Most of Your Website Traffic Was Never Human. I Checked.

The CBC ran a piece this week about what AI summaries are doing to the people who make things for the web. The example that stuck with me was Deb Perelman, who has run the food blog Smitten Kitchen for twenty years and now finds chunks of her own recipes appearing in the AI summary above the search results, published without her permission and without anyone needing to visit her.

Cloudflare’s own report is blunter than the coverage of it. Between June 2025 and April 2026, human traffic fell across every major industry by at least 35 per cent, and the most heavily crawled categories lost as much as 40 per cent in under a year. More than half of all internet traffic is now non-human. And the share of crawler requests made for AI training, not indexing, went from 22 per cent in spring 2025 to 52 per cent by June 2026.

One number in there explains who holds the power. Google accounts for roughly 88 per cent of referral traffic. Whatever arrangement exists between websites and search, it is effectively an arrangement with one company.

I spent thirty-three years in enterprise IT, twenty of them as Director of Computer Operations at Trader Joe’s, before I started writing full time. So when an article tells me a number has moved, my instinct is to go and measure the thing myself instead of taking the reported figure. This week I did, on my own site, and what I found was worse and more interesting than the article suggests.

What happens when you count properly

My site has been reporting between 400 and 1,800 visits a day and I could never explain the variation. A post that landed well and a post nobody read produced numbers I could not tell apart.

This week I put proper filtering in front of it. Not the usual exclusion of things that politely announce themselves as robots, but blocking traffic arriving from data centres: Azure, AWS, OVH, Hetzner, Linode, Alibaba and a dozen more. Real readers do not browse from a rented virtual machine in Frankfurt. They arrive from Comcast, from Verizon, from a phone on a train.

Roughly a third of what I had been counting came from one cloud provider’s address ranges. Another slice was my own site-crawling script, which uses an ordinary browser identity and had been inflating my figures for months. What remained was about half of what I thought I had.

Cloudflare’s own continental measurement puts it more starkly than my site does. Across North America, 70.5 per cent of requests for HTML content are machines. Fewer than three in ten are people.

Why the number stopped meaning anything

Here is what I keep thinking about. A shop counts people through the door because footfall used to correlate with sales. If a hundred people came in, some bought something, and the ratio held steady enough to plan around.

Now imagine seven in ten of those people are photographing your shelves for a catalogue you will never see, published by someone who will never mention where the pictures came from. Your footfall counter still works perfectly. It counts every single one of them. The number it produces has simply stopped having any relationship to whether you sold anything.

That is a website in 2026. The counter is fine. What it counts changed underneath it.

And the machines are not incidental visitors. They are how most people will encounter your work from now on, because they read it, summarise it, and hand the summary to someone who never arrives.

Does this mean writing for the web is finished?

No, and the doom version of this argument is as unhelpful as pretending nothing has changed.

What broke is one arrangement: you write something, a search engine finds it, the search engine sends you a reader, and some fraction of those readers become money. That deal held for twenty-five years and a lot of businesses assumed it was permanent.

It was never permanent. It was a distribution agreement with a company that owed you nothing, and it lasted exactly as long as sending you the reader served their purposes.

Mozilla’s president, Mark Surman, puts the worry precisely: if the incentive to make original independent work disappears, what is left comes from large companies, from governments, or from the AI itself. That is a real risk.

But the response is not to write less. It is to stop building on somebody else’s distribution.

What replaces traffic as the thing worth measuring

Citation replaces the visit.

When somebody asks an assistant about ghostwriting contracts, or book structure, or whether their story is worth telling, the answer gets assembled from sources. Being one of those sources is now the equivalent of ranking, and it does not produce a visit you can count. It produces a mention you mostly cannot see.

That sounds like a downgrade until you notice what it rewards. A search engine ranked pages. An answering system cites entities, and wants to know who you are and whether anything outside your own website confirms it.

Which is why the work I have been doing for the last year stopped being housekeeping and became the actual strategy. A Wikidata entry. Structured data that says the same thing on every page. Author identifiers that tie the writing to a person, not a URL. The same figures repeated in the same form everywhere a machine might look. None of that was worth much when search was sending readers, and all of it matters now that machines are doing the reading.

