The jug did not get fuller. The pouring got narrower

What actually happened to website traffic this year#

Every client asks the same question eventually, usually in a quiet voice: is it us, or is it everyone? A year of comparative website traffic data landed this week that lets you answer it properly for once, and the answer is more uncomfortable than either option.

Those movements come from Semrush's own trending websites data, which publishes the figures per site and updates them monthly. ChatGPT gained 48.38 per cent, YouTube 36.6 per cent. Bing lost 50.43 per cent, roughly half its measured traffic in a year, and Amazon shed 15.31 per cent.

Read the aggregate and the picture changes character entirely. Search Engine Land calculates that the top 150 sites grew only 6.1 per cent between them, that the top ten hold 68.6 per cent of all that traffic, and that the top three alone hold 54.3 per cent. Those three aggregate figures are Search Engine Land's own arithmetic over the Semrush data rather than Semrush's published totals, and it is worth saying so.

Before anyone reaches for a conclusion, one discipline is worth keeping. Website traffic figures at this scale are modelled from panels rather than counted at the server, so the sensible reading is of direction and rank rather than of decimal places. A fox trusts the scent, not the exact number of paces to the quarry.

Growth of 6.1 per cent is not a growing web#

Hold those two numbers side by side, because together they are the whole argument. Total measured growth across the biggest 150 sites in the United States was 6.1 per cent. The share held by the top ten was 68.6 per cent. A website traffic market barely growing, with more than two thirds of it in ten hands.

This is why so many perfectly competent content teams spent the year feeling as though they were failing. They were not necessarily doing worse work. They were operating in a website traffic market where the growth was being collected somewhere else, and where a flat year was, in relative terms, a decent one.

Where the growth actually went#

Semrush's own clickstream research puts numbers on the destination. Outbound referral traffic from ChatGPT to the wider web rose 206 per cent during 2025, and roughly 170,000 unique domains now receive referrals from it in a month. That is real ai web traffic arriving at real websites, and it is growing fast from a small base.

The same research contains a caution that rarely gets quoted alongside the good news. Web search was enabled on 34.5 per cent of ChatGPT queries by February 2026, down from 46 per cent in late 2024. A shrinking share of a growing whole, which means the referral engine is not scaling as simply as the headline suggests.

The distributional point is the one to carry into a planning meeting. If website traffic overall is barely moving while the largest properties absorb most of it, then the median site is losing ground even while the aggregate looks calm. That is a very different brief from the one most content teams were given in January.

Three explanations, and only one of them is comfortable#

The obvious culprit for the website traffic squeeze is artificial intelligence eating the click, and the evidence for it is real but narrower than the headlines suggest. It deserves stating carefully, because the difference between a measured effect and a moral panic is where a content strategy either works or wastes a year.

The best-instrumented study is from the Pew Research Center. Working with 900 United States adults on a probability-based panel, it observed 68,879 Google searches during March 2025, of which 12,593 produced an AI summary. The behavioural gap it found is stark and specific.

Pew also found that users clicked a link inside the summary itself in just 1 per cent of visits. So the summary is not redistributing the click, it is absorbing it. SparkToro reaches a compatible conclusion by a different route, reporting that 68.01 per cent of Google searches ended without a click in the first four months of 2026, against 60.45 per cent in 2024.

Google's own statement is worth reading in full rather than through a summary of it, and it says the company continues to send billions of clicks to websites every day. The honest position is that the aggregate and the individual experience genuinely diverge, and that publishers are not imagining a traffic decline simply because a platform-level total held steady.

The third explanation is the least discussed and possibly the largest: consolidation that has nothing to do with AI at all. Bing losing half its traffic is not an AI story. Amazon shedding 15 per cent is not an AI story. Some of what looks like an AI-driven reshaping of organic search traffic is simply large properties taking share from other large properties, as they always have.

Which of the three explanations dominates will differ by vertical, and that is precisely why a category comparison beats a web-wide average. A publisher losing website traffic to summary boxes and a retailer losing it to a larger retailer need entirely different responses, and the aggregate number cannot tell them apart.

How to read your own numbers without lying to yourself#

The practical value of a year like this is that it gives your website traffic a denominator. A 4 per cent fall reads as failure in isolation and as an outperformance against a category that fell 12. Without the comparison you are managing on feelings.

That was posted this week and it captures the operational squeeze precisely. The measurement job expanded, the budget did not, and the new surface is far harder to instrument than the old one. Anyone selling a tidy solution to ai search traffic measurement is selling ahead of the evidence.

Know what the panel can and cannot see#

Every figure in this article rests on clickstream panels, and they deserve their caveats out loud. Semrush describes its basis as our panel of over 200 million real, anonymized internet users across more than 190 countries and regions, and says it partners with hundreds of clickstream providers recording billions of events monthly. It also states that it receives terabytes of data from a panel of various third-party data providers every one or two days.

Two honest limits follow. Semrush publishes no margin of error for these estimates, and it has recently expanded its traffic data coverage by around 20 per cent, which sits underneath a year-on-year comparison in a way nobody has fully unpicked. Website traffic panels model, they do not count. Treat a 6.1 per cent aggregate as a direction, not a measurement.

Keep the working somewhere durable, too. A team that can show how its website traffic moved against its category, month by month, has an argument that survives a change of marketing director. One that recalculates the story each quarter from memory is starting the same conversation from scratch every time, in an ever darker undergrowth.

What a content team should actually change#

The temptation after a year like this is to chase the winner, pour everything into being cited by chatbots and treat the old channel as finished. That would be a mistake of timing. Referral volumes from AI surfaces are growing quickly and remain small, and building a business on them today is denning down in a burrow that is still being dug.

The steadier move is to stop treating organic search traffic as one number. Split it by surface, by intent and by whether the query could plausibly be answered without a visit. A recipe lookup and a procurement comparison were never the same asset, and only one of them was ever going to survive a summary box.

Practically, that means treating website traffic as a portfolio rather than a single line. Some of it is durable, some was always borrowed from a search engine's patience, and knowing which is which before the next algorithm change is the whole of the craft. The trail is there to be read if you take the time.

There is also a regulatory thread worth watching, because it may hand you data you currently cannot get. The Competition and Markets Authority secured a commitment that publishers can opt out of their content powering AI features in Google search, and its conduct requirement obliges Google to give publishers clear and detailed metrics on user engagement with their search content in search generative AI features.

That last obligation matters more than the opt-out. The central frustration of this whole year has been that nobody could see what AI surfaces were doing to their own traffic. A regulator has now written down that publishers are entitled to know, which is the first crack of moonlight in a very dark hedgerow.

Meanwhile the trend everyone is chasing is itself in flux. Similarweb reports ChatGPT's share of worldwide AI web traffic falling from roughly 76 per cent in June 2025 to about 53 per cent by May 2026 as Gemini gained ground. Optimising a whole content programme for one assistant, in a market redistributing that quickly, is the kind of confident error that costs a year.

So: measure against the category, not against your own best month. Split the surfaces. Keep the brand asset that no summary can replace. And if a traffic decline this year is genuinely a category effect rather than a failure of craft, say so plainly to the board, with the denominator in your hand. That conversation goes far better with evidence than with apology, and we are glad to help build it if you want a second pair of eyes.

Frequently asked questions#