ChatGPT Fell Below 50% of AI Users. Your One Test Is Not Enough

ChatGPT fell to 46.4% of global AI assistant users by the end of May 2026, the first time below half, while Gemini's share of AI referral traffic more than doubled. The practical consequence is that checking how one assistant describes your business now tells you very little about the others, which read the web differently.

The referral picture tells the same story from another angle. SimilarWeb has ChatGPT's share of AI referral traffic sliding from 76% to roughly 53%. Other trackers put the decline more gently. None of them show it holding.

Almost everyone tested once, in one place

Nearly every business that has done anything at all about AI visibility did the same thing. Opened ChatGPT, typed "recommend a company like mine," read the answer, and either felt fine or felt slightly sick.

Eighteen months ago that was a reasonable proxy for the whole market. It now describes under half of it, and the half it describes is the half that is shrinking.

Fragmentation is the real problem, not the decline

One dominant assistant is a solvable situation. You learn what it rewards and you do that.

Four assistants with genuinely different behavior is a different kind of problem, because they do not read the world the same way. Each builds its answer from a different part of the internet.

• Gemini leans toward official company sources.
• Perplexity leans heavily on forums and user accounts, taking 46.7% of its top citations from Reddit.
• ChatGPT draws far less from forums, around 11.3%, with Wikipedia far more prominent.

The same business can be confidently recommended by one and effectively invisible to another, for reasons that have nothing to do with the quality of its work.

So what does "we checked how AI describes us" actually mean now?

Much less than it sounds. Checked where? The answer that reassured you came from one system with a falling share, and it tells you almost nothing about the three that are growing.

The failure mode here is worth stating bluntly. A business can run the test, get a flattering answer, conclude it is fine, and be entirely absent from the assistant its next ten customers actually use.

Chasing all four is also a trap

The instinct after reading the list above is to build four strategies. Do not. You cannot optimize four moving targets with different logics and different release cycles, and anyone selling you a package that claims to is selling certainty that does not exist yet.

What transfers is duller and more durable. Every one of these systems is trying to do the same job: establish what you do, who you serve, and what happened when you did it, then answer a question with something concrete. They differ in where they look, not in what they need.

A practical baseline to run this month

1. Ask the same three questions of four assistants: what does this company do, who is it for, and who are its alternatives.
2. Record the answers verbatim, with the date. This is your baseline.
3. Mark every answer that is wrong, vague, or missing you entirely.
4. Look at what the accurate answers had in common. Usually it is a page that states something plainly.
5. Repeat quarterly. The shares above will have moved again by then.

The businesses that handle this period well are not the ones with the cleverest tactics. They are the ones who noticed that "we asked ChatGPT and it was fine" stopped being a meaningful sentence, and started checking properly.

Why the numbers disagree between reports

Anyone comparing these studies will notice the figures do not line up, and it is worth understanding why before quoting any of them.

• Assistant users counts people using the product, which is what the 46.4% figure measures.
• Referral traffic counts clicks sent to websites, a much smaller and differently distributed population.
• Citation share counts which sources get quoted, which is different again.

A business can therefore be told that ChatGPT holds 76%, 53%, or 46% depending on which question was asked. All can be accurate. The consistent finding across every methodology is directional: one assistant is losing share and the others are gaining.

What a shrinking leader means for planning

A market with one dominant player rewards depth. You learn its behavior, you build for it, and the investment compounds.

A fragmenting market punishes that same depth, because the specific tactics that worked for the leader do not transfer, and the effort spent learning them is not recoverable.

Shifting share is the practical reason to resist buying assistant-specific optimization right now. Not because it never works, but because the half-life of a platform-specific tactic in a fragmenting market is short, and you are paying for something that may be obsolete before it is finished.

How often should you re-run the check?

Quarterly is a reasonable rhythm for most businesses, with one exception: re-check immediately after anything that changes how you are described publicly.

That includes a website rebuild, a repositioning, a name change, a merger, or a significant new service line. Those are the moments when the description a model holds and the description you would give diverge, and the gap can persist for months if nobody looks.

Keep the records. A dated log of what four assistants said about you is genuinely useful six months later, and almost nobody has one.

The cost of getting this wrong quietly

There is no alert when an assistant stops naming you. No ranking drop, no notification, no competitor visibly above you. The customer asks, receives three names, and picks one.

That silence is the genuinely new risk. Every previous shift in how people find businesses came with a signal you could see in a report. This one does not, which means the only way to know is to look on purpose, repeatedly, across more than one system.

Businesses that check once and feel reassured are not being careless. They are applying a habit formed when checking once was enough.