One marketer quoted in the CBC piece describes adding substantial text to the About page, and says plainly that people will not read all of it but the AI definitely will. He is right, and a year ago that would have sounded absurd.

What this changes for someone writing a book

Counterintuitively, it makes the book more valuable, not less.

A blog post is a claim on a page. A published book is a durable, citable artifact with an ISBN, a publication date, a named author and a record in library systems that no algorithm change deletes. When an AI system is deciding whether somebody knows what they are talking about, a book is one of the strongest signals available, and it keeps working whether or not anyone visits your website.

I have written 113+ books under my own name and ghostwritten 54+ more for other people. The ghostwritten ones taught me the part that matters here: those books work for their authors regardless of whether those authors have a website at all. The book is the asset. The site was always the shop window.

So the person most exposed to what is happening is not the author. It is the business that spent a decade building an audience on rented land, first Google’s, then a platform’s, and now finds both landlords have stopped forwarding the post.

What I would do about it

Four things, in the order I would do them.

Find out what your traffic really is. Not your analytics number, which counts whatever reaches it. Filter out data centre traffic and check your raw logs against your dashboard for a few days. You will not enjoy the answer and you need it, because every decision downstream depends on knowing which half was real.

Fix your identity before your content. Consistent facts about who you are, in structured form, everywhere a machine reads. Same name, same credentials, same numbers. This is dull work and it is the single most useful thing available right now.

Build the direct relationship. A mailing list, a podcast, a book in someone’s hands. Anything where the connection does not pass through an intermediary who can decide tomorrow to stop passing it on. This is the same argument I make to clients about tooling, for the same reason: own the part you cannot afford to lose.

Write the thing only you can write. A summary can replace an article that assembled public information. It cannot replace the eleven years you spent inside an industry, or the decision you made when payroll was uncertain. Whether your story is worth telling was never a question about search traffic, and the answer has not changed.

The choice Google is offering, and why it is not one

This part deserves more attention than it gets.

Google uses a single crawler to do two different jobs. It indexes your site for search, and it collects material for training and for the AI Overview that answers the question without sending anyone to you.

One crawler, two purposes, one switch. So a publisher who does not want their work summarised has exactly one lever available, and pulling it removes them from search entirely. Keep the arrangement that no longer sends readers, or leave the index that 88 per cent of referral traffic flows through.

That is not a choice. It is the shape of one, offered by the only party with anything to lose from you making it.

The UK is the exception, and it is worth watching. In June 2026 the Competition and Markets Authority required Google to give publishers a way out of AI Overview without being deindexed, and Google is trialling exactly that, with a promise that using the control will not affect ordinary search rankings. Which is an admission that separating the two was always possible and simply never offered.

Is anyone going to fix the economics?

Possibly, and the shape of it is already visible.

Wikipedia is the interesting case. Lane Becker, who runs Wikimedia Enterprise, describes a single bot pulling fifty articles in under a second and wanting every page in every language inside half an hour. That costs real money in servers, for content that is free to read. Wikimedia now sells bulk access to Google, Microsoft, Amazon and Meta, turning the scraping problem into a revenue line.

At the other end, publishers are considering the nuclear option. The CEO of USA Today said the traditional Google search business model is dead and signalled a willingness to block Google outright. Reddit and others are reportedly weighing the same thing. Cloudflare is moving too, and from September 15 it starts blocking multi-purpose crawlers by default unless site owners opt out.

A settlement will come out of that, and it will probably look more like licensing than like the free-crawl arrangement we have had. What it will not do is restore the referral traffic. That is gone regardless of how the payment question resolves.

The counter still works

I spent today deliberately making my own traffic numbers smaller, and it felt wrong the entire time. Years of watching a number go up, and I sat there removing most of it on purpose.

But the shop analogy holds all the way through. I was not losing customers. I was switching off a counter that had been including everyone who came in to photograph the shelves, and finding out how many people had been buying. That number is much smaller and it is the first one I have had in years that means anything.

The web is being rebuilt for machines. Human attention did not disappear, it moved somewhere that does not produce a visit you can count. So count something else, and put your work where being read by a machine still leads back to you. If you want the detail on how that side works, the writing hub collects the pieces on entity, structure and authorship, and my ghostwriting service page explains how a book gets built from what you already know.

The counter still works. Just stop believing what it tells you.

Frequently Asked Questions