Hi everyone, this is Neo.

SEO folks are living in a state of whiplash right now, and my inbox shows the same two conflicting anxieties every day:

• The optimists’ screenshot: “See? Similarweb says 95% of ChatGPT users still use Google — Google isn’t dying, stop the doom-posting!”
• The pessimists’ screenshot: “See? Bocconi University studied 45,000 households and found traditional search queries dropped 17% after ChatGPT Search access — pivot now or die!”

Same timeline, two headlines, feels like parallel universes. Who’s lying?

I read the underlying research carefully, and here’s the answer: neither side is lying. They just aren’t counting the same thing.

1. First, understand: the traffic world runs on three separate ledgers

This is the master key to the whole debate, so please remember it — “people,” “queries,” and “clicks” are three independent ledgers, and they can move in completely opposite directions at the same time.

Ledger	What it counts	Typical source	Typical headline
People	How many individuals appear in a platform’s audience over a period	Similarweb audience overlap	“95% of ChatGPT users also use Google”
Queries	How many searches each household or user runs	Bocconi University clickstream study	“Traditional search queries fell 9.4%–17% after ChatGPT Search access”
Clicks	How many searches actually lead to an external website	Field experiments, Search Console	“AI Mode cut outbound click-through by 18.8 percentage points”

How can all three be true at once? Because one person can keep using Google every single day while steadily handing a specific class of questions to ChatGPT. The 95% overlap answers “are people leaving?”; the 17% decline answers “are questions moving?” Between those two answers lies the entire AI-era traffic anxiety gap.

2. The people ledger: Similarweb says AI layers on top of search

Start with the optimists’ source. The most-quoted numbers from Similarweb’s 2026 Generative AI Landscape report:

• Between March and May 2026, 461 million of ChatGPT’s roughly 494 million global users — about 95% — also appeared in Google’s audience. That 95% overlap rate held steady from September 2025 through May 2026;
• Generative AI platforms averaged 9.5 billion monthly web visits over the period — up 70% year over year. Growth is genuinely ferocious;
• But keep the comparison in view: traditional search engines still drew 3.3 billion monthly unique visitors — five times the 655 million across all AI chatbots combined.

Similarweb reads this as AI being added to search rather than replacing it.

But before you screenshot the 95%, notice the report’s other numbers that never made the headlines:

• In May 2026, only 6.8% of US ChatGPT answers contained a link to an external website (up from ~1.6% in June 2025) — AI is resolving a growing share of questions in place, while still rarely pointing people elsewhere;
• 65% of the URLs ChatGPT cited were deep content pages, yet 58.8% of AI referral traffic landed on homepages — what gets cited and what gets clicked are often not the same page;
• Similarweb itself concedes the overlap figure excludes API usage, desktop apps, and in-app AI features — precisely the usage modes most likely to replace search.

And here’s the killer caveat: the bar for counting as a “Google user” is embarrassingly low. Anyone who used Google once that month — a maps lookup, a login, store hours — counts. They could be routing most of their research questions to ChatGPT and still sit inside Google’s audience numbers. Pew’s February survey of 5,119 US adults adds texture: 49% have ever used a chatbot, 42% use one to find information, 60% read AI summaries atop search results — but “ever used” is a staggeringly broad metric that says nothing about intensity.

The people ledger can only tell you if anyone left. It can’t tell you what they’re doing on Google anymore, or how deeply they use it.

3. The query ledger: Bocconi’s causal study — the same people, searching less

Now the pessimists’ source — and it’s the most rigorous causal evidence I’ve seen on whether AI actually steals traditional search volume (the paper, titled “Answering Without Referring: How AI Search Rewrites the Web’s Economic Bargain,” hit arXiv on July 8, 2026).

Here’s the clever part: the researchers exploited a natural experiment. OpenAI rolled out ChatGPT Search to users in three staggered waves — paying subscribers in October 2024, free users in December 2024, anonymous users in February 2025. Who got access when was decided by OpenAI’s product schedule, not by whether a household liked AI — which cleanly isolates cause and effect. The data: Comscore desktop clickstreams from 45,000+ US households, October 2024 to July 2025.

The results:

• After gaining ChatGPT Search access, households cut traditional searches by an average of 3.14 per week — about 9.4% off the 33.51 baseline;
• The decline deepened over time: 17.0% after 20 weeks;
• A clever control (logged-in users who got access in December 2024 vs. anonymous users who got it in February 2025) showed a smaller drop — 4.9% on average, 8.2% after 20 weeks — which the authors read as early adopters simply being more AI-inclined, inflating the naive estimate. Honest methodology.

Now the part that matters most for independent sites — where the decline hit:

• The biggest referral losses were all informational categories: academic sites -32.8%, reference sites -26.5%, developer sites -15.1%, news sites -13.4%;
• Referrals to marketplaces and entertainment sites barely moved;
• In the same panel, Google’s household reach held steady at ~49% while ChatGPT grew from 4.6% to 7.9% — Google kept its people while per-household query volume fell;
• And the number that should chill every content publisher: ChatGPT conversation sessions produced an outbound click just 5.2% of the time, versus 31.1% for Google queries — and 74.4% of ChatGPT-using households never clicked out to any website across ten months.

One-sentence summary of this ledger: AI isn’t stealing Google’s people; it’s stealing the informational questions out of the search box — and those are exactly the questions that used to feed content sites and publishers.

