI assistants?
The best-documented number is ChatGPT's user count, though it changes fast enough to go stale within months. OpenAI reported on February 27, 2026 that ChatGPT had passed 900 million weekly active users, up from roughly 800 million in October 2025. Later in 2026, various outlets floated figures closer to 1 billion, but those are harder to pin to one clean, official statement than the 900 million figure from February.
| Point in time | Weekly active users (ChatGPT) | Source |
|---|---|---|
| October 2025 | approximately 800 million | OpenAI, via reporting from late February 2026 |
| Late February 2026 | over 900 million | OpenAI, own disclosure |
| Mid-2026 | approaching 1 billion (less clearly confirmed) | Various outlets, no single clean primary source |
The lesson here is less about the exact number and more about the pace: a hundred million more users in five months, by OpenAI's own account. A figure you quote today is out of date within a quarter. Always note when you read it, not just what it says.
How many people actually see Google's AI Overviews?
For Google's AI Overviews (the AI-composed answers at the top of search results), the clearest figure is older: on October 28, 2024, Google's own blog announced the feature had expanded to more than 100 countries, reaching "more than 1 billion global users every month," phrasing that describes expected reach after rollout rather than a separately measured usage figure. That's now more than a year old, and coverage has almost certainly grown since, but we couldn't attach a comparably fresh user count from Google itself to a specific date. Treat the 2024 figure as a floor, not a current reading.
What the Gartner prediction does and doesn't say
The most-cited number in GEO content is probably this one: "25% of search traffic will disappear by 2026 because of AI chatbots." That comes from a Gartner press release dated February 19, 2024, attributed to VP Alan Antin, and it was already the toned-down version of an earlier, even bolder Gartner claim (a drop of 50% or more). The detail that usually gets lost when this gets repeated: this was a two-year forecast, not a measurement. Critics at the time pointed out that the assumption that traditional search engines would stand still in the meantime is shaky, and that Gartner's own research showed users distrust chatbots for exactly the high-stakes queries where trust matters most.
Now that 2026 has actually arrived, we found no clear, independently measured confirmation that search traffic actually lost that 25%. That doesn't mean the underlying trend is wrong (AI answers are clearly absorbing some share of informational search behavior), but it does mean a number from a two-year-old forecast isn't the same thing as a number from a after-the-fact measurement. Say "Gartner predicted in 2024," not "research shows."
What we actually know about the effect of GEO techniques
The paper that put the term generative engine optimization on the map in 2023 (Aggarwal et al., published on arXiv and later presented at KDD 2024) remains the most concrete source for what GEO techniques actually do. The authors tested nine content interventions and report a visibility boost of up to 40% in their test setup, with up to 37% improvement in a live test on Perplexity. The strongest effects came from adding citations and statistics; keyword stuffing, by contrast, did almost nothing.
Three caveats that secondary articles tend to drop: it's a ceiling ("up to" 40%, not an average), the effect varies heavily by domain, and it was measured inside a research benchmark, not on a random business website. We wrote this up before in our own GEO checklist, and we're repeating it here deliberately: it's the only study in this article that measures an actual mechanism rather than just a usage count, which is exactly why it deserves the most nuance, not the least.
AI-referred traffic is growing fast, from a small base
Adobe Analytics has tracked, since October 2024, how much traffic to US retail sites arrives via links inside AI answers (drawn from more than a billion visits to US retail sites) and publishes on it periodically. The pattern is clear: explosive percentage growth that gradually flattens out.
| Period | AI-referred traffic growth (year over year) |
|---|---|
| November 2025 | +769% |
| December 2025 (holiday season) | +673% |
| Q1 2026 | +393% |
| May 2026 | +138% |
| July 2026 | +62% |
Source: Adobe Analytics, via repeated Adobe reporting across 2025 and 2026.
