Real images for app and site builders

Agents build content sites now: travel guides, catalogs, directories, field guides. The text generalizes fine. The images do not. Every page needs the one specific thing the page is about, and a generator would quietly fake all of them; readers who know the subject would know. Below is the honest version of that problem, live on this page: a small field guide, exactly as an agent scaffolds it, and then the same build after twelve retrieval calls filled its image slots.

Flip it yourself. The placeholder state is real, that is what every content build looks like the moment before images. The filled state is real too: every photo came back from a Lightdrift query naming the species, and each one is the actual fish, from National Park Service stream surveys, public domain, license chip rendered from the rights answer.

variegate darter (Etheostoma variatum), full body, adult
rainbow darter (Etheostoma caeruleum), full body, adult
greenside darter (Etheostoma blennioides), full body, adult
snubnose darter (Etheostoma simoterum), full body, adult
shield darter (Percina peltata), full body, adult
Roanoke darter (Percina roanoka), full body, adult
common shiner (Luxilus cornutus), full body, adult
redbreast sunfish (Lepomis auritus), full body, adult
rock bass (Ambloplites rupestris), full body, adult
redfin pickerel (Esox americanus), full body, adult
mottled sculpin (Cottus bairdii), full body, adult
American eel (Anguilla rostrata), full body, adult

Why this needed retrieval

Look at the darters. Six of the twelve entries are small striped fish from the same family, several from the same genus, and the differences that matter, the orange spot rows, the dorsal band, the snout, are exactly what a generator blurs and exactly what keyword stock search cannot rank. This is the long tail again: the more specific the page, the fewer real photographs exist, and the more it matters that you find the actual one. A guide like this is only trustworthy if every image is the species it claims.

The loop that filled the slots is the same two-endpoint loop from the rest of this blog. The agent already knows what each page is about, so the page subject is the query. The top result arrives with its rights answer, the build script keeps commercial-safe results and renders the license chip from the same field, and the credits section below was generated, not written.

for (const page of guide.pages) {
  const res = await lightdrift.search({
    query: `${page.common} (${page.latin}), full body, adult`, k: 3 });
  const img = res.results.find((r) => r.rights.commercial);
  page.image = { src: img.file, credit: img.rights.attribution };
}

Twelve pages, twelve queries, $0.24 of retrieval, and a guide where every image can face a reader who knows fish. That is the difference between a site that looks finished and a site that is finished.