AI has not replaced the buying path. It has compressed the middle of it and moved that middle somewhere you cannot see. Buyers arrive already researched, already narrowed and sometimes already wrong about you. The stage that decides everything is shortlist formation, and it now happens inside a conversation you were not part of.
The sector-specific version of this argument for regulated businesses is how AI is changing the B2B buyer path. This page is the general map and the measurement problem.
The old shape and the new one
The traditional model, whatever you called it, assumed a visible sequence. A buyer searched, read something, came back, compared, downloaded, subscribed, enquired. Each step left a trace, and the traces were what marketing optimised.
What happens now, increasingly:
- Buyer describes a problem to an assistant in their own words, often at length
- Assistant asks clarifying questions and narrows the requirement
- Assistant returns two or three options with reasons attached
- Buyer asks follow-ups, comparing the options inside the same conversation
- Buyer visits one or two sites, briefly, to verify rather than to learn
- Buyer contacts one
Steps 1 to 4 are invisible to you. They are also where the decision effectively gets made. By step 5 the buyer is verifying a conclusion, not forming one.
Two things follow, and they are both uncomfortable.
Being absent from step 3 removes you entirely. Not disadvantaged, absent. There is no position eleven in a generated answer.
Your website’s job changed. It used to teach and persuade over several visits. Now it frequently gets one short visit from someone who already believes something about you, and its job is to confirm or correct it fast.
Where you can still exert influence
You cannot get into the conversation. You can change what the conversation has to work with.
Before the conversation: be in the source pool. Everything on this site about citation and corroboration is really about this one thing. It is decided months earlier by whether you are resolvable, extractable and corroborated. See AI visibility strategy for the sequence and ecosystem validation for the off-site half.
During the conversation: answer the qualifying questions. Assistants narrow by qualifiers. Budget, size, sector, constraint, timescale. If your published material answers “who is this right for and who is it wrong for”, you get matched precisely. If it says you serve everybody, you get matched vaguely and dropped when the qualifier tightens.
At the visit: confirm fast. The verification visit is short and it usually lands on a deep page rather than the homepage. Every page needs to work as an entry point, with the answer near the top, the proof visible and a route to contact that does not require navigation.
At the enquiry: capture the source. One field. It is the only reliable measurement you have.
The measurement problem, stated honestly
This is where most articles on the subject become vague. It is worth being precise instead, because the vagueness is what leads teams to report numbers they should not trust.
GA4 under-reports AI referrals, badly. Many assistant surfaces strip or omit the referrer. Some traffic arrives as direct, some as unassigned, and none of it is labelled. Build the channel group covering chatgpt.com, chat.openai.com, perplexity.ai, claude.ai, copilot.microsoft.com, gemini.google.com and you.com, and then treat the number it produces as a floor rather than a count. The growth of unassigned traffic is itself a clue, which has its own page: why is my unassigned traffic growing in GA4.
Attribution models are misleading here. Last-click credits the final touch, which is often direct or brand search, because the buyer already knew your name by then. First-click credits whatever happened before the assistant conversation, which may be nothing. Neither sees the stage that mattered.
What actually works, in order of reliability:
| Method | What it tells you | Cost |
|---|---|---|
| A source question on the enquiry form | The clearest signal available. Add “how did you come across us” as free text | Minutes |
| Sales asking how the shortlist was built | Qualitative, immediate, and it improves the call anyway | Nothing |
| Citation testing on a fixed prompt set | Whether you are in the pool at all, monthly, same prompts | An hour a month |
| Brand search volume | Rises when assistants name you, because people then search the name | Free in Search Console |
| Direct and unassigned traffic trend | Indirect, noisy, but directionally useful alongside the rest | Free |
The full metric treatment is in how to measure AI search visibility and GEO metrics.
Falling traffic is not automatically a problem
Worth saying plainly, because it causes bad decisions.
Assistants absorb informational queries. The visits you lose are disproportionately the ones that were never going to convert: definitions, how-it-works, early curiosity. What remains skews later stage and better qualified.
So a site can lose 30% of its sessions and increase enquiries, and a dashboard measuring sessions will report a crisis. Before concluding anything, split the trend by intent. If informational pages are down and commercial pages are flat or up while conversion rate rises, the mix improved. That is the argument in zero-click search and why citations matter.
