AI may shape the shortlist before you see a click

Your analytics may miss the first decision

Imagine a procurement director looking for a specialist supplier. She asks an AI assistant which vendors can meet a particular technical requirement, operate in two countries and support a tight implementation window. The assistant gives her three names, summarises their apparent strengths and flags one concern about each.

She then visits two websites, downloads a case study and asks a colleague for a reference. A week later, one of those suppliers sees an ordinary direct visit in its analytics and a normal opportunity in its CRM.

The AI may have made the first commercial cut, but almost none of that influence is visible in the referral data.

That is the emerging measurement problem. If AI becomes part of how buyers research, compare and interpret suppliers before contact, the number of clicks arriving from ChatGPT, Gemini, Claude or another assistant may be a useful metric. It is not the same thing as the number of buying decisions those systems influenced.

The referral number is not the influence number

The strongest independent evidence currently argues for restraint rather than hype. A 2026 Marketing Science paper analysed 12 months of first-party data from 973 e-commerce websites with a combined $20 billion in revenue. The researchers compared more than 50,000 ChatGPT-referred transactions with 164 million transactions from traditional channels. ²

Across the August 2024 to July 2025 period, organic LLM traffic represented less than 0.2 per cent of visits. Its conversion rate and revenue per session were better than paid social, but lower than every other traditional channel in the study. ²

That is a useful correction to claims that AI search has already replaced conventional discovery. It had not, in that dataset. The study is also narrower than the management question. It measures people who clicked from a large language model into an e-commerce site. It cannot measure a buyer who used an AI assistant to build a shortlist and later arrived by direct navigation, branded search, a colleague's link or a salesperson.

The commercial implication is therefore slightly awkward: direct AI referrals can be small at the same time as AI influence is growing upstream.

B2B buying already shows the split

Gartner's May 2026 B2B buyer survey makes the distinction more visible. In a survey of 645 buyers, conducted in August and September 2025, 45 per cent said they had used generative AI during a recent purchase, mainly to gather information about vendors and products. Buyers reported using an average of seven information sources. ³

Yet 69 per cent said they preferred to validate AI-generated insights with sales representatives. At the same time, 67 per cent preferred a rep-free experience and 70 per cent preferred a completely digital, self-service buying experience. ³

Those numbers come from a consultancy with a commercial interest in advising sales leaders, so they should not be treated as neutral market law. But the pattern is plausible and commercially important: AI can move into the early research and filtering stages without making the human seller irrelevant.

The seller's role can become more concentrated. Instead of repeating public information, the valuable interaction is increasingly about validation, trade-offs, exceptions, risk and confidence.

The market data disagree, which is itself a signal

Later retail data look much stronger than the peer-reviewed 2024 to 2025 picture. Adobe Digital Insights reported that, by March 2026, AI-referred visitors to US retail sites converted 42 per cent better than non-AI traffic. Adobe later reported a 60 per cent conversion advantage in July 2026. Adobe's 29 October update said that by September 2026, AI-referred US retail visits converted 90 per cent better and generated 83 per cent more revenue per visit than non-AI traffic. These are Adobe's own observations, not a universal benchmark. ^{4 8 9}

That does not invalidate the Marketing Science paper. The periods are different, the data sources are different and the definitions are not identical. Adobe is also commercially interested: it sells analytics and AI-visibility products to the organisations that might act on these findings.

The safe conclusion is not that AI referrals are weak, nor that they are universally superior. It is that the channel is changing quickly enough for a single headline conversion number to age badly.

For management, volatility matters. Budgeting around one vendor benchmark is premature. Building the ability to observe what AI systems say about your business, how buyers use those systems and how influenced demand later appears in your pipeline is not.

Search is becoming a representation problem as well as a ranking problem

The infrastructure is moving in the same direction. Google has added AI-led commerce features across Search, Gemini and Maps and has introduced Merchant Center tools intended to show brands how they perform in AI-driven search. Anthropic launched blueprints for retailers to build shopping and merchant agents in early September. European regulators are already examining how Google's AI-search opt-out affects publishers and competition. ^{5 6 7}

This is not evidence that autonomous purchasing is mature across the economy. Research, comparison and recommendation are further along than delegated purchasing, and many buyers still want human validation before a consequential decision.

