For twenty years, a website had two readers: people, and Google reading on their behalf. You designed for the first and structured for the second, and the two rarely argued.
Now there's a third. ChatGPT, Gemini, Perplexity, and Claude read your site on behalf of a buyer who may never actually arrive. They don't see the hero image, and they don't notice the animation. They just extract facts, compare them with facts from your competitors, and hand a summary to somebody who will use it to make a decision.
How Much Has Already Moved
The numbers are further along than most boards think.
TrustRadius found that 90 per cent of B2B buyers used tools like ChatGPT in their research last year, and 72 per cent encountered a Google AI Overview during evaluation. Google's AI Mode passed a billion monthly users within a year of launch. And Pew measured what happens next: on visits where the search page shows an AI summary, people click a result just 8 per cent of the time, compared to 15 per cent when there isn't one. Only 1 per cent click a source inside the summary itself.
Scott Brinker calls buyer-side agents "the real disruption". Most AI in marketing works for the marketer, where the gains are mostly incremental, whereas this works for the customer and changes who gets found in the first place. His research with Frans Riemersma found that while 63 per cent of marketing teams are adapting their content for the shift, only 14 per cent are set up to measure it.
For a while, at least, that gap is the opportunity.
Why the CMS Is Growing Again
In Brinker's 2026 landscape, content management and web experience was the fastest-growing category of all, up 21.4 per cent in a year after a decade of being treated as plumbing. Ecommerce platforms came second. As The Martech Weekly noted, the website now serves "six or seven audiences", agents among them, and they read Salesforce's move for Contentful as one of the largest suite vendors conceding that content has to be available through an API, without the templates, because the newest readers don't have a screen.
That's the same argument that made headless content management sensible for mobile a decade ago. The difference now is that the mobile app was yours, and the agent isn't.
What an Agent Actually Reads
An agent is a poor audience for brand, and a good audience for facts. It wants to know what you sell, what it costs, where it's available, what the terms are, and whether the claim on page four agrees with the claim on page nine. When your facts disagree, it doesn't ask which one is current. It just picks one, or else it picks a competitor whose facts do agree.
Tony Byrne's MarTech 2030 treats this as a core job: structuring product specifications, compliance information, and pricing data so that customers' AI agents can read them. Real Story Group goes further and describes what they call a "content warehouse", where every claim is stored with its approval state, its source, its expiry date, and the uses it's permitted for, so that when a sentence leaves your site it carries its own provenance. As they put it, "The sentence cannot safely travel alone."
Almost nobody has one of these yet, and RSG is clear that you won't buy one as a finished product. Instead, you have to extend the DAM, extend the CMS, or build a content services layer across both. All three routes start from the same place, which is content modelled properly, with one source of truth for the facts that matter. The CMS was never the problem. The content model was.
An agent doesn't admire your homepage. It checks whether your facts agree with each other, and leaves if they don't.
Don't Rebuild for a Visitor Who Isn't Buying Yet
Still, there's a case for calm. The same research showing that agents are everywhere in the research phase also shows people keeping hold of the purchase itself. Bain found that only 7 per cent of consumers would trust a general AI platform to shop end to end, against 25 per cent for retailers. Gartner puts the share willing to let AI make a purchase decision at 11 per cent. Or as Jon Miller put it on Brinker's blog about B2B: "Agents compress research time, but people sign contracts."
And the rules are still unsettled. A German court held Google liable this year for false statements in an AI Overview about two publishers. The Martech Weekly's reading was that a mention in an AI answer isn't automatically good for you, and that the value of optimising for these engines could change on a single ruling.
So the sensible position isn't a rebuild for the agent. It's a site that happens to be legible to one, because it was built on structured content and consistent facts, which were worth having anyway.
Five Questions Before the Next Rebuild
Is our content modelled as things or as pages?
Are your products, services, locations, people, prices, and claims stored as structured entities, each with one home, or are they just paragraphs typed into templates?
Do our facts agree everywhere?
That means price, availability, specifications, and making sure the claim in the brochure matches the claim on the site. If three teams can each edit the same fact in three places, an agent will find the disagreement before you do.
Could the site serve its content without its templates?
In other words, could another system consume it directly through an API? If the honest answer is no, you have a brochure, not a content platform.
Do we publish structured data and keep it current?
That includes schema markup, product feeds, and a sitemap that reflects what's actually live. Agents read what's machine-readable first.
How will we know it's working when the click disappears?
You'll need to look at branded search volume, direct traffic, and holdout tests. The Martech Weekly's advice is to borrow brand-lift methods, because click attribution isn't coming back.
The Short Version
Your website has a third reader: AI agents acting for buyers who may never visit. They read structure and facts, not design, which is why content management is the fastest-growing category in martech again. Don't rebuild for the agent. Instead, build on a proper content model and a single source for the facts that matter. Do that, and the agent will be able to read it, and so will everyone else.