Collect three kinds of evidence first: where people drop out (your analytics funnel), what they do on the page (free session recordings and heatmaps, from a tool such as Microsoft Clarity), and what they tell you (a one-question exit survey, chat logs, enquiry emails). Then ask AI to rank the likely reasons from that evidence, and fix and measure one at a time.
The common mistake is skipping the evidence and asking an assistant to "review my website and tell me why it doesn't convert". You'll get a confident, generic list, such as add trust badges, create urgency and simplify navigation, that may have nothing to do with your visitors. AI is very good at sorting and summarising what your visitors did and said. It is poor at guessing it. And some visitors were never going to buy, so the goal is to find the ones who wanted to and couldn't.
Find the step where people leave
Before asking why, find where. Every online sale passes through a few steps, and a funnel report shows how many people make it from each to the next. In Google Analytics 4 it's a funnel exploration (Explore, then Funnel exploration, with steps such as viewed an item, added to cart, began checkout and purchased). Shopify's analytics shows similar steps in its conversion reports.
Throughout, the example is a hypothetical independent bookshop that sells online as well as in the shop, including special orders for books it doesn't stock. A month of its funnel:
| Step | Visitors | Carried on to the next step |
|---|---|---|
| Visited the site | 6,200 | 40% viewed a book |
| Viewed a book page | 2,480 | 12.5% added to basket |
| Added to basket | 310 | 55% reached checkout |
| Reached checkout | 170 | 44% completed |
| Completed an order | 74 |
Two steps stand out. Only one in eight people who looked at a book added it to the basket, and more than half of those who started checkout didn't finish. Checkout drop-offs are usually about surprises: costs, delivery times, forced account creation. Book-page drop-offs are about the page: price, availability, whether the visitor was a browser or a buyer. Those are different problems with different evidence, so the owner decided to look at both, starting with checkout because the people there had already shown intent.
If your tracking isn't reliable yet, fix that first, because every later step depends on these counts. Setting up conversion tracking properly covers the basics.
Watch what visitors actually do
Session recordings replay a visitor's scrolling, clicking and typing (with typed details masked), and heatmaps show where clicks and scrolling concentrate across many visits. Microsoft Clarity does both, and according to Microsoft's Clarity FAQ it's free with no traffic limits and no sampling. Recordings are kept for 30 days, so review them regularly rather than saving it for later.
Setting it up takes about 15 minutes: create a project, add the tracking code (many site builders and shop platforms have a Clarity integration), and wait a few days for recordings to build up. Then look at the signals Clarity flags on its dashboard:
- Rage clicks: repeated rapid clicks on one spot, usually something that looks clickable and isn't, or a button that doesn't respond.
- Dead clicks: clicks that do nothing.
- Quick backs: visitors who open a page and go straight back, often because it wasn't what they expected.
- Excessive scrolling: hunting up and down for information that's hard to find.
Filter the recordings to the step you care about, such as sessions that reached checkout but didn't purchase, and watch 20 of them. It takes about an hour at double speed. Copilot in Clarity can summarise a page's heatmaps (first clicks, last clicks, rage clicks, dead clicks and scroll depth) and answer questions about your dashboard in plain English. Microsoft's own guidance is to double-check what it says against the data, which is good advice for any AI summary.
On privacy: Clarity masks what visitors type into input boxes in every masking mode, and Microsoft says full recording needs valid consent where the law requires it. Mention it in your privacy notice and connect it to your cookie banner. If you're unsure what your site needs, ask your data-protection adviser.
Turn an hour of recordings into patterns
Watching recordings is easy; remembering what you saw across 20 of them isn't. Jot one line per recording as you go, then let AI find the patterns:
Below are my notes from watching 20 session recordings of visitors who
reached checkout on my online bookshop and didn't buy. One line each.
Group them into patterns. For each pattern: a label, how many sessions
show it (list the session numbers), and what the visitor seemed to be
trying to do. Keep patterns separate if you're unsure they're the same.
Don't suggest fixes yet.Part of what came back (illustrative):
Delivery cost shock (8 sessions: 2, 3, 5, 9, 11, 14, 17, 20): scrolled to the delivery options at the final step, paused, then left or went back to the basket. Account wall (4 sessions: 1, 6, 12, 19): clicked "guest checkout", which isn't obvious below the sign-in form; two gave up on the password screen. Special-order uncertainty (3 sessions: 4, 8, 15): went back to the book page from checkout and scrolled to the availability line.
Check it by rewatching one session from each group. The numbers are only as good as your notes, but patterns of eight out of 20 are hard to argue with.
Reading the heatmap on the page that leaks most
The bookshop's other leak was the book page, where seven in eight viewers never added anything. Recordings of browsers are less revealing than recordings of people at checkout, because many were only ever browsing. A heatmap across a few hundred visits is more useful here, and splitting it by device is the single most useful filter.
