AI in e-commerce: why it changes buying behavior but stays invisible in tracking

In March 2026, the logistics provider Ryder asked 1,160 online shoppers in the US how they shop. 64 percent of them said they use AI tools for it.^[1]

Look at our clients’ analytics accounts and you see something completely different. The share of traffic that demonstrably comes from AI sources sits below one percent in some of them. That gap is exactly what we talked about in episode 16 of our podcast Buzzsozial.

At first glance this looks like one big measurement error. Either people in the studies are ticking boxes at random, or web analytics is missing whole streams of visitors. But the studies are right, and so are the analytics accounts. They simply show two completely different phases of a purchase, and those no longer happen on the same platform.

So what is actually behind those 64 percent?

Look at the Ryder study more closely and it quickly becomes clear where that high number comes from. Ryder counts everything that has AI in it somewhere.

Germany looks much the same. Here too the focus sits entirely on research. A vendor study by the agency kernpunkt and commercetools showed in March 2026 that a good 57 percent of German respondents have already used chatbots or AI assistants in online shops. 65 percent can readily picture using them for comparisons and product search.^[2] For consumer electronics specifically, a study by the consultancy Oliver Wyman and the industry body gfu found in August 2026 that AI was involved in 44 percent of the purchases examined in Germany.^[3]

So the technology has clearly arrived at scale. But what all these numbers measure is above all the research phase. The search for the right product has moved into the chat, while the purchase happens somewhere else.

Research in the chat, purchase on Google

Inside the chats, almost everything now happens without a single click. A link gets clicked at the very end at best, and usually only once the buying decision has effectively been made, or when the chat cannot supply one specific detail.

And often enough even that does not happen. We notice it in our own team, in the way we shop ourselves. A colleague sends ChatGPT a voice message on the way to work, asking for the best over-ear headphones. The AI comes back with two models and weighs up the pros and cons. A few hours later the same person sits down at their computer, types the name of the product they picked into Google and orders from the site with the cheapest offer. For the shop, that visit turns up in analytics as a perfectly normal organic click. That AI actually set up the deal is something nobody sees in the attribution.

Sometimes the digital chain breaks off entirely after the chat. According to the Oliver Wyman study, 32 percent of electronics purchases still end in a physical store.^[3] If the advice beforehand came from an AI, that route shows up in no digital attribution at all.

So even though this is quite hard to measure, we know the channels are shifting massively. Salesforce analyzed this across more than 1.5 billion shoppers. Between August 2025 and May 2026, classic search lost 15 percent as a discovery channel. AI search as the first step of the shopping journey, meanwhile, shot up by a full 200 percent year over year.^[4]

We see exactly that with our own clients, of course. Right now, as of September 2026, clicks from informational queries are collapsing by 40 to 50 percent in some accounts. In return, 30 to 40 percent of all impressions now come from AI features. And this hits strong brand terms too. As Sebastjan put it on the podcast, a stark discrepancy between the tools is opening up right now. Sistrix still shows you in position one for your own brand. But in Google Search Console those exact rankings start to jump around or slip to position two, because the AI answer pushes itself above the first organic result. The bitter consequence is that the shop loses clicks even when somebody searches for the brand directly.

Less traffic, but noticeably better conversion rates

The traffic that does still trickle through from the chats to the shops is tiny, but it has a completely different quality. The software company Adobe measured that AI visits to US retailers shot up by almost 400 percent in the first quarter of 2026 against the year before. The interesting part is that in March 2026 those visitors converted 42 percent better than the rest of the traffic. A year earlier they were still doing 38 percent worse.^[5]

That makes complete sense, because anybody landing in a shop today through a link from the chat has the tedious comparing behind them and knows exactly what they want. We observe the same in our client projects. Conversion rates are rising noticeably, even if that does not offset the massive loss of clicks on classic guide content one for one.

Will we soon leave the whole purchase to AI?

There is a simple reason why buyers still come into the shop themselves to click “buy” at the end. We are reluctant to hand over that very last step. Anyone who thinks that all the hype around “agentic commerce” means AI will soon take over the entire purchase is skipping one decisive point. Most people are simply not ready for it yet. According to the Oliver Wyman study, just five percent of Germans would let an AI agent complete a purchase entirely on its own.^[3] A good 36 percent can at least imagine it in theory, going by the kernpunkt survey.^[2] But between theory and your own bank account there is still a world of difference.

The hesitation shows in the US too, where new tech topics usually move a notch faster. The software vendor Akeneo asked around 1,000 consumers there in May 2026. 22 percent said they had bought something on the basis of an AI recommendation. Yet only a paltry 3 percent name it as their strongest purchase trigger. A clear majority of 62 percent would rather compare the prices themselves in the end, instead of trusting the technology blindly.^[6]

Where is the sticking point? Trust, above all. In Germany, more than 56 percent of respondents worry about paid or influenced recommendations from chatbots.^[2] And anybody who does not trust the recommendation certainly will not hand the AI their credit card. Whether fully autonomous purchases really become everyday behavior over the next few years therefore stays a big question mark. Especially since EU rules on transparency and market dominance could slow the process down further.

Retail is nevertheless preparing in the background. In a representative Bitkom survey from 2025, half of the retailers in Germany expect customers to leave product and price comparison entirely to an AI by 2030.^[7] Many shop operators are therefore already planning for the autonomous agent as a completely new type of customer.

New metrics: what really counts in web analytics now?

When traffic stops being the hard currency in e-commerce, shops need different metrics to judge their success. In our analysis we therefore look at three big levers above all:

1. Visibility and citations: Does the brand turn up in the answers of the language models at all? Whoever is missing here barely exists in the early research phase. To make that measurable, we built our own tool, Buzz­View. With it we track exactly whether and how often a shop gets recommended and linked in ChatGPT, Perplexity or Google AI, or whether the answer only gives general information without naming a brand.
2. Positioning in the content: Visibility alone is not enough, though. It makes a big difference which attributes a large language model (LLM) ties to a brand, whether that is the “cheap alternative”, the “expensive premium product” or the solution to one very specific problem. Picture somebody looking for a stroller that is extremely maneuverable and fits neatly into a car trunk. The race is decided in the chat. If the models do not name your product for exactly those attributes, you have lost the deal. We evaluate these associations and sentiments directly through our Buzz­View audits.
3. Channel attribution by elimination: Since the actual AI traffic often is not directly measurable in normal analytics, we work by elimination and look for correlations. We lay the visibility data from Buzz­View over the classic analytics numbers. Then we check very concretely whether clicks and purchases from organic brand search rise to the same degree as impressions in AI answers.

Where shops should start now

The most urgent job for shops right now is to enrich their own shopping feed with data and connect it directly to programs like the ChatGPT merchant program. At the same time, the product copy in the online shop has to be written so that an LLM can extract and quote the key arguments easily. Because language models rarely trust a single manufacturer page blindly, we also build up independent third-party sources for our clients on purpose. That means mentions on large portals, comparisons, discussions on Reddit and videos. They help the AI place a product correctly.

Special formats come after that. We spend no time writing dedicated llms.txt files for crawlers or building separate, AI-optimized page versions for every URL while the product data itself is not clean.

And in the end this comes full circle to the surveys from the start. The studies are right when they measure heavy AI use in online shopping. And the analytics accounts are right when they show barely any direct traffic from those sources. The AI advises at length, but the human buys in the end through channels that tracking does not connect directly to the chat.

Anyone waiting today for measurable AI traffic to get big is waiting for the wrong number.

Sources