For years a website was thought of as a place people arrive at, and everything we built around it assumed as much. The home page as a shop window, the navigation as a route, the conversion at the bottom of the funnel. The metrics of its success revolved around the same idea: sessions, bounce rate and time on page. All of that has now changed.
What it means to be a source
A destination exists so that people reach it. A source exists so that people use it, cite it and refer to it. That is what is happening more and more with the content of a website.
Rand Fishkin, founder of Moz, a tool that defined SEO for a decade, and now at the helm of SparkToro, has put it plainly: Google’s summaries, AI Mode, Gemini, Claude, ChatGPT, Perplexity and Copilot “all depend on the information they crawl from your website”. That it should be him saying it has a certain irony, because he has spent years dismantling his own industry’s fads, including much of what is now sold as AI optimisation. Your page is not visited: it is read, summarised and cited, and whoever receives that answer… may never set foot on your domain.
The underlying shift is confirmed by the Reuters Institute’s Digital News Report 2026, which, looking at 48 markets, found that news websites and apps have fallen 12 percentage points since 2020 as a way of accessing information, down to 51%, while social media and video reach 54% and overtake them, and television, for the first time. People enter sites less, but the content of those sites keeps circulating. How? It simply does so through others.
Yet nobody with any judgement is dismantling their website. What has changed is not whether it is worth keeping, but what it is for. According to the Content Marketing Institute, 32% of B2B marketing leaders plan to increase their investment in owned media during 2026, the third area with the strongest intention to grow, behind AI tools and events, and in large organisations the figure is 29%, tied with content personalisation and behind only AI tools.
Four figures that sum up the website's change of role: it loses weight as a front door and gains weight as an investment and as a reference.
−12 pp
is how far news websites and apps have fallen since 2020 as a way of accessing information, down to 51%
Reuters Institute, Digital News Report 2026
54%
of people access information through social media and video, which overtake news websites and television for the first time
Reuters Institute, Digital News Report 2026
32%
of B2B marketing leaders plan to increase their investment in owned media in 2026
Content Marketing Institute, 2026
77%
of the time, the vendor leading the buyer's preference list before first contact wins; it used to be 83%
6sense, Buyer Experience Report 2025
Don’t judge your website by visits alone
If your website has started working as a source and you keep judging it by visits, you are grading the new job with the old exam. Organic sessions, time on page or pages per session describe a destination reasonably well and describe absolutely nothing about a source, because they measure who arrived, not who used you.
And the consequence is hard to swallow: the better your website works as a source, the worse the metrics on your dashboard may look. Every time an assistant answers well thanks to your content and the user walks away satisfied, you record one visit fewer. Success and failure produce the same downward line, and from the dashboard they are indistinguishable.
The dashboard’s two blind spots
On top of measuring the wrong quantity there is a more awkward problem: the tool, set up as it is, has two documented blind spots.
The first concerns which page gets the credit. According to Similarweb’s analysis of the 2026 generative landscape, 65% of the URLs ChatGPT cites sit two or three levels deep, while 58.8% of the traffic that assistants refer as a whole lands on the home page. The page that earned the citation is almost never the one that records the visit, so in your analytics the article that did the work looks flat and the home page rises for no apparent reason.
The second concerns where people appear to come from. Cloudflare documented that traffic referred from Claude’s native app carries no Referer header, and adds that it assumes the same of other native apps. Without that origin, the visit enters your analytics as direct traffic, with no source. Nobody has rigorously measured how big that hole is, because the figures in circulation come from vendors selling the fix, but the mechanism is confirmed and points in a single direction: part of the discovery you generate today is invisible to your own organisation.
Put the two together and a conventional dashboard can show you direct traffic growing and organic falling, and let you conclude exactly the opposite of what is happening.
What to measure instead
The answer is not to stop measuring, which would be the worst outcome of all. It is to change the noun: stop counting arrivals and start counting use and demand.
Four indicators
Four indicators do that job reasonably well.
How many people search for you by name, how much direct traffic you receive and how both evolve against what you publish. Fishkin proposed a simple way to look at it: a single chart with direct traffic, visits from people typing the address, social referrals and branded search volume, to see whether they move together. He himself warns that correlation does not imply causation, even if it is a clue. It is less precise than last-click attribution and considerably more honest, because that never measured what it claimed to measure.