The owner asked Copilot in Clarity to summarise the heatmaps for the book page template on mobile. A trimmed version of the kind of summary it gives (illustrative):
Most first clicks on mobile land on the cover image, which opens a larger image. Scroll depth drops sharply after the cover: roughly half of visitors don't reach the price and "Add to basket" button. Dead clicks cluster on the grey "Special order" label beneath the title.
Each claim was checkable. The scroll map confirmed that on a typical phone the cover filled the whole first screen, pushing the price and button below it. The dead clicks on "Special order" made sense too: visitors expected the label to explain what special order meant, and it did nothing. Two cheap changes followed from that, a smaller cover on mobile with the price beside it, and a tappable "Special order: usually 3-7 working days" line. Neither came from a generic checklist; both came from watching where thumbs actually went.
A tip for small sites: heatmaps need volume, so pick the template that most visitors see (the product page, the service page) rather than a single low-traffic page, and give it a fortnight of data before reading anything into it.
Ask the people who left, and the people who nearly did
Recordings show what happened; they don't show what people were thinking. Two short questions fill that gap:
- An exit question on the basket or checkout page, shown once to people who are about to leave: "Was anything stopping you from ordering today?" with a free-text box. Survey widgets from tools like Hotjar, or your site builder's own, can do this.
- A post-purchase question on the order confirmation page: "What nearly stopped you ordering?" Buyers are generous with answers, and they describe the same obstacles as the people who gave up.
Keep it to one question. Every extra field cuts responses. After a month the bookshop had 46 answers, including "is the $4.95 delivery per book or per order?", "didn't know if the special order would arrive before the 12th", "wanted to collect in the shop" and "cheaper elsewhere". Paste the answers into an assistant and ask for themes with counts and quotes; analysing customer feedback surveys with AI has a fuller method for larger sets.
Mine the questions already in your inbox
Every "do you have…?" email, chat message and phone question is a visitor who got stuck and cared enough to ask. Most of them are sitting in your inbox already. Export or copy three months of pre-sale enquiries and ask:
These are pre-sale questions customers sent my online bookshop over
three months. Group them by the information the customer couldn't find
on the website. For each group: count, two example questions, and the
page where the answer should have been.The bookshop's biggest group, 31 questions, was some version of "if I order a book you don't have in stock, when will it arrive?" The website said "usually 3-7 working days" in the delivery FAQ, a page almost nobody visited. The answer existed; it was just in the wrong place.
Let AI rank the reasons, with the evidence attached
Now combine the four sources: the funnel, the recording patterns, the survey themes and the inbox groups. This is where AI earns its keep, because it can hold all four at once and cross-reference them.
Here is evidence about why visitors to my online shop don't buy:
1) funnel numbers, 2) patterns from 20 recordings, 3) exit and
post-purchase survey themes with counts, 4) pre-sale question groups.
Rank the likely reasons people leave. For each reason: which evidence
supports it (cite the source numbers), how strong the evidence is
(strong / moderate / weak), which funnel step it affects, and what
would confirm or rule it out. Only list reasons the evidence supports.The ranked result, after the owner's corrections:
| Reason | Evidence | Strength | Step |
|---|---|---|---|
| Delivery cost unclear or a surprise | 8 of 20 recordings; 14 survey answers | Strong | Checkout |
| Special-order timing unclear | 3 recordings; 9 survey answers; 31 emails | Strong | Book page |
| Guest checkout hard to find | 4 recordings; 3 survey answers | Moderate | Checkout |
| No click-and-collect option | 5 survey answers; 6 emails | Moderate | Checkout |
| Price higher than big online retailers | 6 survey answers | Weak to moderate | Book page |
The correction: the first version ranked price second, as "strong", on the strength of six survey answers, two of them long and angry. Six out of 46 is real but not dominant, and nothing in the recordings pointed to it. Assistants give vivid comments more weight than quiet patterns, so check each "strong" rating against the counts. Price also isn't something the bookshop can fix by matching large retailers, so it went to the bottom of the list, to answer with reasons to buy locally rather than with discounts. If trust is part of the picture, online shop mistakes that hurt trust and conversions covers the usual culprits.
Test one fix at a time, even on a small site
Small shops rarely have the traffic for formal A/B tests, which split visitors between two versions at once. The practical alternative is a before-and-after comparison with a written change log, so you know exactly what changed and when:
| Date | Change | Step watched | Before (4 weeks) | After (4 weeks) |
|---|---|---|---|---|
| 2 Oct | "Delivery $4.95 per order, free over $40" on every book page and in the basket | Checkout completed | 44% | 58% |
| 30 Oct | Special-order books show "Usually with you in 3-7 working days" under the price | Added to basket | 12.5% | To be measured |
Rules that keep the comparison honest: one change per period; compare equal lengths of time; note anything else that happened (a newsletter, a sale, a school holiday); and remember that seasonal businesses change on their own, so compare with the same weeks last year if you can. The bookshop's first fix moved checkout completion from 44% to 58% over four weeks, about 24 more orders a month at its traffic level. That's worth far more than the hour the change took.
For people who still leave with books in the basket, abandoned cart emails written with AI are the natural safety net, but fix the reasons first. A reminder email doesn't help if the delivery cost is still a surprise.