Don't read this table as "AI traffic is now a quarter of all traffic." Growth of hundreds of percent mostly tells you how small the starting base was: traditional search still accounts for the large majority of traffic for most sites. What the table does show is that the direction is settled, and that the percentage growth rate slows as the base gets bigger, which is exactly what you'd expect from a channel going from almost nothing to something real.
The zero-click number that keeps circulating
A good example of a statistic with a genuinely traceable origin that lost its caveats along the way: "two-thirds of Google searches end without a click." That traces back to SparkToro (Rand Fishkin) panel research from 2020, based on Jumpshot, Datos, and SimilarWeb panel data, landing at around 64.8%. SparkToro itself warned that stitching together different panel datasets makes year-over-year comparisons shaky. Google has also publicly disputed how the study gets framed.
The figure has circulated in various forms since, sometimes without the year attached, sometimes with a slightly different percentage from a later retelling. Tellingly, two providers that independently tried to measure how often a US Google search ends without a click in 2026 landed on figures roughly 45 percentage points apart. Not because either one is measuring nonsense, but because "a click" and "a search session" are genuinely hard to define in a way everyone agrees on. Use numbers like this as a direction, cite the original source and year, and be extra cautious the moment a figure circulates without anyone being able to say where it came from.
So what do we actually know for sure?
| Claim | Status |
|---|---|
| ChatGPT had over 900 million weekly users in late February 2026 | Measured, official OpenAI disclosure |
| AI Overviews reached over 1 billion users/month in late 2024 | Measured (dated), official Google disclosure |
| GEO techniques can boost visibility by up to 40% | Measured in an academic benchmark, no guarantee for your site |
| AI-referred traffic to retail sites is growing fast | Measured (Adobe), but from a small base |
| Search traffic will drop 25% by 2026 due to chatbots | Prediction (2024), not yet independently confirmed |
| Two-thirds of searches are zero-click | Dated, methodologically disputed figure, cite with its source |
What to actually do with this
Not every number on this list carries the same weight, and that's the whole point: the difference between a measurement, a prediction, and a number that just circulates should decide how much you lean on it in your own content or business case. Want to know where your own brand actually stands, instead of relying on market numbers that describe everyone at once? That's exactly what we do: we ask the buying questions your actual buyers put to ChatGPT, Gemini, Claude, and Perplexity, and measure the real answers for brand mentions. Curious what that looks like for your brand? Request access.
Further reading
- What is GEO? for the broader explanation of the field.
- The GEO checklist for what to actually do with the research effect covered in this article.
- What is AI visibility, and how do you measure it? for how to measure your own standing periodically, instead of relying on market-wide numbers.
Frequently asked questions
How many people use ChatGPT weekly? By OpenAI's own figure, it passed 900 million in late February 2026, up from roughly 800 million in October 2025. Reports of a number closer to 1 billion circulated later in 2026, but those are harder to trace to a single official statement. Treat numbers like this as an order of magnitude, not an exact count.
Is it true that AI chatbots will cut search traffic by 25%? That's a Gartner prediction from February 2024, not a measured outcome. Gartner expected traditional search volume to drop a quarter by 2026 because of AI chatbots and other virtual agents. Now that 2026 has arrived, we found no clear independent measurement confirming it actually happened, and critics flagged weak assumptions in the forecast at the time.
How much does GEO optimization actually improve visibility? The paper that introduced the term generative engine optimization (Aggarwal et al., 2023) reports a visibility boost of up to 40% in its test setup, with up to 37% improvement in a live test on Perplexity. That's a ceiling from an academic benchmark that varies heavily by domain, not a guaranteed result for your site.
Is the "two-thirds of searches end without a click" statistic still reliable? Its origin is traceable (SparkToro/Rand Fishkin panel research from 2020-2021), but the figure usually gets repeated without its caveats. Google has publicly disputed how it's framed, and two providers that tried to measure the same thing in 2026 landed on figures about 45 percentage points apart. Treat it as a direction, not a fixed percentage.