As of 5 November 2026, the public survey data has moved the first step of product research off the search box. An impact.com study of 1,128 AI-active US shoppers found that 47 percent began their most recent product search with an AI-generated answer, split between standalone AI platforms at 33 percent and AI Overviews at 14 percent, while 26 percent still started with traditional search. The number that matters for ecommerce teams is not the 47 but the 72: the share of respondents who said they click through to a brand or retailer site before deciding. The answer layer now decides which product pages get read at all, and that changes how listings, reviews and price are measured.
What the 47 percent figure actually measures
The 47 percent is a starting-point share, not a traffic share, and the difference decides how the number should be used. A shopper who opens an assistant first still ends up on a product page in most cases, but the page arrives after the shortlist has already been formed by a machine. That means the product listing is no longer competing only against other listings in a grid; it is competing against whatever the answer layer already decided to summarise.
The composition matters just as much. A third of AI-active shoppers chose to open an assistant, while 14 percent met an AI Overview without asking for one, and those two placements reward different content. An assistant pulls on whatever it can retrieve and reconcile across sources, while an Overview summarises what the search index already ranks. Treating them as one channel hides the fact that one is pull and the other is push.
Why the first click moved upstream
For a category manager, the shift changes what a demand signal means. When product research starts inside an answer, the questions a shopper asks never appear as keyword volume, and the queries that do appear are shorter and later in the journey. Category growth can look flat while the decision has already been made somewhere that the search console cannot see, which is why teams that only read keyword data keep arriving late to a shift that started weeks earlier.
Where the funnel breaks on cost
The economics of the new entry point are less flattering than the behaviour data. An October 2026 working paper covering 1,898 websites over 60 weeks found that paid visits from ChatGPT generated 0.89 dollars of revenue per session in US ecommerce, against 2.45 dollars for paid Google Search, with conversion at 0.37 percent versus 1.17 percentReal Internet Sales: revenue per visit by traffic source. The gap is roughly three to one on conversion, not a rounding difference.
Read carefully, the paper does not say AI traffic is worthless. It says the traffic behaves like an early-stage referral: fewer sessions convert, and the ones that do arrive with a narrower set of questions already answered. Budgeting it against search CPCs therefore flatters it, while budgeting it against upper-funnel discovery spends is closer to the truth. The honest comparison is revenue per session after returns, not cost per click.
Citation sets are not stable enough to audit once
The reason a single audit misleads is mechanical. Exposure Ninja tracked 59 UK prompts every day across ChatGPT, Perplexity and AI Overviews, collecting 7,965 responses, and found that the same set of cited sources reappeared on consecutive days in only a small fraction of casesWebbiquity: the state of AI search in eight findings. A citation snapshot is a photograph of a moving object, and the platform with the most movement is the one most teams check first.
| Platform tracked | Days with an identical citation set | What it means for measurement |
|---|---|---|
| ChatGPT | 2 percent | A weekly check is close to meaningless; measure the rolling share of prompts where a product page is cited |
| AI Overviews | 6 percent | Source rotation is high, so a single win on one prompt should not be read as a durable position |
| Perplexity | 27 percent | The most stable of the three, which makes it the cheapest place to test whether a content change moves citations at all |
The same study puts AI referrals at 0.14 to 1.08 percent of site visits, with the higher end concentrated in a few categories, while Adobe's retail data cited alongside it found 42 percent better conversion for AI-referred visits. Those two numbers are not in conflict: the volume is still small and the quality is higher, which is exactly the profile of a channel worth instrumenting before it is worth scaling.
The review layer decides who gets quoted
When an answer is assembled, the cheapest evidence to summarise is what shoppers already wrote. That makes customer review text a citation asset rather than a customer-service by-product, and it changes what a review programme is for. A review analysis that groups complaints by product pain point produces the sentences an answer layer can lift; sentiment analysis at the listing level produces a number nobody quotes. Retailers that run only the second get mentioned less than rivals with thinner catalogues.
What to measure every week
The practical response is a short weekly loop rather than a quarterly report. Pick a fixed set of twenty to fifty prompts that match how shoppers actually describe the category, record which sources are cited and whether a product page is among them, then reconcile that against sessions and revenue from AI referrers. The loop only works if the prompt set stays fixed while the sources rotate, because the variable being measured is the source, not the question.
FAQ
Does an AI answer replace the product page?
Not for most shoppers. 72 percent of the respondents in the impact.com study still clicked through to a brand or retailer site before deciding, so the product page remains the place where the purchase is confirmed. What changed is that the page is reached after a shortlist has formed, which makes the first screen of the listing and the review block more decisive than the browse grid.
Why are the cited sources so unstable?
Because the answer layer re-retrieves and re-ranks every time a prompt is run, and small changes in phrasing or index freshness change what it can reconcile. In the Exposure Ninja tracking, identical citation sets appeared on only 2 percent of consecutive days for ChatGPT and 6 percent for AI Overviews. That instability is the reason a one-off audit tells you almost nothing about position.
Which metric should an ecommerce team add first?
Revenue per session from AI referrers, measured after returns and segmented by platform. It is the only figure that can be compared directly with paid search, where the working paper found 2.45 dollars per session against 0.89 dollars for ChatGPT visits. Session counts and citation counts both move too easily to anchor a budget decision.
Is a channel at 0.14 to 1.08 percent of visits worth the work?
It is worth instrumenting, which is a different decision from scaling it. The volume is small, but Adobe's retail data points to 42 percent better conversion on AI-referred visits, so the sessions that do arrive carry more value per visit. The sensible sequence is to build the measurement first and let the revenue case decide the investment.
