Your boldest claim is the one nothing checks

We keep a checker that reads our own marketing pages after every build and complains when something on them has stopped being true. This morning we found that one of its older rules had been reading half of what it was pointed at.

The rule covers a row of short statistics near the top of several pages: a figure in bold, then a few words of explanation after it. It exists because one of those figures used to say “10s every film” and stayed on the page for weeks after the product began offering four different lengths. So the rule hunts for absolute words - every, all, always, any, each - and when it finds one it demands that whoever wrote it name the list the word is claiming to cover.

It was hunting in the explanation. It was not hunting in the bold figure. An absolute in the small grey half got caught; the same absolute in the large black half went straight through. We proved it by putting the words “Every film” into the bold slot of a real page and running the check, which passed, cleanly, and printed a tidy summary line saying so.

The prominent half is the unread half

Nobody decided the bold half mattered less. The opposite: it is set in bold because it matters most. The gap opened because checking is done on sentences, and that half is not a sentence. It is two or three words with no verb in them, so it reads as decoration to anything built to parse prose - including, it turns out, a person proofreading quickly.

This is worth sitting with, because it inverts what you would assume. The claim a customer is most likely to read, remember and repeat back to you in a complaint is the claim your process is least likely to have examined. Prominence and scrutiny are supposed to travel together. In practice they are inversely related, and the mechanism is as dull as word count.

In a shop the unread half is not even text

Our version of this was recoverable, because the bold figure was still characters in a file. In a shop it usually is not. The prominent half of the claim has been flattened into an image, and at that moment it leaves the reach of every tool you own. You cannot search it, spell-check it, find-and-replace it, or list it. It will not appear in an export. It is a picture of a sentence, and pictures do not get audited.

• The badge on the product photo. “Free shipping”, “handmade”, “ships in 24 hours”, set in a corner of the main image. The day the free-shipping threshold moves, you edit the shipping page, the cart copy and the footer. The badge is on forty photographs and it now contradicts all three.
• The number on a sale graphic. A banner reading forty per cent off outlives the sale by however long it takes somebody to notice, and it is the first thing on the page.
• The size chart saved as an image. Correcting a sleeve measurement in text is typing. Correcting it in a chart means finding the original file, and if the person who made it has gone, the wrong number is now permanent.
• The end card on a video. A price, a discount code or a web address burned into the last two seconds. The code expires; the film keeps offering it, in the most-watched position a video has.
• Text inside a thumbnail. Chosen to win the click, written once, and never revisited, because it lives in a different system from the listing it is advertising.
• Anything printed. Inserts, labels, stickers, packaging. These are ordered months ahead, which means they encode a claim from a version of the shop that no longer exists.

The common thread is not carelessness. Each of these was correct when it was made. They drift apart afterwards because one half of the claim is cheap to edit and the other half is expensive, and the cheap half is the only half anybody remembers to visit.

The test is whether a search box can find it

There is a quick way to inventory this and it takes about twenty minutes. Pick your three loudest claims - the ones you would be embarrassed to be wrong about. For each one, go to your admin, open the search box, and type the claim. Not a paraphrase: the words themselves.

What comes back is the half of your shop that is maintainable. What does not come back still exists, is still being read by customers, and is now on a list. That list is the deliverable. It does not need to be long or clever; it needs to name each uncheckable claim and say where the editable original lives, so that the next time the underlying fact moves, the move has somewhere to propagate to.

The reason to write down the location rather than just the claim is that the expensive half is expensive mostly because of a lookup. Knowing that the free-shipping badge is a layer in one named file, in a named folder, turns half a day of archaeology into a ten-minute job. Most claims that go stale in an image go stale because re-exporting it was briefly annoying.

What not to do about it

The tempting response is to ban text in images. Do not. Text in images works - it is prominent, it survives being screenshotted, and it travels into contexts where your careful product copy does not follow. Removing your most effective surface to solve a bookkeeping problem is a bad trade, and you will quietly reintroduce it within a month because it works.

The second tempting response is to promise to remember. This is the one everybody picks, and it fails for a specific reason: the moment you need the memory is the moment you are busy changing something else. You are editing the shipping threshold, not auditing photographs. The claim you need to recall is attached to the thing you are not looking at.

What works is less interesting. Decide where each loud claim is allowed to live, keep the list of the places a search box cannot reach, and treat that list as part of the claim rather than as a chore attached to it. Our checker now reads both halves of its stat, which is the same fix at a smaller scale: not more vigilance, just pointing the check at the part of the sentence nothing was reading. It is a close relative of a count that goes stale because nothing holds it to the set it counts, and of the way an overlay decides more than you typed.