Two leaks, stacked.
In our own study of 39 UK buying guides, 54% of recommendations linked only to Amazon or a retailer, never to the brand. That is the first leak. The Guide Leak Index has the figures.
The second is newer. AI assistants answer "what is the best" questions largely by reading those same guides. So a brand can win the guide, win the AI answer, and still watch the customer land somewhere else.
To be straight about it: nobody has yet measured the two together for UK brands. The second leak is a reasonable reading of separate studies, set out below, and not a proven number.
What is known.
- The buyer still leaves the chat.
OpenAI withdrew its own in-chat checkout in March 2026, and Shopify says buying through its AI channels is open to US buyers only. For a UK shopper, an AI answer ends in a link to somebody's site.
Source: Retail Week, 25 March 2026; Shopify, 18 June 2026 - Most "best product" answers lean on a guide that earns commission.
Two studies put it at 70% and 86% of answers. Both are from firms that sell AI tracking, and neither is UK only, so treat the figures as a direction, not a measurement.
Source: Suff Digital, 24 September 2026; GetCited, reported by PPC Land - The same question gives a different list each time.
Across 2,961 runs, there was less than a 1 in 100 chance of getting the same list of brands twice. Anyone who tells you your "AI ranking" is selling a number that does not exist.
Source: SparkToro, 28 January 2026 - Product cards mostly come from the open web, not from a feed you submit.
One study of 812,190 ChatGPT product cards found 87% were built from pages the assistant had read, and 13% from merchant feeds. The study is from a firm that sells tracking tools.
Source: Profound, 23 July 2026 - Many shoppers do not click the AI's link at all.
In a survey of more than 2,000 UK adults commissioned by Adobe, 43% said they open a new tab and search for the product themselves.
Source: Marketing Week (paid content from Adobe) - Visitors who do arrive from AI now buy more often.
Adobe reports that in August 2026 visitors sent by AI to UK retail sites converted 20% better than other visitors. In March they converted 11% worse.
Source: Adobe, reported 23 September 2026
Check your own brand in an hour.
- 1. Ask three assistants
ChatGPT, Google's AI Mode and Perplexity. Use a private window, signed out, and start a new chat every time.
- 2. Use four questions
Best [your product type] in the UK. Best [product type] for [your customer's main need]. [Your brand] or [your main rival]? Where to buy [your best seller] in the UK.
- 3. Ask each one five times
Answers change from run to run, so one answer tells you nothing. Sixty runs takes about 45 minutes.
- 4. Write down four things each time
Were you named? Who else was? Where did the link go: your store, Amazon, a retailer, a guide, nowhere? Which sources did it cite?
- 5. Open the guides it cited
Find the three it leans on most and click the buy links next to your product. This is where the sale is really decided.
Read the result honestly. Say "named in 3 of 5", never "ranked second". Five runs is a small number: a brand that appears in a third of all answers can still show up zero times in five by chance. What you get is what one person saw on one day, which is enough to see where the links go and not enough to prove a trend.
What is worth doing about it.
- Fix the guides first. They feed the assistants and they are run by people you can talk to. Ask each publisher for a direct link, with a proper affiliate arrangement behind it.
- Make your own product pages easy to read. Price, size, ingredients, delivery and returns in plain text on the page, since most product cards are built from what the assistant reads there.
- Be sceptical of anything sold as a ranking. The lists change every time. Pay for work that changes where links go, not for a dashboard.
I did the first of these on an online store my partners and I later sold. Going to the guide publishers one by one built an affiliate programme that became a third of sales.
Every source was read on 12 October 2026. Research was AI-assisted. How this site is made. Found an error? Email hi@aiintelligence.com and it will be corrected.