4. The clicks ledger: what happens when everyone is forced into AI Mode

This ledger is the closest thing to a rehearsal of the future: what if AI Mode becomes the default search experience? A preregistered field experiment posted to arXiv on August 18, 2026 (Wang, Gleason, Bart, Wilson & Metaxa; N=1,100 US Chrome users) answers it.

The design was clean: three days observing participants’ natural search behavior, then random assignment into three groups for seven days — one unchanged (with AI Overviews), one with all AI features hidden, and one where every search was routed through AI Mode.

The results are striking:

• The AI Mode group’s outbound click-through rate collapsed by 18.8 percentage points (95% CI: -22.2 to -15.3); conversely, hiding AI features raised click-through by 8.8 points;
• Daily search sessions in AI Mode fell by ~0.92 from an average of 4.0 — and the decline was largest among the heaviest Google users;
• Session duration actually rose by 0.43 minutes — people spent more time inside the AI experience and clicked out to the web less;
• By category: news clicks fell 12.5 points, Reddit 21.2, Wikipedia 9.9;
• Forced AI Mode users started voting with their feet: use of Bing, DuckDuckGo, and Yahoo rose 11.2 points, while trust, satisfaction, and perceived usefulness all declined;
• Of 309 AI Mode users, 47 (15.3%) reported difficulty reaching specific websites — one complained that when they did land on a site, “it would tell me about the website instead.”

A corroborating earlier experiment (Agarwal & Sen, June 2026) found that when an AI Overview appeared, outbound organic clicks dropped 39.8% and zero-click searches rose 34.5% — same direction, same story.

And what does Google say? At I/O in May 2026: AI Overviews at 2.5 billion monthly active users, AI Mode past 1 billion in its first year, Liz Reid citing record query volumes, and Pichai’s summary: “When people use our AI-powered features in Search, they use Search more.”

Note: none of those claims has independently verifiable user-level data behind it. And the two academic experiments point the opposite way from Google’s narrative — are users “searching more,” or are they lingering longer in AI while pointing at the outside world less and less? The data favors the latter.

5. Translating this for independent site owners

Let me translate the research into three sentences of action:

First: your people aren’t leaving, but your “entry points for questions” are migrating. Don’t be numbed by the 95% overlap, and don’t be terrorized by the 17% decline. The truth: B2B, tutorial, news, and industry-research content is losing its staple diet of Google referrals; ecommerce, transactional, and entertainment content is still standing in the safe zone. Which side your niche falls on determines how worried you should be and how fast you should move.

Second: “AI plants the seed, Google closes the sale” is today’s most common path — and your analytics can’t see it. A user gets your brand recommended in ChatGPT, then searches your brand name on Google to buy — in Search Console that visit shows up as “organic” (this is the branded-search lift). No tool attributes it back to the AI session. Your dashboard may be uglier than reality, or prettier — either way, judging AI channels by referral traffic alone is like navigating by a broken compass. Add two measurement dimensions: AI visibility (your mention and recommendation rates in ChatGPT/Perplexity and friends) and branded-search volume.

Third: informational content fights to be cited; transactional content fights to be trusted. If your content teaches people things, change the goal from “rank #1” to “be the source an AI answer cites” — structured data, verifiable facts, clear attribution of claims; this is the GEO/AEO work I keep writing about. If your content gets people to spend money, congratulations — Bocconi’s data shows transactional search is the most resilient category. Obsess over conversion details, review trust, and branded-search capture on Google; that’s worth more than chasing any AI platform du jour.

Neo’s take

A few final thoughts.

The “95% vs. 17%” fight is really two different questions being argued as one. The pessimists citing query decline to argue “Google is dying” are wrong — the people are still there, transactional search is intact, and Google’s ad machine is barely scratched. But the optimists citing 95% overlap to argue “nothing is changing” are equally wrong — the migration of informational queries is real and causally documented, and content-site referral traffic is visibly shrinking.

I see this as a change in the quality of traffic, not just its quantity. Overall volume may not be collapsing, but the composition is shifting: more activity now resolves inside AI sessions (producing no clicks), less traffic flows out to external sites (clicks are scarcer and pricier), and the clicks that do reach you arrive with far higher purchase intent — Seer’s analyses put ChatGPT referral conversion well above Google organic, in some cases around 9x.

This “fewer but better” traffic era is both bad and good for independent sites. Bad, because pure volume plays are ebbing away. Good, because you no longer need to fight 100,000 competitors for the same first-page keyword — what you need is to become a brand that AI trusts and humans remember. The first gate is technical; the second is content quality and genuine reputation — and both are within your control.

There are still real gaps in the research: every dataset covers only a slice (no study tracks the same individuals across Google, AI Mode, and ChatGPT at the task level over a long horizon); Bocconi’s data ends in July 2025, before AI Mode scaled; and all three papers are unreviewed preprints. When ads start saturating AI Mode (Google began testing two AI Mode ad formats in May 2026), the click and trust numbers will move again. But the direction is already clear enough.

Keep the three ledgers separate and no scary headline will ever yank you around again:

• Watch the people ledger to judge whether long-term brand investment is worth it;
• Watch the queries ledger to see where demand in your category is heading;
• Watch the clicks ledger to understand what’s actually happening to your traffic structure.

Data doesn’t lie. But data that only shows you one ledger does.

I’m Neo, and I write about independent site SEO. If you’ve been staring at your analytics with a knot in your stomach, come back and tell me: which ledger is falling for you — people, queries, or clicks? Separate the books and half the anxiety disappears.