How to diagnose product page performance: a 5-step framework

The product page is where a visitor decides whether to buy. Ads, emails and search results only bring people to it, and checkout only processes a decision that was already made. If the page leaves a question unanswered, nothing later in the journey can win the sale back.

That makes product pages the cheapest place to find more revenue. Getting 10% more visitors to add to cart costs nothing per visitor, while getting 10% more visitors usually means a bigger ad budget. Most stores also have a handful of products that carry most of their revenue, so fixing two or three pages can move the whole store.

The hard part is knowing what to fix. Redesigning a page because it "feels weak" is guesswork. This article covers a diagnosis in five steps, from broad numbers down to individual shoppers:

Each step narrows the question. Metrics tell you which page, funnels tell you where shoppers leave, and heatmaps show where attention goes on the page. Recordings and chat conversations tell you why.

Start with the numbers any web analytics tool reports, and look at them per product page rather than for the whole site. A site-wide conversion rate averages your best and worst pages together and hides both.

Then split each metric by traffic source and device. A page can convert well for returning visitors from email and badly for first-time visitors from a Meta ad, and those two groups need different fixes. A large gap between desktop and mobile usually points to a layout or speed problem rather than a product problem. Run the page through Google PageSpeed Insights to rule out slow loading on phones.

You are looking for two kinds of pages: ones with high traffic and a low add-to-cart rate, and ones with a good add-to-cart rate but few purchases. The first group loses people on the page itself. The second group loses them later, and the funnel shows where.

In Bigdelta, these numbers sit next to orders and customers, so a page's revenue per visitor comes from the actual orders rather than from an estimated value attached to a pageview.

A funnel follows the same visitors through an ordered set of steps and shows how many make it to each one. For a product page, a useful funnel looks like this:

The number to watch is the drop between two neighbouring steps, not the final conversion rate. A big fall between steps 1 and 2 means people leave before they engage at all, so look at the first screen: images, price, load time. A fall between 2 and 3 means they engaged but weren't convinced, which usually means missing information, unclear variants or doubts about delivery and returns. A fall after step 3 is often about shipping costs or payment options, not the page.

Build the same funnel for several segments and compare them. Useful splits:

If visitors from one campaign fall off far more than others, the ad may be promising something the page doesn't show. That is a campaign fix, not a page fix.

Bigdelta builds funnels from first-party, server-side tracking and keeps the same visitor across sessions. That matters for products people research over several days, because a shopper who views the page on Monday and buys on Thursday stays one person in the funnel instead of becoming one drop-off and one unexplained purchase.

The funnel tells you where people leave, not why. The next two steps look at what they did on the page.

A heatmap adds up the behaviour of every visitor to a page into one picture. It's the quickest way to see how a product page is used before you watch individual sessions. Two types matter most:

On a product page, look for these patterns:

View the heatmap separately for mobile and desktop, since the layouts differ and a section that's visible on desktop can sit three screens down on a phone. If your tool lets you, compare the heatmap of visitors from the funnel step with the biggest drop against those who continued.

A heatmap shows what the crowd does. It hides the order of actions and the individual hesitations, which is what recordings are for.

A session recording replays what a visitor did on the page: where they scrolled, what they tapped, where they paused and when they left. Watched at random, recordings waste hours. Watched for one funnel step, they answer the question the funnel raised.

So filter first. Pick the step with the biggest drop and open recordings of visitors who reached it but not the next one. In Bigdelta you can open these straight from the funnel step, so you watch the 20 people who picked a size and then left, not 20 random visitors.

While watching, look for patterns that repeat across several sessions:

One odd session proves nothing. When five of twenty sessions show the same behaviour, you have a hypothesis worth testing.

Also watch a few sessions from visitors who did buy. Comparing them with the ones who left shows what the buyers found that the others didn't, for example that buyers scrolled to the reviews and leavers never got that far.

Recordings still have a blind spot. You can see a shopper hesitate over the size chart, but you can only guess what they were wondering. They didn't type the question anywhere.

An AI chatbot on the product page closes the gap recordings leave. A shopper who hesitates over the size chart can type "I'm 182 cm, is L too long?" and get an answer from your product data. You get the question in writing.

That helps in two ways.

First, the shopper gets the missing information at the moment they need it, instead of leaving to search for it. A good product chatbot answers from the store's own catalogue and policies: sizing, materials, compatibility, stock, delivery times, returns. Algoshop, for example, is an AI sales chatbot for Shopify and WooCommerce stores that recommends products, checks stock and can add items to the cart from inside the chat. It also answers in the shopper's language, which matters if you sell across borders.

Second, and more useful for diagnosis, every conversation is a record of what your page failed to explain. Read them grouped by product, and the same questions come up again and again:

Each repeated question is a page change you can make: a line in the description, a comparison table, a delivery estimate next to the price. Algoshop surfaces the top questions shoppers ask, so you don't have to read every transcript to find them.

The chat also tells you which questions come up right before people leave. Put that next to the recordings from Step 4 and you can match the behaviour (pausing on the size guide) with the question behind it ("runs small or true to size?").

The five steps work best as a short routine on your top-revenue products rather than a one-off audit.

The last step is the one most stores skip. Change one thing at a time and check the funnel after a week or two of normal traffic, so you know which edit helped. Stores with plenty of traffic on the page can go further and A/B test the change with a Shopify Rollouts experiment (Grow plan or higher), which compares the new version against the live store. With less traffic, a before-and-after comparison of the funnel step is the realistic option. When the chatbot stops getting the same question about a product, that's a sign the page now answers it.

Bigdelta and Algoshop cover the five diagnosis steps. Two free tools help with checking page speed and testing your changes: