All stories · Article
AI can read your CRO data now. Why customers buy is still your job
AI now reads heatmaps, surveys and funnel drop-offs faster than most people, so the part of CRO a Shopify founder still has to own is knowing why customers buy.
Share: WhatsApp · Telegram · X · Facebook · LinkedIn · Email
@jordanhill11 puts it simply: CRO has two sides, and you only need to be good at one.
Reading heatmaps and surveys is now machine work
The first side of CRO is data, and Jordan Hill, who does CRO for Shopify brands, says AI is swallowing it fast. He gives three examples. Heatmap tools now connect to AI models through MCP, and in his view the model scans a site faster than anybody and gives better recommendations than most CROs. Customer surveys used to take days to read and summarise. Now AI spots the trends and the sentiment in seconds.
The third example is the path shoppers take through the store. AI can read your analytics, find the biggest drop-offs in the funnel and suggest fixes on the spot.
That is good news for founders. Hill points out that this work used to depend on one person reading the numbers and making the right call, which he calls quite risky. The hours it takes have also dropped sharply.
Some operators already run the whole loop this way. @hanifproduktif lays out a workflow that connects Claude Code or Codex to Meta Ads MCP, then to analytics and the landing page. The AI writes ad copy and creative variations, A/B tests hooks, angles and offers, finds where visitors drop off and tests fixes. Losing ads get cut, winners get budget, and the cycle repeats.
Shopify is pushing the same direction in research. Shopify Engineering describes SimGym, which runs storefront A/B tests with VLM agents in a live browser, using personas built from production clickstreams. Experiment cycles go from weeks to under an hour. This is research the team is presenting at COLM2026, not a feature merchants can switch on. It still shows where the data side of CRO is heading: cheaper, faster, and less dependent on a specialist.
Data can only polish what is already on the page
Faster data still has a ceiling. "But data only ever tells you how to improve what's already on the site," Hill writes. What comes out of it is UX tweaks, small copy changes and elements moved around. Useful work. Limited upside.
The same trap shows up in the tactics people reach for when they treat CRO as a list of add-ons. "CRO isn't adding some timer/review and just praying it works," writes @kasper_dolk, who builds landing pages and runs CRO for brands. He ranks slapping a review onto a page as a low-impact move.
A countdown timer or another review block is a guess dressed up as a fix. It changes the page without asking why the shopper hesitated.
Why someone wants the product is still a human call
The second side of CRO, in Hill's framing, is marketing applied to a website. It means looking at a product and imagining what the site should be saying. It means finding the angles nobody has run yet, the static images, the landing page built around that angle, and the positioning that turns a boring business into something worth buying. In his view, AI can't do that part.
Once AI has done the data, he argues, what's left is working out why someone would want the product in the first place. There isn't always data for that. It runs on experience and on understanding what makes people act.
Jordan Hill ⚙️ CRO for Shopify Brands@jordanhill11There's two sides to CRO, and you only need to be good at one of them.
One side is the data. Historically brands outsourced CRO to agencies because they didn't know how to look at data, pull ideas out of it and come up with tests, so people's idea of a CRO has always been someone super data orientated. Data still has a part to play and you need to understand it, but that side is being eaten alive by AI:
-…Oct 7, 2026 · View on X
Kasper Dolk's camera lens client shows what that looks like on a real page. The store sells lenses with many options, so shoppers landing on the site struggle to be fully sure a lens fits their camera. A review would not answer that. His team built a simple popup where shoppers enter their camera brand and model and get the right mount. It took longer to set up. He counts it as the high-impact option, because it targets the exact point where buyers get stuck.
The same pattern holds in creative research. In a September post about finding ad angles, @thankyouecom says he tried using AI for customer research and it produced winners only 1 in 100 times. That is his own account, not a measured rate, and it is about ads rather than pages. His explanation is the useful part: people who let AI do most of the research and then get stuck have handed off the job of understanding the angle. After reading 130 people describe the same problem, he says, you know their fears and what drives the purchase. His verdict: "manual research wins."
Even the most AI-forward workflow leaves this to a person. Hanif's loop hands almost everything to the model, then ends on one line: "\"Taste\" is the only moat now."
Split the work: AI reads the data, you read the customers
The practical move is a clean division of labour inside each test cycle.
- Give AI the data reading. Heatmaps, survey summaries and funnel drop-offs are now fast and cheap, so let the model produce the list of where visitors leave.
- Read customers yourself, in the places they talk. The sources @thankyouecom uses for ads work for pages too: Reddit threads where people describe the problem in their own words, competitor products on Amazon (what those brands get wrong and what buyers hate), post-purchase surveys from people who already paid you, and organic TikTok.
- Pick one high-impact change per round. Dolk's point is that you can only test so many things at once, so each slot should go to the change that removes a real blocker, not to another badge.
- Ask an agency which side it sells. Hill notes that brands historically outsourced CRO because they didn't know how to read data, pull ideas from it and set up tests. That is the part AI now does cheaply. Before you sign, ask whether you are paying for data reports or for someone who understands your customers and your positioning.
This week, read your post-purchase survey answers yourself, start to finish. Write down the questions buyers kept asking before they paid. Then ask the AI to suggest tests, with that list in front of it. The test plan starts from what customers said, and the heatmap tells you where to put it.