The Research Didn't Stop. The Click Did.
Two numbers frame this. Forrester's Buyers' Journey Survey for 2024 found that 89% of B2B buyers were using generative AI in at least one area of their purchasing process. And when Pew Research Center tracked the real browsing of 900 US adults in March 2025, people who were shown a Google AI summary clicked through to a traditional result on 8% of visits. People who were not shown one clicked on 15%.
Sources: Forrester, Buyers’ Journey Survey, 2024 (as reported by Forrester, November 2024). Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, July 2025.
Pew's study was of consumers using Google, not B2B buyers using a chat assistant, so do not read it as a forecast of your traffic. Read it as a direction. When an answer arrives on the page, fewer people go looking for the source, and the ones who do rarely click the links the answer cites. Only 1% of visits to a page with a summary included a click on a link inside it.
Put those two together and the shape is clear. The early research is still happening, probably more of it than before, because asking a chat assistant is cheaper than reading five articles. It just happens somewhere you cannot see, and it ends without anybody arriving on your site.
Your Content Now Gets Read by a Machine First
The old top of funnel had a simple path. Someone searched a problem, clicked an article, read it, maybe came back, and eventually filled in a form. Every step left a trace in your analytics. A lot of that path now runs through an assistant that reads the web on the buyer's behalf and hands them a summary.
Two things follow. First, the summary is built from somebody's material. If your thinking on a problem is clear, specific and published where it can be read, there is a fair chance it shapes the answer, even when nobody clicks. Second, the summary flattens. Ten articles that all say the same thing come out as one sentence, attributed to nobody. Content that only restates the category is now mostly raw material for an answer that does not mention you.
So the question for the top of the funnel is no longer only “did they read our article?” It is also “when they asked about this problem, did our way of seeing it come through, and did our name?”
When Explainers Cost Nothing, a View Is What's Left
The same tools that answer buyers' questions also write the content. Anyone can now publish a competent explainer on any B2B topic in an afternoon, and many teams are doing exactly that. The predictable result is a web full of articles that say the same thing in the same order, which is precisely the kind of material an AI answer absorbs without credit.
Look at what that does to the economics. A generic “what is X” article used to earn traffic because few people bothered to write a good one. Now it is free to produce, it competes with a summary that answers the question before anyone clicks, and it sounds like everyone else's. It has lost most of what made it worth writing.
What keeps its value is the part a model cannot generate on its own:
- A position someone could disagree with. “Your cost per lead is rising because your whole category is bidding for the same buyers” is a view. “Lead generation is important for growth” is filler that any tool will write for free.
- Numbers that came from your own work. A benchmark from your own customer base, a before-and-after from a real programme, a cost you have actually seen. Original data is something an answer has to take from you.
- Worked examples. The specific sequence of what was done, in what order, and what happened. It is harder to flatten into a sentence than a definition is.
On our own demand generation work, the creation side has always been education, point-of-view content, video and executive presence aimed at the problem the product solves. None of that changes. The bar for what counts as point of view has gone up, because the baseline is now a machine that can produce the obvious version instantly.
Named People Matter More, Not Less
Once anything can be generated, a reader's first question becomes “who is actually saying this?” A founder or senior leader who posts under their own name, takes a position and answers questions in public is the one source an AI summary cannot replace. A buyer can follow them, watch them on video, hear them on a podcast, and form a view of whether they know what they are talking about.
That is also where a lot of early B2B research has always happened: in feeds, in communities, in conversations between peers. None of those is a search result, and none of it goes through an AI summary first. A buyer who has watched your founder explain a problem clearly three times over a quarter is likely to know the name when the shortlist gets written.
Two cautions. A ghost-written post with nothing at stake reads like what it is, and a feed full of AI-drafted posts reads the same way. And it only works if the person actually holds the views. The point is to put a real perspective in front of the market, under a real name, often enough to be remembered.
AI on Your Side of the Funnel
Most of the conversation about how AI is changing top of funnel demand generation is about buyers. The other half is what it does for the team doing the marketing, and here the honest answer is mixed.
| Where AI helps | Where it misleads |
|---|---|
| Researching accounts before outreach: what a company sells, who it hires, what it has announced | Writing the outreach itself, which comes out sounding like everyone else’s |
| Sorting large volumes of first-party signal: repeat visits, pricing-page views, email engagement | Treating an AI-scored “in-market” label as proof an account will buy |
| Drafting variations of a message for different roles, for a person to edit | Publishing those variations unedited, at volume |
| Summarising call notes and what sales has heard, so marketing hears it too | Replacing the conversation with sales that should have happened anyway |
The rule we hold to on signal work applies here too. Intent data is directional. It tells you where to look, not what will close, and it is valuable as prioritisation and wasteful as qualification. An AI model that scores accounts is still a way of deciding where to look first.
And before buying any of it, use what you already have. First-party signals, your own site, CRM, email and sales engagement data, are under-used at most companies and cost nothing extra. A tool that reads them faster is useful. A tool that replaces them with a guess about the whole market is not.
Measuring a Top of Funnel You Can't See
This is the part of how AI is changing top of funnel demand generation that costs teams their budgets. If the top of the funnel was judged on blog traffic and content downloads, and those numbers are falling, the obvious conclusion is that the content stopped working. Sometimes it has. Often the education is still landing, just inside answers and feeds that never send a visit, and the numbers that would show it are not on the dashboard.
The measures we use for the creation side of demand generation were never built on clicks, which is why they hold up better now:
- Branded search volume over time. If more people search your name, more people know it, wherever they learned it.
- Direct traffic to key pages. People typing you in after an answer or a conversation, not clicking an ad.
- Self-reported attribution. A free-text “How did you hear about us?” field on the form. It is now one of the few places a buyer will tell you an AI assistant recommended you.
- Sales acceptance rate and win rate by source. Buyers who arrive already educated tend to be better qualified, and these two show it.
- Cost per opportunity, not cost per lead. Leads get cheaper and worse together far more often than they get cheaper and better.
One more thing changes how you read those numbers. Creation compounds over two to four quarters and shows up first as branded search and inbound quality, before it shows up as pipeline. Falling top-of-funnel traffic inside that window is not, on its own, evidence the work is failing. Agree that with whoever owns the budget before the first quarterly review, not during it.
The reporting rules do not change either: platform-reported values labelled as such, the attribution window stated on every view, early-stage data flagged as early-stage, and no adding up of numbers from different platforms and calling the total revenue.
What to Change This Quarter
- Ask the answer engines. Type the five questions your best buyers ask before they buy into the assistants they use. Read what comes back. Note whose framing you hear, and whether your name appears. Do it again next quarter and compare.
- Retire the generic explainers you were planning. Spend that budget on one piece with a real position or your own data in it.
- Put a named leader on a schedule. Two posts a week from a person who holds the views beats ten anonymous articles.
- Add the free-text attribution field to every form that matters, and read the answers every month.
- Move branded search and direct traffic onto the main dashboard, next to pipeline, and take blog sessions off it as a headline measure.
- Use AI where it saves research time, on accounts and signals, and keep a person on anything a buyer reads.
