Late last month the industry spent a week debating a ratio, and seems to skip the number it sits on.
Three pieces of research landed within days of each other and carried an entire conversation. Marie Haynes looked at 12 months of GA4 data across 100 sites and more than 145 million sessions, and found that ChatGPT sent 0.62% of the sessions Google organic did, rising to 0.73% with every LLM included. The SEO Works benchmarked 95 sites and found organic search delivered 39 times the traffic of AI assistants. Kevin Indig ran 632 tests across 12 sites looking for AI visibility to show up later as branded search or direct traffic, and couldn’t find a reliable relationship. All three are careful pieces of work, generous with their data and candid about their limits, and I’d trust the numbers in each.
The conclusion the wider industry drew from them traveled faster than the work itself: AI sends almost nothing, so GEO is a distraction, so get back to Google. That reading answers a reasonable question: Which channel sends more visitors today, and the answer is plainly Google. What it leaves out is the size of the pool both channels draw from, and that pool is getting smaller.
Kevin, as it happens, published the clearest evidence of that a few days later. In his Similarweb analysis of 10,000 domains in each of the U.S. and France, organic search visits fell 13.2% in the U.S., and 24.4% in France between August 2024 and August 2026, and publishers lost roughly 41.5% of their organic traffic in the U.S. over the same two years. Paid search visits grew in the same data, and Alphabet’s third-quarter filing shows paid clicks still up 8% year over year, so the open web is losing visits from search while Google’s own paid inventory keeps growing, and no ratio between two referral channels will show you that. Kevin notes that paid growth replaced only about a third of the organic visits lost in the U.S., and that AI referrals, the fastest-growing channel in his sample, still account for about one visit in 200. Nothing on the search side of that picture refills the pool. His follow-up last week points to where some businesses are rebuilding it instead: email grew about 16% in his US sample, the steadiest gain of any channel, because it reaches an audience a business already owns.
Comparing AI referrals to Google organic tells you which of two inbound streams is smaller. It cannot tell you whether the total reaching your business is still enough to sustain it, and that second question is the one that decides budgets, headcount, and whether a traffic-dependent model survives the next three years.
The Ratio Holds Still While The Pool Underneath It Drains
The visits are not moving to a new channel so much as leaving the click altogether. A preprint from Bocconi University researchers Shi, Zhu and Gu, built on Comscore desktop clickstream data from more than 45,000 U.S. households, found that ChatGPT produced an outbound click in 5.2% of conversation sessions, compared with 31.1% of Google queries, and that wider access to ChatGPT Search reduced traditional search use by 9.4%. It’s a desktop-only preprint, but the direction matches everything else in hand: activity is moving to interfaces that answer in place and send fewer people anywhere.
Google’s own results page shows the same pull. Similarweb panel data for U.S. searches in early 2026 put the clicks reaching the open web at roughly 232 for every 1,000 searches. Last week I looked at Alphabet’s third quarter, where Google reported query growth at a five-year high, by a count it hasn’t defined, alongside its slowest search revenue growth of the year. More searching and fewer visits can both be true at once, and that pairing is the shrinking pool seen from Google’s side. Google’s head of Search said last year that total organic click volume to websites was relatively stable, but that claim came without numbers, and the Q3 call offered no figure on clicks sent to websites at all.
The decline is uneven, and that is the practical point. Kevin’s data shows the typical small site roughly flat while large sites and publishers absorbed most of the loss. The number to watch is your own pool: total inbound from every source, month over month, set against impressions for your category queries, so you can see whether demand is still there while the visits fall away. Nothing about tracking the pool needs a new vendor, though, and you can chart it this afternoon from your own analytics..
I argued recently that being indexed now earns a place in an answer rather than a visit. The shrinking pool is what that shift looks like in your reporting. Past a certain point, no mix of sources refills it, and the useful question changes from where your visits come from to whether the visit was ever the thing you needed. For most businesses it wasn’t. The visit was a step on the way to a booking, a purchase, or a signed form, and if that step is disappearing, the work moves to whatever completes the transaction without it.
When The Visit Disappears, The Transaction Still Has To Happen Somewhere
Marie’s piece points exactly here. Alongside her data, she encouraged readers to spend their time preparing for the agentic web, naming UCP, Universal Cart, WebMCP, and intelligent eyewear, and I agree with her. Google is pointed the same way: on its Q3 earnings call, Sundar Pichai said agents will operate as a unifying layer across Google’s products, booking reservations, filling in forms, scheduling appointments, and making purchases. Agents acting on a searcher’s behalf are the most credible offset to a shrinking pool, not because they send traffic, but because they finish the job the traffic used to exist for. That changes the unit you are working toward, from the visit to the completed action, meaning the booking, order, or request an agent finalizes on someone’s behalf.
The clearest evidence of how agents choose comes from a McGill study by Davood Wadi and Yu Ma. They randomized the order of 100 hotel listings across 5,000 agent sessions on four large language models and compared the results with human data from Expedia. The agents searched more deeply than people and never once declined to book. Position still affected which listings they opened, but weakly and in an unexpected shape: the middle of the page was inspected least, not the bottom. Rank carried through to the final booking for only two of the four models. And every model converged on the same listing, the one nothing else beat on review score at its price, which took 78.2% of all bookings. The authors’ own conclusion is that for agentic search, the attributes displayed on a results page matter more than placement within it.
The same researchers found a second effect worth knowing before you assume agents are neutral shoppers. In a companion study, agents told they worked for a booking platform chose sponsored listings at roughly nine times the rate of agents told they worked for the traveler, 25% against 2.8%, with every other attribute matched. Whose agent it is shapes what it buys, and on the same Q3 call Google’s chief business officer said he’s excited about the advertising opportunities in an agentic world. A strong offer remains the price of entry everywhere, but on agents a platform deploys, sponsorship can still outrank it, so it pays to know which agents your buyers are actually using.
An AI agent does not need to find your business first. It needs to find nothing that beats your business on the attributes it can read, because agents converge on the undominated offer regardless of where that offer sits in a list.
That is also why the plumbing alone won’t carry anyone. A checkout protocol makes it easier to buy from you once an agent has decided to, and does nothing to make it decide. OpenAI ran a version of this experiment in public: It launched Instant Checkout inside ChatGPT in September 2025 and pulled back from it in March 2026, moving purchases back to merchants’ own sites and apps. The infrastructure matters, and it comes second.
Here is where that infrastructure stands today. Google’s Universal Cart launched in the U.S. on Oct. 15, with checkout running on the Universal Commerce Protocol, and on the Q3 call Google said UCP now enables direct checkout for hundreds of thousands of brands. That count almost certainly includes merchants enrolled through commerce platforms such as Shopify, because a public scan in May found 26 sites publishing a UCP profile of their own across more than 3 million. Both numbers can be true, and the gap is telling: If you sell through a platform, you may already be eligible without building anything, and if you don’t, very few competitors have built it themselves either. WebMCP, which lets a page expose actions such as booking or checking availability as tools an agent can call directly, is still an origin trial in Chrome, and Marie herself notes it may be early for most businesses to invest heavily there. Meanwhile, Cloudflare’s new defaults block agent traffic on ad-supported pages for new domains and for Free-tier sites that never changed their settings, so plenty of businesses are turning agents away without knowing it.
Earlier this fall, I declined to publish a checklist for measuring what AI does to your numbers, because there was no ground truth to score it against. This list is a different kind of thing. Every item is checkable, since an agent can either read your offer, trust it, and complete an action with you, or it can’t. The order runs offer first and plumbing last, with an honest maturity label on each.
The Agent-Ready Short-List
- Track your pool, not the ratio. Chart total inbound sessions from every source each month, beside impressions for your category queries and segmented by page type, and decide in advance what level of decline triggers action. Now.
- Create an undominated offer. List the three to five attributes an agent can read in your category, usually price, rating, delivery, and terms, then look for any competitor that beats you on all of them at once, and either close that gap or state your trade-off plainly. Now.
- Publish reputation an agent can read. Keep review score and review count current, in structured form, and consistent across the platforms agents check. Now.
- Publish policies as data, not prose. Give returns, shipping, warranty, cancellation, and service levels their own structured fields, so an agent can compare your terms the same way it compares your prices. Now.
- Publish complete feeds. Retailers should fill every Merchant Center attribute for price, availability, and shipping, which Universal Cart reads directly, and service businesses should publish packages, price ranges, service areas, and availability in the same machine-readable spirit. Now.
- Publish the same facts everywhere. Name, pricing, locations, hours, and specifications should agree across your site, your profiles, and the directories agents cross-check, because a conflict reads as a risk. Now.
- Configure your bot rules for signed agents. Let agents that sign their requests reach the pages where buying decisions happen, keep the stricter checks on account and payment steps, and review your Cloudflare AI settings, since blocking Training there also blocks multi-purpose crawlers such as Googlebot. Now.
- Code your commerce path. Check first whether your commerce platform has already enrolled you, then use its agentic storefront where one exists, or publish a UCP profile at /.well-known/ucp and apply through the Merchant Center integration hub, and keep product data clean for ChatGPT’s shopping results. Next, and US-only and selection-gated for now.
- Code WebMCP tools for one to three core actions. Booking, quoting, and checking availability are the strongest candidates, with the human path left fully intact. Pilot, while it remains an origin trial.
- Measure agent actions, not sessions. Log the orders and bookings that originate with agents, plus signed-agent requests in your bot logs, so you can see the offset arriving even when your analytics show nothing new. Next.
None of this brings back the old volume, and it isn’t meant to. The businesses that come through a shrinking traffic pool will be the ones an agent can read, trust, and complete a transaction with, and that work starts with the offer rather than the protocol.
The debate over whether AI sends enough traffic has a fair answer: It doesn’t, and it probably won’t. The better use of a week of argument is to ask how much of your business still depends on the visit at all.
- Google Search Queries Hit A 5-Year High, And Search Revenue Didn’t Follow
- Agentic Commerce And The New Rules Of Google Ads
- Habitual Publisher Traffic Is Collapsing
This post was originally published on Duane Forrester Decodes.
