Every shelf metric your team owns describes a moment before the shopper arrived. Share of search describes the results grid. The Buy Box describes which offer got shown. Agentic commerce describes what happens when no grid is rendered at all.
The product detail page is different, and the difference matters. It is the only surface in the entire funnel that you write, you control, and the shopper actually reads. It is also the surface your team measures worst.
The short answer
A product detail page test measures whether a page closed a sale. It cannot measure what the shopper came to the page to find out. Those are different questions, and only the second one tells you what to change.
The reason is structural. By the time somebody loads your PDP, the comparison has already happened somewhere you cannot see - in a search grid, a review thread, a recommendation, an AI answer, a friend's message. The visit is usually not discovery. It is verification: the shopper arrives with a provisional decision and a short list of things that would kill it. Your page either resolves those things or it does not.
Conversion tells you the aggregate outcome of that check. It never tells you what was on the list.
What the benchmark data actually says
Baymard Institute has been scoring product page UX against the leading US and European ecommerce sites for years. Its most recent state-of-PDP analysis (updated 18 March 2026) is built on 30,000+ manually reviewed product page usability scores across 155+ benchmarked sites, distilled into 115+ product page guidelines.
The headline is not flattering:
- Only 48% of leading desktop ecommerce sites have a "decent" or "good" product page UX performance.
- On mobile that falls to 38%.
- In apps it falls further, to 36%.
So on the device where most shopping now happens, roughly three sites in five are running a product page that Baymard rates mediocre or worse - and these are multi-million-dollar operations, not hobby stores.
The specific gaps are more useful than the score, because they are all information gaps rather than aesthetic ones:
- 81% of sites omit price per unit.
- 67% do not show an estimated total order cost on the product page.
- 44% have no link to the return policy.
- 89% do not respond to negative reviews.
- 63% lack navigation for customer review images.
Read that list again with the verification framing in mind. Price per unit, total cost, return policy, and what unhappy buyers said are not decoration. They are precisely the four things a shopper checks when they are trying to talk themselves out of a purchase. The industry has built beautiful pages that are silent on the questions people came to ask.
The vendor number nobody has audited
Here is the other half of the problem. Amazon's own merchant-facing page for A+ Content states that "Basic A+ Content can increase sales by up to 8% - and well-implemented Premium A+ Content can increase sales by up to 20%." Both figures carry a footnote reading, in full, "Amazon internal data."
No methodology. No control definition. No sample. No time window. No statement of what "well-implemented" means, which is the load-bearing word in the sentence.
This is not a scandal - it is an advertisement, and advertisements are allowed to say "up to." The problem is what happens next. That number gets pasted into an internal business case, loses its "up to," acquires a decimal point, and becomes the expected return on a six-figure content program. A team then measures the launch against it, sees 2%, and concludes the creative was weak. The honest conclusion is that they never had a comparable number to begin with.
The general rule is worth writing on the wall: a platform's estimate of the value of the platform's own feature is a marketing input, not a research finding. It belongs in the same bucket as network-reported ROAS, which we covered in retail media measurement.
The three things a PDP test cannot see
| What the test reports | What it silently assumes | What is actually happening |
|---|---|---|
| Variant B converted 4% better | Both variants faced the same shoppers | Traffic mix shifts with campaigns, season, and rank; the shopper populations differ |
| Adding the size chart lifted conversion | The chart answered the question | The chart may have delayed the exit long enough for a different element to land |
| The long description underperformed | Shoppers rejected the content | Most shoppers never scrolled to it; you measured placement, not content |
| No variant beat control | The page is fine | Every variant missed the one unanswered question, so all of them lost equally |
The fourth row is the expensive one. A flat test result reads like evidence that the page is optimized. It is equally consistent with a page that is failing everybody in the same way - and a test can never distinguish those two, because the winning variant is only ever the best of the options you happened to imagine.
There is a second blind spot that shows up on the P&L rather than the dashboard. A page can raise conversion by overselling, and the cost arrives months later in the returns line. The National Retail Federation's 2025 Retail Returns Landscape, produced with Happy Returns and based on a summer 2025 survey of 2,006 consumers and 358 ecommerce professionals at US merchants above $500 million in revenue, put total 2025 returns at $849.9 billion, a 15.8% return rate. Online is materially worse: 19.3% of online sales were expected to come back.
Roughly one in five online sales reversing is not a logistics statistic. It is a listing accuracy statistic. Every returned item is a shopper whose expectation, formed by your page, did not survive contact with the product.
And the reputational cost has been rising. NRF found 71% of consumers say they are less likely to shop with a retailer again after a poor returns experience, up from 67% in 2024, and 82% now cite free returns as an important consideration when shopping online, up from 76%.
What to research instead
The question a PDP test cannot answer is simple to ask directly: what did you need to know, and did the page tell you?
That is a conversation, not a click. The instrument has to be able to hear "I wanted to know if it fits in a standard cupboard and I gave up looking," which no multivariate test will ever surface, because nobody thought to write a cupboard-dimension variant.
A well-formed listing study has five parts:
1. Recruit on the behaviour, not the outcome. Talk to people who viewed and did not buy, people who bought, and people who bought and returned. The third group is the most under-used sample in ecommerce and the cheapest to reach, because you already have their contact details and a reason to make contact.
