Google is not going to send you the traffic it used to. That is no longer a prediction. It is arithmetic on numbers that have already been published.
The Association of Online Publishers ran an Artificial Intelligence Publisher Impact Study across eight UK publishing groups, covering 10.8 billion pageviews over six months. Organic Google referrals fell 7.1% per quarter. Hold that rate steady and pageviews delivered by Google are cut in half by the third quarter of 2027.
That single figure is the most useful planning input a small publisher has right now. Not because it is precise, but because it turns a vague fear into a schedule. If half your search traffic disappears within roughly two years, every decision about what to publish next quarter looks different.
What the numbers actually say
The decline is not one study having a bad year. It shows up everywhere the data is measured.
More than 60% of Google searches now end without a click, and in the first four months of 2026 the zero-click rate hit 68%. When an AI Overview appears above the results, 83% of those queries end with no click at all. In AI Mode, the figure is 93%. That is the mechanism behind everything else in this article: the answer is delivered on the results page, so the visit never happens.
Chartbeat data across more than 2,500 news sites, published by the Reuters Institute, shows Google Search referrals down 34% between December 2024 and December 2025. Google Discover fell 21% globally and 29% in the United States over the same window.
Size matters, and not in your favor. Sites doing 1,000 to 10,000 daily pageviews lost about 60% of their Google traffic across two years. Mid-size sites at 10,000 to 100,000 lost 47%. The smaller you are, the harder this has already hit you.
One honest caveat: these figures come from different studies in different markets, and your own numbers will not match any of them exactly. The direction is consistent across all of them, which is the part worth planning around.
The content that dies first
This is where the AOP study gets uncomfortable, because it breaks the decline down by what you publish.
Guides and how-to content took the worst of it, down more than 20%. Technology lost 28%, finance and business 23%, politics 7%. Meanwhile opinion and commentary traffic rose 4%, interviews rose 1%, and news sessions rose 2%.
The pattern is easy to read. Anything a model can answer completely from what it already knows is being answered without you. Procedural content is the most answerable content there is. The moment somebody asks how to do a thing, the machine explains how to do the thing, and there is nothing left for a click to add.
Yes, this article is a guide. We are describing our own category on purpose. The difference between a guide that survives and one that does not is whether it contains something the model cannot produce on its own: numbers from a specific study, a decision someone actually made, a result someone actually got.
If your traffic model is "we rank for how to do X", that model is first in line.
Discover is not the rescue plan
Plenty of sites treated Discover as the hedge. When search softened, Discover still fired, so the answer was to write for the feed. That escape route closed this year.
In 2026 Google shipped its first core update aimed only at Discover, running from 26 January to 2 March. The stated goals were less clickbait, more locally relevant content, and more original in-depth reporting. The measurable result was a narrower field: the number of unique domains appearing in the top 1,000 US Discover placements dropped from 172 to 158. Local news publishers gained. Niche sites covering technology, music and film lost clicks and impressions.
So Discover is now a second algorithm that can independently drop you, on top of the one already shrinking. It is a distribution channel worth having, not a hedge against search.
AI referrals are real, and they are tiny
The optimistic story is that traffic moves from Google to ChatGPT and the rest. Half of that story is true.
AI referral traffic grew roughly 16 times in two years, and it converts far better than organic search, by a factor of four to twenty depending on whose measurement you use. Buyers arriving from an AI answer have already had their question answered, so they arrive late in the decision rather than early.
The other half of the story is the volume. AI platforms account for about 0.32% of all website traffic, up from 0.24% a year earlier. ChatGPT is roughly three quarters of that sliver, with Gemini and Perplexity splitting most of the rest.
Do the arithmetic before you build a strategy on it. A channel worth a third of a percent of the web cannot backfill half of your search traffic by 2027. Treat AI referrals as small, high-quality and growing, and track them separately in your analytics so you know when that changes. Do not treat them as the replacement.
Where the effort is actually worth spending
Given all of the above, here is where we would put the work, in order.
Own the way people reach you. Email, a Telegram channel, push, an app, a community. Anything where you can reach a reader without an intermediary deciding whether they see you. The metric that matters is the share of your sessions that arrive direct or returning. If that number is under 10%, your business is a rented audience.
Publish what a model cannot generate. Original data from your own operation, tests you ran yourself, interviews, and clear opinion. Those are the exact categories that grew while everything else fell. Not a coincidence: they contain information that did not exist until you published it.
Get specific and local. Discover now rewards local relevance and Google says so openly. Broad international coverage of a general topic is the most crowded, most substitutable content on the internet.
Optimize to be quoted, not clicked. Put your claims in plain sentences, with dates, sources and numbers attached. A model that reproduces your finding and names you builds recognition even when nobody clicks. This is uncomfortable, and it is still better than being neither read nor cited.
Build revenue that does not scale with pageviews. Paid products, services, memberships, affiliate deals that follow fewer but better visitors. If every dollar you earn requires another thousand sessions, the trend line above is a countdown.
Measure AI traffic separately. Split chatgpt.com, perplexity.ai, gemini and copilot out of your referral report today, so that in a year you can see whether the sliver became a channel.
And what we would stop doing: chasing keyword volume with how-to posts, thin roundups, and rewrites of other people's news. Those pages compete directly with the machine that now sits above the results.
The honest summary
Search will keep sending traffic. It will send roughly half of what it sends today, sooner than most publishing plans assume, and it will send less of it to small sites than to large ones.
Nothing in the data suggests a clever trick reverses that. What the data does suggest is that the sites still standing in 2027 will be the ones whose readers can find them without asking a search engine first, and whose pages contain something that was not already in a model's training data.
That work is slower than SEO was. It is also the only part of the stack nobody can update out from under you.