The genuine risk is different and worth watching: if you monetised informational traffic through advertising or volume, that model is under real pressure and no amount of AI visibility work restores it.
What changes on the website
Five practical consequences.
Every page is a landing page. Verification visits land deep. Each page needs the answer near the top, a visible route to contact, and enough context to stand alone.
Answer the disqualifying questions. Minimum engagement size, sectors you do not serve, what you will not do. This feels like turning business away and it does the opposite, because it gets you matched accurately by systems that narrow on qualifiers.
Proof has to be on the page, not a click away. A buyer verifying a recommendation is checking whether the claim survives contact. A case study behind a form does not exist for that purpose.
Publish prices, or a range, or the basis. The most common verification question, and one of the most common reasons a shortlist gets cut. MarGen publishes its own; five of the seven UK agencies we track publish nothing.
Nurture sequences assume a path that fewer people take. They still work for the people who enter them. They are no longer where the consideration stage happens for most buyers.
The sales conversation
Small changes, disproportionate return.
Ask early how they built their shortlist and whether they used an assistant. Most people say so readily, and increasingly they say which one.
Expect firmer prior views. A buyer who has spent 20 minutes with an assistant arrives with a formed position. Sometimes it includes something inaccurate about you, taken from a stale source. Correcting that early in the call is easier than discovering it at proposal stage, and it tells you which source needs fixing. That diagnosis links straight back to AI giving wrong information about my company.
Feed it back into the content programme. The questions that come up repeatedly on calls are the questions assistants are being asked, and they are usually more specific than anything a keyword tool will surface. That loop is the practical version of prompt cluster research.
AI-Powered Customer Journeys: Common Questions
How has AI changed the customer buying path?
It compressed the middle and hid it. Buyers used to move through several searches, several sites and a comparison stage you could see in analytics. Now an assistant does the research and returns a shortlist, so the buyer arrives already informed and already narrowed, and the stages where you used to influence them happened somewhere you cannot observe.
Can you track a customer path that goes through an AI assistant?
Only partially, and pretending otherwise leads to bad decisions. Many assistant surfaces strip the referrer, so GA4 under-reports AI referrals badly and should be treated as a directional floor. The reliable signals are a source question on your enquiry form, sales asking how the shortlist was built, and citation testing on a fixed prompt set.
At what stage do AI assistants influence buyers most?
Shortlist formation. That is the stage where being named or absent decides everything downstream, because a business missing from the shortlist is never compared, never visited and never contacted. Earlier awareness and later negotiation are less affected. This is why citation share matters more than any traffic number.
Does AI search reduce website traffic?
Often yes, and that is not automatically a loss. Assistants answer the informational queries that produced browsing traffic, so what remains is fewer, later-stage, better-qualified visitors. Judge it by enquiry quality and conversion rate rather than sessions, because sessions will fall while revenue can rise.
How do I influence a buyer I cannot see?
By supplying the material the assistant uses. If your business answers the qualifying questions clearly, states its limitations honestly and is corroborated by independent sources, you get included in the shortlist the buyer is handed. You are not influencing the person directly, you are influencing what is available to the system advising them.
What should change on my website because of this?
Assume arrival at any page by an already-informed buyer. That means every page needs to stand alone, carry a route to conversion, and answer the qualifying questions rather than assuming a nurture sequence. The linear funnel where visitors enter at the top and progress is now the exception rather than the rule.
Do sales teams need to change how they qualify?
Yes, and the fix is small. Ask how the shortlist was built and whether an assistant was used. Buyers arrive with a firmer view formed elsewhere, sometimes including inaccuracies about you, so surfacing the source early lets you correct the record and tells you which of your content is doing the work.
Is the traditional marketing funnel dead?
No, but the middle of it is now often invisible and much shorter. Awareness and decision still exist. The consideration stage that marketing teams built content programmes around increasingly happens inside an assistant conversation, which changes where the work goes rather than removing the need for it.
Where to Go Next
The measurement half: how to measure AI search visibility · GEO metrics · zero-click search and citations
Get into the shortlist: AI visibility strategy · ecosystem validation · AI brand trust signals
Related: how AI is changing the B2B buyer path in regulated sectors · conversational marketing AI · search intent matching
Have it delivered: MarGen is a UK GEO agency with published pricing, a published method and documented results.