Reasons that turn up again and again in small businesses
The specific reasons differ, but a handful appear in business after business. Some illustrations of what the evidence tends to show:
- A bakery taking cake orders online: customers choose a design, then discover at the final step that the earliest collection date is ten days away. Recordings show people clicking through the date picker and leaving. The fix is stating lead times on the product page.
- A dog groomer with online booking: the booking system demands an account and password before showing available times. Rage clicks on the "see times" button that leads to a sign-up form give it away. Showing availability first, then asking for details, usually helps.
- A bicycle repair shop: the service page lists what each service includes but not the price, so visitors go back to search results. Quick backs from the service page are the signal, and the pre-sale inbox is full of "how much is a service?"
- A café selling gift cards: the gift card page works on a laptop but the amount selector doesn't respond on some phones. Dead clicks on mobile only is the clue; checking the heatmap by device finds it in minutes.
- A florist: the same-day delivery cut-off appears only at checkout, after the customer has chosen a bouquet. Exit answers say "needed it today". A banner with the cut-off time on every page solves most of it.
Notice that none of these are about persuasive copy. Most lost sales on small sites come from missing information and small frictions, which is good news, because they're cheap to fix. Where the words on the page do matter, product descriptions that answer buyers' questions and landing pages built around one clear offer are the next places to look.
What the evidence can't tell you
- Whether the traffic was ever going to buy. If a new ad campaign sends thousands of people who wanted something else, conversion falls and nothing on the site is broken. Check where the drop-off visitors came from before changing pages.
- Offline reasons. Some people browse online and buy in the shop. For the bookshop, the survey answers asking for click-and-collect suggested a fair number did exactly that.
- Anything from a handful of sessions. Three recordings are an anecdote. Twenty start to be a pattern.
- What people will do, as opposed to what they say. Survey answers about price are often a polite way of saying "I wasn't convinced". Weigh them against behaviour.
Run the whole cycle, funnel, recordings, questions, ranking and one fix, every two or three months, or whenever the funnel numbers shift. Each round gets quicker, because the tracking, the survey question and the prompts are already in place.
Finding out why visitors don't buy: more questions
How much traffic do I need before this is worth doing?
The recordings and survey steps work with very little traffic, because you're looking for patterns in behaviour, not statistics. Twenty recordings of people abandoning a basket will usually show the problem. Measuring whether a fix worked needs more: aim for at least a few dozen orders in each before-and-after period, or run the comparison for longer.
Is session recording legal on a small business website?
It's widely used, but it counts as collecting data about visitors, so treat it like analytics. Say so in your privacy notice, keep masking switched on so typed details aren't captured, and use your cookie banner where consent is required. Microsoft says Clarity needs valid consent for full features in places that require it. If you're unsure, ask your data-protection adviser.
Can I just ask ChatGPT to review my website instead?
You can, and it will give you a tidy checklist of general advice. The trouble is that it's guessing from what sites in general get wrong, not from what your visitors did. Use that review as a list of hypotheses to check against your recordings and survey answers, and act only on the ones the evidence supports.
What conversion rate should a small online shop expect?
There isn't a reliable single figure: rates vary widely by product, price, traffic source and season, and published averages mix very different businesses. Compare yourself with your own past months and track each funnel step separately. An improvement in the step where most people leave matters more than matching any benchmark.
Further reads
- Can a Small Online Shop Use AI to Compete With Bigger Brands? — Ways a small shop can compete once the leaks are fixed.
- How to Audit Your Website's SEO With AI in One Afternoon — Check whether the right visitors are finding you in the first place.
- How to Measure Whether AI Is Improving Your Marketing Results — Measure whether changes are improving results over time.
- How to Add an AI Chatbot That Captures Leads on Your Website — Answer the questions visitors leave over, while they're on the page.
- How to Create Customer Personas With AI From Real Data — Turn what you learn about buyers into grounded personas.
- Instant Roof Estimates on Your Website: Do They Win More Jobs? — How satellite-measured instant roof estimators work, why they raise lead volume but change lead quality, what they cost, and how to test one on your own site.
- How to Write Your Website Copy With AI: Home, About, and Services — Page-by-page prompts for your Home, About and Services copy, built from a fact sheet, with sample outputs, a claims check and a 20-minute customer test.
- Best AI Landing Page Builders for Small Businesses (2026) — Verified prices, traffic caps and AI features for the landing page builders small businesses actually use, with picks for three different businesses.
- How to Build a Quiz Funnel With AI to Capture Leads — Design the outcomes first, let AI draft the questions, test the scoring with fake personas, and give every result its own follow-up emails.
- Performance Max for Small Businesses: 9 Costly Mistakes — Nine Performance Max mistakes that drain small ad budgets, where each shows up in the account, and the setting or habit that fixes it.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Microsoft Learn, Clarity frequently asked questions (pricing, retention, masking, Copilot in Clarity); Google Analytics help on funnel explorations; Shopify help on its conversion reports.