The thirty thousand product pages Baymard reviewed, and what they said about reviews.

In 2009, a small Copenhagen research firm called Baymard Institute began recording usability sessions of people buying things online. They have not stopped. As of their 2025 benchmark update, Baymard had run more than 130,000 hours of moderated UX testing across roughly thirty thousand product page experiences on major e-commerce sites. The findings are the most empirically grounded body of evidence we have on how buyers behave on a product page. They are also almost never quoted by the review platforms whose widgets the data describes.

This essay reads Baymard's findings on reviews carefully, then asks a follow-up question Baymard does not ask. The findings describe how a human buyer reads a review widget. The follow-up question is about a second reader who has arrived on the page in the last twenty-four months: the AI crawler. The shape Baymard's data describes is, structurally, almost the opposite of what that second reader needs.

Baymard's session recordings show that buyers spend disproportionate time on the review section of a product page. In their 2024 report on product-detail page UX, more than 88 percent of test users consulted reviews before adding to cart on any product priced above a threshold (the threshold varied by category; for skincare, around $30; for electronics, around $100). For the subset of buyers who reached the reviews, the median time spent reading them was 32 seconds. The longer-tail (the slowest 10 percent of readers) spent more than two minutes inside the review block.

This number contradicts the operator's default assumption. The dashboard shows the brand a click-through rate on the "read reviews" button or a scroll-depth metric to the review section. The numbers tend to be low. The buyer who is reading reviews is usually not interacting with the buttons. She is reading. Scroll-depth metrics undercount actual reading time because most buyers scroll past the reviews, scroll back, and then dwell.

Baymard's qualitative observations from the same study: buyers read reviews to find people who match their use case ("woman in her 40s with combination skin"), to find the negative reviews specifically (eight out of ten test sessions filtered or sorted for one-star reviews), and to look for words that match a concern they have already named to themselves ("does it sting," "does it pill under sunscreen").

The pattern matches what we have argued in the long tail begins inside the review. The reviews are where the buyer goes to verify a specific question. The question is in her language, not the brand's. The widget either contains her question or it does not.

The second Baymard finding contradicts the first. Across the same 2024 study, 41 percent of test sessions failed to locate the review section on the product page within the first thirty seconds of arrival. The failure modes Baymard documented: reviews placed below a long block of brand-written marketing copy, reviews placed below an expanded ingredients table, reviews placed inside a collapsed accordion captioned with a label other than "reviews," reviews placed entirely below the fold on mobile.

The discovery problem is structural. Buyers want reviews; pages do not surface them. Baymard's recommendation, repeated across multiple benchmark updates since 2017, has been to place the review block within the first two screens, label it explicitly, and pre-expand it. The recommendation is widely ignored. Of the top 50 US e-commerce sites Baymard benchmarks, fewer than half meet the recommendation as of the 2025 update.

A specific Baymard finding worth quoting: on mobile, the median scroll distance to the review block on a top-50 US product page is 4.2 screen heights. Buyers who do not scroll past three screens (roughly 27 percent of mobile sessions in the study) never reach the reviews. They make the buy decision on brand-written content alone.

Baymard's third finding is the most pointed and the least repeated. Review widgets, particularly the polished aggregated widget shapes shipped by major platforms, are read by buyers as marketing artifacts and discounted accordingly. In session recordings, buyers verbalised mistrust of widgets that: showed an unrealistically clean star distribution (the 4.8-with-no-1-stars pattern), placed positive reviews above negative ones with no apparent sort logic, included reviewer photos that appeared stock-staged, or buried the negative reviews behind pagination or filters.

The Baymard observation, in the 2023 report: "Users frequently described aggregated review widgets as 'something the brand controls,' even when the underlying data was user-generated. The presentation of the widget, more than the content of the reviews, determined whether the buyer treated the reviews as credible."

This is a structural observation about UI shape. The widget shape implies brand control. The buyer reads the shape, then discounts the content. The reviews are real and the buyer is treating them as if they are not.

Baymard's specific recommendation for trust recovery: show a realistic distribution (let the 1-stars stand at the top of the negative tab), show reviewer names in full where the reviewer consented, show date and verified-purchase status prominently, allow the buyer to sort by "most critical" with a visible button, and link to the reviewer's other reviews where the platform supports it. Each of these is a structural change to the widget. Almost none of them are defaults.

In reviews are language not inventory, we argued that the widget treats reviews as countable units of inventory rather than as a body of writing. Baymard's trust finding is the empirical version of the same observation. The buyer treats the widget the way the widget treats the writing: as a polished surface, not as testimony.

Read together, the three findings name a contradiction the widget category has not solved. Buyers want to read reviews (finding one). Buyers cannot find reviews on the page (finding two). When buyers do find the widget, they discount what is in it (finding three).

The widget's job, on a typical e-commerce site in 2026, is to surface user-generated language to a buyer who wants to verify a concern. The widget does all three jobs badly. It is hidden, it looks like marketing, and it presents the writing as a polished aggregate rather than as individual paragraphs of human testimony.