2. Get the arrival state before you show anything. What did they already believe about the product when they landed, and where did that belief come from? This is the only way to separate what your page did from what the review site, the recommendation, or the AI answer did.
3. Reconstruct the check list. Ask what they were trying to confirm and what would have stopped them. Do it before showing the page, so the page cannot prime the answer.
4. Then expose the page and probe the gap. Now you can ask what they looked for and could not find. The difference between the list they gave you in step 3 and the list the page answers is your actual backlog, ranked by how often each item appears.
5. Close the loop with returners. Ask what they expected and where the expectation came from. Point at the element. That is a direct line from a sentence on your listing to a line item in your reverse logistics cost.
Here is how that maps onto a structured study, using all six of Koji's question types alongside the AI conversation:
| What you need | Question type | Example |
|---|---|---|
| The arrival belief | open_ended | Before you opened the page, what did you already think this product was? |
| Whether comparison happened elsewhere | yes_no | Had you already compared this against something else? |
| The pre-page check list | multiple_choice | Which of these did you want to confirm before buying? |
| What actually mattered most | ranking | Rank price, fit or size, delivery, materials, and reviews |
| Confidence after reading | scale | After reading the page, how confident were you that it would suit you? |
| The unresolved question | single_choice | Which question did the page leave unanswered? |
The open_ended item in row one is where the value sits, and it is the one traditional survey tools handle worst. A static form takes "I wanted to check the size" and stops. Koji's AI moderator hears the same answer and asks what they were comparing it against, where they looked, and what they would have accepted as proof - and it does that with every respondent, at 2am, without a moderator to schedule.
Where Koji fits
Traditional listing research is a scheduling problem before it is a research problem. UserTesting and dscout mean recruiting panels, booking sessions, and waiting; Qualtrics and SurveyMonkey give you speed but flatten every answer to the depth of the box you drew. Dovetail helps you organise findings you have already paid a moderator to collect.
Koji collapses the loop. You write the brief, Koji generates the interview guide, and AI-moderated voice interviews run at whatever hour your shoppers are actually free. Every transcript is analysed thematically as it lands, so the unanswered questions surface as a ranked list rather than a folder of recordings. There is no moderator whose phrasing drifts by interview 30, and no researcher headcount standing between a listing hypothesis and an answer. Teams go from question to insight in hours, not weeks, and no research background is required to run one.
You can pair it with the structured questions above to get quantified counts alongside the verbatim reasons, and route the study through the cart abandonment research guide when the drop-off is at checkout rather than on the page. For the physical-pack equivalent of this work, see packaging concept testing; for the words themselves, content testing.
What to do in the next quarter
- Pull your top 20 listings by revenue and check them against the five Baymard gaps above. Price per unit, total cost, return policy link, review images, and negative-review responses cost engineering days, not quarters.
- Stop quoting the platform's own uplift figure in business cases. Replace it with a measured baseline from your own catalogue.
- Run one listing study against your highest-return SKU. Returns give you a pre-qualified sample and a hard cost to compare against.
- Add "what did you expect, and what told you that?" to your returns flow permanently. It is the cheapest continuous listing research that exists.
The page is the one part of the digital shelf you actually own. It is worth knowing what it is failing to say.
Related reading
- Share of Search and the Digital Shelf (2026)
- Winning the Buy Box (2026)
- Agentic Commerce Research (2026)
- Customer Research for Ecommerce Brands
- Choice Architecture and Defaults (2026)
- Product Returns Research (2026) - why the reason code measures your refund policy, not the cause.
Frequently Asked Questions
Is A/B testing a product detail page worth doing at all?
Yes, for choosing between options you have already generated. A/B testing is a good selection mechanism and a poor discovery mechanism. It tells you which of the variants you imagined performed best; it cannot tell you about the variant nobody imagined, which is where most of the upside sits. Use interviews to generate the candidate list, then test to pick the winner.
Why can a product page test come back flat?
Two very different situations produce the same flat result. Either the page is genuinely well optimized and your variants were marginal, or every variant failed to address the one question shoppers actually had, so they all lost equally. A test cannot distinguish these, because the comparison is only ever between the options you wrote. Talking to non-buyers separates them immediately.
Should we trust Amazon's 8% and 20% A+ Content figures?
Treat them as marketing claims, not measurements. Amazon's own page footnotes both to "Amazon internal data" with no published methodology, no control definition, no sample, and no time window, and both are phrased as "up to." They are fine as a directional argument for investing in listing content and unsuitable as a forecast in a business case. Measure your own baseline instead.
How do product returns relate to listing research?
Directly. The NRF put the 2025 online return rate at 19.3%, and a large share of those are expectation failures rather than defects - the product was fine, but it was not what the page implied. That makes returners the single most informative sample you can interview about your listings, because the mismatch is specific, recent, and already cost you money.
How is this different from cart abandonment research?
Cart abandonment research asks why somebody who had decided to buy did not complete checkout - the failure is in cost, trust, or process. Listing research asks why somebody never decided at all, or decided wrongly. Both matter, and they point at different teams: abandonment findings usually go to checkout and payments, listing findings go to merchandising and content.
How many interviews do we need for a listing study?
For a single listing, 15 to 25 conversations across viewers, buyers, and returners will surface the recurring unanswered questions; the same two or three items start repeating quickly. What matters far more than the count is sample composition - 25 buyers will tell you your page is excellent, because it worked on them. Non-buyers and returners carry the information.
