You are optimizing for four engines that send you nothing

You have seen the slide. Five logos in a neat row: ChatGPT, Perplexity, Gemini, Claude, Copilot. Underneath, a line about optimizing across the full AI search ecosystem. Maybe you made the slide. I have made a version of it.

It is a comforting slide. It suggests a diversified strategy, several bets, no single point of failure. It also, according to a fairly large pile of GA4 data, describes a world that does not exist.

Here is the number. Across 166 GA4 properties spanning 10 industries, measured from November 2024 through May 2026, ChatGPT accounted for 92.4% of trackable AI referral traffic. Not a plurality. Not a strong lead. Ninety two point four percent, out of 6.77 million sessions.

The slide with five logos on it

I want to be careful here, because there is a bad version of this argument and I do not want to make it. The bad version says the other engines do not matter, delete them from the plan, put everything into one basket. That is not what the data supports and it is not what I think.

The good version is narrower. It says your effort allocation should look something like your traffic allocation, and right now for most teams it does not. Teams are running five-engine visibility audits, buying five-engine tracking, and writing five-engine reports, for a distribution where one engine is doing nearly everything. That is not diversification. That is spreading a limited amount of attention across four things that are currently rounding errors.

And the total is not a rounding error, which is the part that makes the concentration matter. Monthly AI referral sessions across those properties went from 65,249 in November 2024 to 644,478 in May 2026. The channel grew 9.9x in nineteen months. ChatGPT alone went from 47,606 sessions to 610,910, a 12.8x climb. This is a real channel getting real volume, and it happens to be one channel wearing five logos.

What the AI referral traffic share data says

Underneath the headline number, the movement between engines is where it gets interesting. These are not five roughly similar competitors drifting slowly. They are diverging hard, in both directions.

Gemini is the one I would watch, and not because of its current numbers. Its referral share is small. But it is the engine wired into the surface with a billion monthly users, and referral traffic is a lagging measure of a product that mostly answers in place. A low referral number from Gemini does not mean low influence from Gemini. It means Gemini is not sending clicks, which is a different sentence entirely.

That distinction runs through this whole dataset and it is worth saying plainly. Referral traffic measures the engines that hand you a visitor. It does not measure the engines that mention you to someone who then types your name into a browser, which is exactly the attribution gap that shows up as direct traffic. So read 92.4% as ChatGPT's share of AI visitors sent, not as ChatGPT's share of AI influence. Those are two different metrics and only one of them is currently measurable.

There is a second reason the concentration matters, and it is about risk rather than allocation. A channel where one intermediary controls 92.4% of the flow is a channel with a single point of failure, and everyone who lived through a Google core update knows how that story goes. If OpenAI changes how ChatGPT surfaces links, or moves further toward answering without citing, the AI referral line in your analytics does not dip. It falls over. That is an argument for keeping the small engines on a watch list, but it is also an argument for not building revenue forecasts on a channel this concentrated until you understand what it converts at. Diversification you have not sized is not protection, and the source concentration research suggests this pattern shows up at several layers of the stack at once.

Perplexity fell. Copilot fell off a cliff.

Copilot going from 8,651 monthly sessions to 339 is the most striking line in the set. A 96% decline. That is not a competitor losing ground, that is a referral channel closing. And Perplexity, which spent 2025 as the engine everyone in GEO circles talked about most, is down 61% from its March 2025 peak and got passed by Claude in March.

Sit with the Perplexity number for a second, because it says something uncomfortable about how our industry picks its priorities. Perplexity was the engine with the most GEO commentary written about it per unit of traffic it actually delivered. It was interesting. It was citation-forward, it showed its sources, it made for good screenshots in a deck. It was easy to write about. It was never the volume.

None of this means Perplexity is dying or Copilot is irrelevant as a product. Microsoft moved Copilot deep into Windows and Office, where its usage does not produce a web referral at all. What died was the referral pathway, not necessarily the usage. But if your reporting counts referrals, and your strategy is built on what your reporting counts, then for your purposes the distinction is academic and the number is 339.

The number nobody measured is the one that matters

Here is my favorite thing about this dataset, and I mean that sincerely. The authors were explicit that they did not answer the most valuable question: conversion rate by platform. They named it as the single most valuable open question and said it remains uninvestigated.

That is an unusually honest thing to publish, and it points at the actual work. Because 92.4% of sessions is not 92.4% of revenue unless every engine converts identically, and there is no reason to assume that. A Claude user researching a technical purchase and a ChatGPT user asking a casual question are not the same buyer. We have seen the conversion gap between LLM referral traffic and paid search run wide enough that session share alone is a poor proxy for value.

The industry penetration spread deserves its own beat. SMB at 1.71%, insurance at 1.51%, finance at 1.19%, health down at 0.17%. That is a tenfold difference between the top and bottom of the same dataset. If you work in fintech, your AI channel is meaningfully more developed than a health brand's, and any advice calibrated to the aggregate is calibrated to nobody.

What to do with a 92% engine

So, practically. You are not going to ignore four engines because of one report, and you should not. But you can stop pretending your effort is evenly justified across them.

The reallocation is easier than it sounds because most of the four-engine work is duplicated effort anyway. Auditing whether your pricing page is extractable is one job, not five. Checking that your retrieval crawlers can reach you is one job. What actually multiplies across engines is the monitoring and the reporting, and that is precisely the part worth thinning out. Watch the leader closely, watch the rest cheaply, and put the hours you get back into the content and access work that pays out regardless of which logo is winning.

The fundamentals underneath all five engines barely differ anyway, which is the quiet good news in a concentrated market. Being extractable, being current, being accessible to retrieval crawlers, being corroborated somewhere other than your own domain. That work pays out in whichever engine happens to be winning in eighteen months, and it is the same work that produced 40% of leads arriving from AI engines in our tax savings platform engagement.

“A five-logo strategy for a one-logo distribution is not diversification. It is four-fifths of your attention spent on hedges you never sized.”

One caveat on the data itself, because it deserves it. This is 166 GA4 properties, not a census, and GA4 only sees engines that pass a referrer. Any engine answering in place, or stripping referrer data, is systematically undercounted here. That almost certainly understates Gemini and Copilot's real influence. The 92.4% figure is a fact about measurable referral sessions, and it is the right number to plan a referral strategy around, but it is not a claim about who is shaping buyer opinion.

So do the boring thing this week. Open your analytics, split AI referrals by engine, and compare your own distribution to 92.4%. Then split conversion the same way. If your split looks like the aggregate, weight your effort accordingly and stop apologizing for it. If it does not, you have found something genuinely worth knowing that no report could have told you. Either way you will be planning against your own numbers instead of a slide with five logos on it, which is the whole point of building a proper reporting and analytics layer in the first place.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.