Why We Show You the Screenshot the Robot Saw

Every GazeSite report includes the screenshots the audit was based on — the page as it rendered in our browser, desktop and mobile, at the moment we graded it. This seems like a small feature. I’ve come to think it’s one of the most important decisions in the product, and the reason is trust.

An audit is a series of judgments. Judgments are only as good as their evidence, and a judgment whose evidence you can’t inspect isn’t really a judgment you can use. It’s an assertion. If a report tells you “your headline is buried below the fold on mobile” and you can’t see what the auditor saw, you’re stuck in an argument with a black box. Maybe it’s right. Maybe it loaded some broken variant of your page. Maybe your A/B test served it the old design. You have no way to know, so you either take it on faith or dismiss it, and people mostly dismiss it. Showing the screenshot changes the epistemics entirely. The claim comes stapled to its evidence. You look at the picture, you look at the finding, and you can verify one against the other yourself. The report stops asking to be believed and starts asking to be checked, which is a much stronger position.

There’s a deeper reason, though, which took me longer to appreciate. The screenshot answers a question most site owners have never actually confronted: what does my page look like to a machine that isn’t me?

You have never seen your website the way a first-time visitor sees it. This sounds like an exaggeration but it’s nearly literal. You visit with a warm cache, an active session, dismissed cookie banners, subscribed-newsletter cookies, and a monitor and browser you chose. The audit visits with none of that. Clean profile, cold cache, no cookies, nothing dismissed, a standard viewport. It sees the cookie consent wall in full. It sees the newsletter popup you stopped noticing years ago. It sees the layout at a phone width you never test because your phone is newer than your customers’ phones. When people are surprised by their screenshot — and they often are — the surprise is almost never that our browser rendered the page wrong. It’s that the page really greets strangers this way, and the owner is seeing the greeting for the first time.

Sometimes the screenshot is the finding. A page that renders blank because a script failed. A hero image that never loaded, leaving text floating on nothing. A consent dialog that covers the entire viewport with a button that doesn’t work at mobile widths. In these cases the six reviewers’ notes are almost beside the point; the picture does the arguing. And this cuts both ways, honestly. When something goes wrong on our side — a timeout, a page that blocked the fetch, a rendering hiccup — the screenshot exposes us too. If we graded a half-loaded page, you can see that we graded a half-loaded page, and you know to weigh the findings accordingly. An audit tool that showed only conclusions could quietly paper over its own bad captures. Publishing the evidence keeps us honest in both directions, and I consider that a feature, not a risk.

The screenshot is also what makes findings arguable in the productive sense. Reviewers, human or machine, are sometimes wrong. When a finding comes with its evidence, disagreeing with it is cheap and concrete: you look at the same image and say no, the price is visible right there. Without the evidence, disagreement degenerates into arguing about what the page probably looked like, which is a waste of everyone’s time. I’d rather ship a report where every claim can be cross-examined than one that’s protected from cross-examination by hiding its sources.

There’s a practical benefit that has nothing to do with trust: the screenshot is a timestamp. Reports accumulate, and each carries the page as it was that day. When a re-audit shows a finding fixed, the before and after screenshots are the proof, and they’re strangely satisfying — the visual equivalent of a test going from red to green. When something regresses, the pair of images usually tells you when and roughly why. Your website’s history is otherwise unrecorded; sites change silently and nobody keeps the old versions. A stack of dated screenshots turns out to be a useful archive that nobody set out to build.

The underlying principle is one I’d apply to any tool that renders judgments: show your work. A conclusion you can trace is worth more than a confident one you can’t, and the difference compounds with every report. Machine judgment is only going to become more common, in this domain and every other, and the habit we should demand of it from the start is that it hands over the evidence along with the verdict. So the robot takes a picture of exactly what it saw, and we show it to you, unedited. What you do with the findings is your call. But you’ll never have to wonder what we were looking at.