Baymard's recommendation set, taken seriously, would dismantle the standard widget shape. Pre-expand the reviews. Show the negative reviews first or alongside. Show real names, dates, and verified-purchase indicators. Allow sorting that exposes the worst reviews. Render in HTML, not in a polished iframe that looks like an ad block.

The widget, dismantled along Baymard's recommendations, starts to look like something else. It starts to look like a list of paragraphs. Each paragraph has a byline. Each paragraph has a date. Each paragraph is sortable, but the default view is reverse chronological with the most recent at the top. The aggregate star count is present but not the dominant visual element. The reviews themselves are the page.

That is the shape Baymard's data points at. It is also the shape the citation engine now needs.

The second reader on a product page in 2026 is not a buyer. It is GPTBot, ClaudeBot, PerplexityBot, Google-Extended, and the half-dozen other crawlers whose user-agent strings have been arriving at e-commerce origins since mid-2023. They are reading the same page the buyer reads. They are reading it differently.

What the second reader needs, in mechanical terms: HTML in the initial document (not loaded by JavaScript), individual review paragraphs as addressable units (not paginated behind a click), authorship metadata in the HTML (name, date, verified-purchase status), and clear topical headings that let the engine route a query to the right block.

What the second reader does not benefit from, even when present: the aggregate star rating widget, the photo carousel, the "AI summary" feature shipped by most platforms in 2024 and 2025, the polished iframe.

The mapping is direct. The Baymard recommendations to recover buyer trust are also the recommendations to be citable. Show the real reviews. Show the names. Show the dates. Show the negative ones. Surface them above the fold. Render them in HTML. The page that does these things is more trusted by the buyer and more readable by the engine. The page that does not is both lower-trust and lower-cited.

The widget shape Baymard's research describes is roughly the opposite of what AI crawlers need. The cure for the trust problem is the same edit as the cure for the citation problem. There are not two design briefs. There is one.

It is worth being precise. Bazaarvoice, Yotpo, Okendo, Junip, and Trustpilot have all, at various points since 2018, published blog content acknowledging Baymard's research. The acknowledgements have generally taken the form of "Baymard finds reviews are important, so install our widget." The recommendation Baymard actually made (rethink the widget shape, show negative reviews prominently, render in HTML) has not been adopted in any of those platforms' default install paths.

The commercial reason is unsurprising. A widget that pre-expands negative reviews and renders in HTML is, for the platform, a less visible product. The platform's brand depends on the widget being recognisably the platform's widget. Render in HTML, in the merchant's own template, and the platform's logo, distinctive star icon, and "Powered by [Platform]" footer become harder to maintain.

The Baymard recommendation is, in a category sense, an argument for less visible review software. The platform incentive is for more visible review software. The two are in tension. The tension explains why Baymard's findings have been quoted in marketing copy but rarely implemented in product.

Three steps an operator can take this week. The order matches the Baymard recommendation priority.

Step one. Find the review block on a current product page. Open it on mobile. Count the screens to scroll. If the answer is more than two screen heights, the review block is below where most buyers will look. Move it. The change is template-level in most stacks (Shopify themes, Magento blocks, headless React/Next compositions) and takes minutes. The buyer payoff is immediate; the citation payoff lags by the engine's crawl cadence, typically two to six weeks.

Step two. Sort the reviews so that the most critical and the most recent are visible by default. If the platform's widget does not allow this, the platform is the problem. The interim fix is a one-page server-rendered list that shows the ten most recent reviews above the platform's widget. The platform's widget can stay as the deeper archive.

Step three. Strip the AI summary feature, if the platform ships one and you have it enabled. The Baymard trust finding and the December 2025 FTC warning letters (covered in the end of the review widget and elsewhere) point in the same direction. A machine-written paragraph above the real reviews depresses trust and creates regulatory exposure. The brand that does not summarise the reviews and instead surfaces them directly wins on both axes.

After those three, the work is at the architectural level addressed in the half life of a product page. The product page that compounds is the page that accumulates real customer paragraphs, dated and signed, readable by humans and by engines. Baymard's data, accumulated since 2009, points at exactly this page. The page has been in plain sight for fifteen years.

A research institute spent fifteen years recording what buyers actually do on product pages. The findings have been published continuously, quoted selectively, and implemented sparingly. Buyers want reviews. Pages hide reviews. Widgets damage the trust of the reviews they do show.

The same fifteen years saw the rise of a new reader of the same pages: a class of AI crawlers whose citation behaviour now drives a meaningful share of high-intent product traffic. The new reader needs roughly the same thing the buyer always needed. Real paragraphs of human writing. Visible. Dated. Signed. In HTML. Sortable. Not summarised by a machine and not buried below the fold.

Two readers, separated by twenty years of product research, arriving at the same product page, asking for the same content shape. The operator who reads Baymard's findings as a content brief, not as a UX nicety, ends up with a page that satisfies both. The page is also, not coincidentally, the page that ranks for the long tail, accumulates link equity, and reads as a document rather than a widget. The brief was always there. The category has not yet built to it.