The flag said PVI IDE

Our city pages carry a photograph of a street at golden hour. The American one showed a fire escape, a row of brownstones, and a Pride flag hanging from a window.

The flag read PVI IDE.

Not small. Not in the corner. Centre-frame, in focus, fully legible, and it had been live for weeks on all 19 United States pages and both Canadian ones.

How it survived

The image is generated, and generated images are bad at text. We knew that. We had checked the image — at thumbnail size, where it is a warm street scene with a rainbow in it and reads perfectly.

Nobody opened it at full width. So a word that is obviously wrong to any human who actually looks sat on twenty-one pages aimed at exactly the people most likely to notice a mangled Pride flag.

That's the bit that stings. It wasn't a subtle failure. It was an unexamined one.

What we changed

Every generated image gets opened at full size before it ships. Not glanced at in a grid, not approved as a thumbnail. Opened, at the size a reader will see it.

Anything with text in it gets read. Signage, flags, book spines, posters. If it contains letters, someone reads the letters. Generated text is wrong more often than it's right, and a thumbnail hides that completely.

Alt text describes the photograph, not the place. Ours used to say "a street scene in {country}" — but one image serves the US and Canada, another Denmark and Sweden. That was a claim about location the picture could not support, made specifically to the audience that has to take it on trust. It now describes the light and the subject and lets the page say where you are.

The wider thing

We use generated imagery because we are two people and cannot commission photography for seventy-one cities. That's an honest constraint and we're not going to pretend otherwise.

But there's a difference between using a tool with known failure modes and using it without checking the failure modes. The flag wasn't a surprise. Garbled text is the single most documented weakness of image generation, and we shipped it anyway because we looked at the picture the way we wanted it to be seen rather than the way it would be.

If you're building with generated assets: the check is not whether it looks good. It's whether it survives being examined. Those are different tests, and only one of them is the one your users will run.

Know someone this would help?

Keep reading

We gated our best writing to grow. Google saw 144 words.

A conversion decision that quietly cost us every reader who might have arrived.

We built 71 city pages. A quarter of each one was ours.

What happened when we finally measured our own programmatic SEO instead of admiring it.