The funnel moved: the buying decision happens inside the answer
The traditional funnel descends through pages and leaves a trail; the AI-answer funnel happens inside a single answer and appears in no analytics. A stage-by-stage comparison, with real observations in Meta's assistant and Amazon's Rufus.
In the traditional funnel, the stages of the decision are different places: an ad, a search, three sites open, a comparison, a conversion. In the AI-answer funnel, the middle of the funnel happens inside a single answer: the question goes in and the shortlist comes out, already made, with two or three brands. This article compares the two funnels stage by stage, with real observations from August 2026 in Meta's assistant and Amazon's Rufus, and explains why most of this funnel appears in no analytics tool at all.
The same purchase, two paths
Somebody wants to choose a private hospital in Lisbon. On the traditional path, they search Google, open the sites of two hospitals and a comparison site, read, close, come back days later, and every step leaves a trail: impressions, clicks, sessions, a pixel at every stage. Marketing teams have lived off that trail for twenty years.
On the second path, the same person asks a question: which is the best private hospital in Lisbon. On 23 August 2026, Meta's assistant answered that question in seconds with an assertion ("the consensus is quite clear") and a shortlist of three names, with ratings and a map. The consideration, the comparison and the shortlist, which on the first path were weeks and dozens of pages, happened inside one answer.
The first funnel still exists. The second grows every month, and it has three properties that make it a new problem, not a variation on the old one.
1. The shortlist stopped being built by the buyer
In the traditional funnel, the buyer builds their own consideration set, site by site, and the brand has dozens of chances to get in along the way: an ad, an organic result, a remarketing impression. In the answer funnel, it is the engine that builds the shortlist, from what it reads at that moment.
In the 23 August observation, the reasoning panel of Meta's assistant showed the process from inside: it consulted 26 sites to answer about private hospitals in Lisbon and cited 2 in the final answer. An article from a university medical faculty and a blog for expatriates. The answer asserted which is the best private hospital in Lisbon on the basis of those two pages, and neither of them belongs to a hospital.
The practical consequence: the work stops being "attract the buyer to my site" and becomes "be present on the pages that decide the answer". Those pages are identifiable, one by one. It is source work, not advertising work.
2. The funnel became invisible to whoever is inside it
Every stage of the traditional funnel is measurable because it happens on properties somebody controls. The answer funnel happens inside the conversation, and the brand's site only sees the end, when it sees anything at all.
The public numbers draw the picture: the click on Google's first position falls to less than half when there is an AI answer block on the page (Ahrefs, a study of 300,000 keywords, February 2026), and at the same time visitors arriving from AI assistants convert above organic traffic (NetElixir, 2025). Fewer clicks, better clicks: which is exactly what you expect from a funnel whose middle moved inside the answer and where only the decided buyer gets out of it.
Whoever measures this funnel with the old tools sees only a drop in traffic. The part that matters, who got into the shortlist and who was left out, appears in no analytics, because it happened on another company's screen. The only way to see it is to measure the answers themselves: ask the questions the buyers ask, record who appears, and repeat under the same conditions, week after week.
3. The stages turned into follow-up questions
The traditional funnel descends through pages; the answer funnel descends through follow-up questions. Rufus, Amazon's shopping assistant, shows this literally: it answered "best air fryer" with a shelf of products and proposed the next steps itself: "Which is the quietest?", "Compare Cosori vs Philips", "Best air fryer for a family of 4". The consideration and comparison stages did not disappear; they came to be suggested by the engine itself.
This changes what being present means. Appearing in the first answer and disappearing in the second is the dominant pattern: in a 2026 study of 69,120 B2B conversations (Clovion), 62% of the brands in the first answer disappear by the buyer's second question. The presence that is worth money is the one that survives the follow-up, and that is only measured by walking the chains of questions, not isolated ones.
4. The bottom of the funnel is moving inside the conversation
In the traditional funnel, the conversion happens on the brand's site. In the answer funnel, the shop window is already inside the conversation: ChatGPT shows products with prices and ratings, Google AI Mode shows product panels in Portuguese, and in the United States payment already happens inside the chat on three platforms. The shelf and the till merge with the answer. In Portugal, today, the shop window exists; the payment comes later, and whoever is in the window when it arrives will have no ground to make up.
5. The funnel gained a memory
The traditional funnel restarts with every search. The answer funnel feeds itself: today's reviews, forum discussions and press coverage are the raw material of tomorrow's answer. A recurring complaint about waiting times, written this month in public reviews, is an unfavourable engine answer in the next cycle. Next quarter's top of funnel is being written now, by third parties, on pages the brand does not control but can know.
What this changes in practice
- Measure the answers, not just the traffic. The new funnel is only visible to whoever asks the engines what the buyers ask and records who appears. AI traffic in GA4 is the tail end; the shortlist is decided before that.
- Work the sources, not just the site. The pages that decide each answer are identifiable. Being on them is worth more than any isolated internal optimisation.
- Survive the follow-up. Question two is where the shortlist is lost. Chains of questions can be measured; isolated questions mislead.
- Treat reputation as top of funnel. What customers write today is what the AI answers tomorrow. Operations and visibility have stopped being separate departments.
The sentence that sums up the argument: the funnel the teams measure ends where the new one begins, inside the answer. That is where we measure.
Frequently asked questions
Is the traditional funnel dead?
No. It still exists for whoever searches in classic Google and clicks. What changed is that a growing share of decisions now runs through a parallel funnel, inside AI answers, where the consideration and shortlist stages happen in a single answer. The two funnels coexist; only one of them is visible in traditional analytics tools.
How do you measure a funnel that happens inside ChatGPT?
By measuring the answers themselves: asking the questions the buyers ask, recording which brands appear, in what position, with which sources, and repeating every week under the same conditions. The traffic arriving at the site from AI assistants is measured in GA4, but that is only the tail end; most of the funnel never generates a click.
What decides who gets into the AI's shortlist?
The sources the engine reads at the moment of the answer, and the model's memory. In an August 2026 measurement about private hospitals in Lisbon, Meta's assistant consulted 26 sites and cited 2: an article from a university and a blog for expatriates. Neither of the pages belonged to the recommended hospitals. Whoever controls those pages controls the shortlist.
Sources for the numbers: Ahrefs (300,000 keywords, February 2026), NetElixir (2025), Clovion (69,120 B2B conversations, 2026), and our own observatory's observations on meta.ai and on Rufus at amazon.es (23 August 2026, dedicated account, real session). Vendor data treated as directional.