Your server logs say which bots come by, how often and to which pages. It is the most direct signal of whether your content is entering the circuit and it depends on no external tool, although the logs have to be read carefully: a bot can claim to be something it isn't, and checking which ones are genuine is a job for whoever administers the server.
You sample it by periodically asking what a client would ask and noting whether you appear, with one caveat that also works as a filter for any tool you are sold: SparkToro and Gumshoe repeated twelve questions between sixty and a hundred times each, and there is less than a one-in-a-hundred chance of getting the same list of brands twice, and less than one in a thousand of getting it in the same order. There are no positions here, only how often you appear across many repetitions.
If fewer people arrive, the state they arrive in matters more. The 2025 report from 6sense, covering some 4,000 buyers, finds that 94% of buying groups rank their list of vendors by preference before speaking to any of them, and that the one at the top wins 77% of the time, down from 83% in previous years. Translated to your contact form: it matters less how many open it and more how many arrive with the decision half made.
What still can’t be measured
There is no reliable way today to attribute a sale to a citation in an assistant, and it is worth saying so out loud before the first vendor who knocks on your door with a promise in hand says otherwise. There are hints: Similarweb observed that people exposed to AI recommendations are more likely to visit the brand’s site afterwards: 7.2% more in the case of American Express and 14.2% in the case of Capital One. But those are huge brands, the window is seven days and the author himself wonders whether those users would have arrived anyway. It works as a clue, not as proof.
That is why we would distrust any tool promising an exact share of voice in language models: with the variability that has been measured, that number cannot exist. What can be observed is a trend across many repetitions, which is considerably less sellable and considerably closer to what is actually happening.
The organisations that were already a source
For some organisations, none of this is a change: it is a recognition.
Neither a hospital nor a university has ever lived off visit volume. They have spent years measuring their website with indicators borrowed from the media because those were the ones available, while their value lay elsewhere: being the reference used to answer a question about a treatment, a programme or a procedure.
For them, being a source is not a metaphor but the literal description of what they have always done. What usually fails is not the authority or the content, but that the website holding them is not built for them to be found, read and cited.
Our position
At Digital Bakers we have spent years saying that what isn’t measured doesn’t improve, so we are not going to suggest now that you stop measuring because the ground has got harder. We propose the opposite: measure what your website does now, even if it is more laborious and less tidy than a sessions chart.
Your website is still the place where what you know lives; what has changed is that you no longer control where it is read. And a source is not judged by how many people come near it, but by how many times people drink from it.
A source is not judged by how many people come near it, but by how many times people drink from it.
If you want a starting point, look at your server logs for the last three months and check which AI bots pass through your website and which pages they take. It is information you already have and it costs nothing to consult. It is usually the first time an organisation sees, in black and white, that its website has been working as a source for some time without anyone having noticed or acted on it.
JM Sanchón, Content Director at Digital Bakers
Sources
- Rand Fishkin, SparkToro, Does Your Website Still Matter in the Zero-Click Era?, 14 August 2026. And How to Measure “Hard-to-Measure” Marketing Channels, October 2022, for the correlation method.
- Reuters Institute, Digital News Report 2026, 16 June 2026. 48 markets. News websites and apps: −12 percentage points since 2020, down to 51%; social media and video, 54%.
- Content Marketing Institute, B2B Content and Marketing Trends 2026 (October 2025, 1,015 respondents) and Enterprise Content and Marketing Trends 2026 (January 2026, 296 respondents at organisations with more than 1,000 employees). Intention to increase investment in owned media: 32% and 29% respectively.
- Similarweb, 2026 Generative AI Landscape Report. 65% of the URLs cited by ChatGPT sit two or three levels deep; 58.8% of assistant-referred traffic lands on the home page. Data from June 2025 to May 2026.
- Cloudflare, The crawl before the fall… of referrals, 1 July 2025. Traffic from Claude’s native app carries no Referer header.
- SparkToro and Gumshoe.ai, 28 January 2026. 12 queries repeated between 60 and 100 times across three tools, 2,961 answers: less than a 1-in-100 chance of getting the same list of brands twice; less than 1 in 1,000 in the same order.
- SparkToro on Similarweb data, 28 June 2026. Exposure to AI recommendations and likelihood of a later visit: +7.2% (American Express) and +14.2% (Capital One), 7-day window, large brands, no causation established.
- 6sense, B2B Buyer Experience Report 2025. 94% rank their list by preference before speaking to vendors; the leader wins 77% of the time, down from 83% in previous years.