As of November 5, 2026, the most detailed public read on how shopping now starts comes from impact.com's May survey of 1,128 AI-active United States consumers: 47 percent began their most recent product search with an AI-generated answer, against 26 percent who started with a traditional search engine. The same study found that 72 percent still click through to a brand or retailer site before deciding, which is why the claim that AI has removed the website from the journey is only half right.
What the 1,128 shoppers actually did
The breakdown inside the 47 percent is the useful part. Thirty-three percent started on a standalone AI platform and 14 percent started with an AI Overview inside a search result, so the two entry points behave differently for a retailer: one is a destination a shopper opens on purpose, the other is a summary the shopper never asked for. Adweek's write-up of the report also records the reasons shoppers gave, led by a preference for a direct answer over a list of links at 52 percent, followed by handling of complex scenario-based needs at 46 percent.
The most awkward finding for measurement teams is the 5 percent who bypass the AI's embedded link and start a new search manually on a brand or retailer site. That behaviour leaves the influence intact and the attribution gone. A retailer that only credits the sessions it can label as AI referrals will under-count the effect precisely among the shoppers who were most persuaded by the answer they had just read.
The shortlist forms before the brand sees it
| Measure | Share reported |
|---|---|
| Started last product search with an AI-generated answer | 47 percent |
| Started on a standalone AI platform | 33 percent |
| Started with an AI Overview in search results | 14 percent |
| Started with traditional search results | 26 percent |
| Click through from AI tool to brand or retailer site | 72 percent |
| Used a brand website to research in the past 30 days | 27 percent |
Read as a set, these numbers describe a handover rather than a replacement. The shortlist is formed in the answer, the confirmation happens on the site, and the two stages are measured by different systems. For a category manager the consequence is concrete: the product page no longer has to win the comparison, because the comparison has already happened, but it does have to survive the last check a shopper makes before paying.
That check is usually about a specific attribute, not about brand story. The shopper who arrived from an answer already knows roughly what the product is and what it costs; what she is confirming is fit, compatibility, delivery date or return terms. Pages that answer those four questions above the fold convert the handover, and pages that repeat the marketing copy of the answer waste it.
Catalog data decides which product gets named
Machine readability is a shelf-stocking problem
The constraint on the answer side is not creativity but structured data. According to AI Best Practices, Adobe found that United States retail product pages average only 66 percent machine readability, the lowest of any page type, and InMobi's scan of more than 8,000 retail sites produced a top agent-readiness score of 61 out of 100. An assistant that cannot read an attribute will not guess it; it will pick the product whose attributes it can read.
Why inconsistent pricing removes a product entirely
The same body of research reports that 41 percent of multichannel retailers have inconsistent pricing across channels and 40 percent have inventory that is not synced in real time. For an assistant assembling a recommendation, a price that does not match the checkout is a reason to drop the item rather than to add a caveat. The commercial upside of getting this right is measurable: the report cites Shopify data showing AI-referred shoppers who reached a product page converted about 80 percent better than organic-search referrals, with the advantage doubling in specification-heavy categories.
Timing sharpens the point. The same source cites a Salesforce projection that 20 percent of e-commerce traffic will originate from AI chat agents during the holiday season, which turns agent readiness from a project into a deadline. Retailers that have launched AI-enabled experiences report conversion gains of 5 to 15 percent, a range that is consistent with the idea that the fix is mostly about removing errors rather than adding capability.
Where customer reviews enter the answer
Sentiment is not the metric that matters
When a catalog is incomplete, the answer draws on what customers have written, and customer review analysis becomes the de facto product description. What survives that transfer is not the average rating but the repeated complaint: a sizing issue mentioned forty times, a battery that dies in cold weather, a part that does not fit the model it is sold with. Those phrases are specific, verifiable and easy for a language model to restate, which is exactly why they travel.
For a brand, that creates a feedback loop worth managing deliberately. Consumer feedback collected on the product page is now an input to a description the brand does not write, and a product pain point that stays unfixed becomes a permanent caveat in every comparison answer. The practical response is to treat recurring review themes as catalog corrections: if buyers keep saying the item runs small, the size field is wrong, and fixing the field changes the answer more reliably than answering the review.
There is a limit to how far this can be pushed. Reviews can corroborate a claim, but they cannot supply an attribute nobody measured, and they carry their own bias towards buyers who had a strong experience. A retailer using review analysis to fill catalog gaps should therefore treat it as a signal to check, not as a source of record, and should keep the proportion of catalog fields sourced from reviews visible to whoever owns data quality.
What to fix before the cut-off
The order of work follows from the evidence. Structured price and stock come first, because their absence removes a product from consideration altogether. Attribute completeness comes second, because it decides whether the product is described accurately or replaced by a review. Measurement comes third, and it needs to include the sessions that arrive without a label, or the picture will understate the channel that is doing the persuading.
None of this requires a new platform. It requires one owner for the feed, one tolerance for price differences, and one weekly check of whether the products that matter are retrievable when the category question is asked. A team that runs those three checks through December will know more about its real funnel than a team that only reads referral reports, and it will be able to show the difference between being found and being chosen.
FAQ
If AI answers start the search, why does the brand website still matter?
Because the click-through has not disappeared. In the impact.com survey, 72 percent of AI-active shoppers said they sometimes, usually or always click through from an AI tool to a brand or retailer site before making a final decision, and only 27 percent had used a brand website to research a product in the previous 30 days. The site is no longer the place where the shortlist is built, but it is still where a large share of decisions are confirmed, which puts the burden on product pages that can answer a question in one screen.
What does machine readability have to do with reviews?
An assistant can only repeat what it can parse and what it can corroborate. Adobe's finding that US retail product pages average 66 percent machine readability means a third of the attributes an assistant needs are missing or buried, so it falls back on other sources, and reviews are the most available of those. A catalog that is silent about fit, compatibility or contents does not stay silent in the answer; it is replaced by whatever customers wrote, which is not the version a brand would choose.
Which review metric actually predicts whether a product gets recommended?
Recurring, specific complaints are more useful than star averages. A product with a 4.6 average and repeated mentions of a sizing problem will be described with that caveat in a comparison answer, while a product with a 4.3 average and no recurring complaint is easier to recommend. Review analysis for this channel therefore looks for repeated phrases tied to a product attribute, and then checks whether the catalog contradicts them.
What should a retailer fix first before the holiday cut-off?
Price and stock consistency for the products that carry the most revenue. The same survey work found 41 percent of multichannel retailers reporting inconsistent pricing across channels and 40 percent reporting inventory that is not synced in real time, which are the two inputs an assistant uses to decide whether an item can be recommended at all. Fixing those two for 200 products is a smaller job than rewriting a catalog, and it addresses the failure mode that removes a product from the answer entirely.