But the management problem is already here. A business is increasingly represented to a prospective customer by machine-generated summaries assembled from public facts, third-party material, reviews, structured data and other sources before anybody from the company has spoken.

Traditional SEO asked, among other things, whether you could be found. The AI-mediated version adds another question: if you are found, what story does the system construct about you?

Do not reduce the problem to GEO

A new market has quickly formed around generative engine optimisation, answer engine optimisation and AI visibility. Some of it will be useful. Some will simply repackage familiar content work around a new dashboard.

The managerial objective should be broader than maximising mentions in AI answers. The real requirement is machine-legible, current and verifiable market information that still differentiates the business for a human buyer.

That means getting the boring things right. Product and service facts should be consistent across important public sources. Availability, geography, technical capabilities and pricing logic should not contradict one another. Claims of differentiation should be supported by evidence that can survive scrutiny. Third-party sources should not tell a materially different story from the company website without somebody noticing.

It also means resisting a new vanity metric. An 'AI visibility score' is interesting only if it helps explain consideration, qualified demand, pipeline or revenue. If a consultancy cannot show how its score connects to buyer behaviour, the metric may be optimising the machine's answer rather than the commercial outcome.

The salesperson becomes a validator of the machine's story

If AI handles more of the buyer's preliminary information gathering, sales enablement should change with it.

A salesperson who opens by reciting features that the buyer has already compared is adding little. A salesperson who can explain why a recommendation is incomplete, distinguish two apparently similar implementation options, provide evidence for a disputed claim or resolve an exception is doing work the buyer still values.

This suggests a different form of sales readiness. Teams need concise evidence packs, implementation boundaries, comparison logic, objection evidence and clear routes for technical or commercial validation. They also need to know what the major AI assistants are saying about the company, because the buyer may arrive with that version of the story already in mind.

The commercial conversation then starts further downstream. That can be an advantage if the machine has prepared the buyer well. It can be expensive if the machine has prepared the buyer badly.

Measure influenced demand, not just referred demand

The measurement fix is not perfect attribution. AI-mediated buying makes perfect attribution even less realistic than it already was.

A more useful approach is to create several imperfect but complementary signals. Ask new customers which research tools they used. Add an 'AI-assisted research' field to CRM discovery. Review call transcripts for references to ChatGPT, Gemini, Claude or AI comparisons. Track branded search and direct visits around periods when AI visibility changes. Compare the questions buyers ask with the questions AI systems are answering. Run periodic mystery-shopping prompts to see how the market is being represented.

None of those measures proves causality on its own. Together, they can stop management from equating 'no AI referral' with 'no AI influence'.

The goal is not to build a new attribution religion. It is to notice when the buying process has moved before the dashboard has.

The AI-Mediated Buyer Journey Audit

Use the following questions in a marketing, sales or commercial review. They are a management screen, not a validated score.

On a narrow screen, each check is shown as a separate card.

What to do now

Pick ten commercially important questions that a buyer might ask before contacting you. Run them through the AI systems your customers are most likely to use. Record not merely whether your company appears, but how it is described, what evidence is used and what is missing.

Repair factual inconsistency before chasing optimisation. If your own site, distributor pages, review sites and industry directories disagree on basic capabilities, the problem is not an AI problem. AI simply makes the inconsistency easier to aggregate.

Instrument the sales process for AI-assisted research. A simple question asked consistently can reveal more about influenced demand than another dashboard that counts only referral clicks.

Finally, equip sellers for validation. If the machine increasingly performs the first pass of research, the human's comparative advantage is judgement at the point where generic information stops being enough.

The first commercial contest may happen before contact

The HBR signal is useful because it points to a change in the customer relationship: AI can increasingly sit between a company and the customer during research and comparison. The stronger independent evidence says not to exaggerate the traffic impact yet. Both can be true. ¹

AI does not need to become the dominant referral channel to become commercially important. It only needs to influence which names make the shortlist, what buyers believe about them and which questions survive to the human conversation.

That shifts the practical objective. An AI visibility strategy should be a buyer-interpretation strategy, not merely a traffic strategy.

If AI helps decide who gets considered, the first commercial contest is increasingly taking place before the company sees the click.