Semrush surfaced thousands of images with missing alt text. Instead of treating it like a manual cleanup task, I built a pipeline.

Every site accumulates image alt-text debt, and nobody is ever excited to pay it down. A Semrush crawl on this project surfaced thousands of missing descriptions, enough that fixing them by hand would take weeks and produce inconsistent results across the board. The question stopped being "who writes these?" and became "what system writes these well enough, consistently enough, to actually ship?"

Alt text touches accessibility, content quality, SEO hygiene, and operational discipline all at once. When one fix improves four layers of a site, it deserves more than a spreadsheet and a Friday afternoon.

The pipeline I built had four stages: pull flagged images from the Semrush audit, analyze each image, generate alt text against a defined ruleset, and push the result into the CMS in a structured format. For the analysis and generation layer I chose Gemini Flash-Lite 3.1 because the task needed speed and cost efficiency more than raw capability. Each image required real interpretation, but the output format was narrow enough that a lighter model performed just as well as a heavier one.

Generating words for an image is trivial. Generating descriptions that stay useful across thousands of images, match editorial tone, avoid keyword stuffing, and arrive in a CMS-ready format required encoding judgment as rules. Maximum character count. No decorative adjectives. No starting with "Image of." Match the page context when available. These constraints did more for output quality than any prompt engineering trick.

The quality bar was specific: descriptive without bloat, useful without over-optimization, consistent with how the rest of the site reads. Vague alt text adds nothing for a screen reader user; overlong alt text clutters the experience; keyword-stuffed alt text reads like spam to everyone. Every rule in the pipeline existed to hold that line, and when early outputs drifted generic, the fix was always tighter constraints on the formatting and review layer, never a "better prompt."

The operational leverage came from the full chain (audit to generation to rules to CMS update), not from the model's raw capability. Every conversation about AI workflows eventually gets to "which model did you use?" but the model mattered less here than the constraints, the formatting rules, and the deployment path into the CMS. The pipeline is the product; the model is a component.

The most transferable insight from this project is that the best AI problems in a production environment are boring. Repetitive, frequently neglected, requiring just enough judgment that you cannot write a regex for them: metadata enrichment, classification, cleanup, QA checks, structured publishing workflows. One systematic improvement in any of these compounds across accessibility, search performance, and content operations simultaneously, which is a better return than most "AI strategy" initiatives deliver.

I keep coming back to alt text as a reference project because it clarified where AI work gets interesting for me: not in the generation layer, but in the infrastructure layer that connects an audit finding to a published fix without a human bottleneck in between. The next version of this pipeline will fold in page-context awareness so the alt text can reference what the surrounding content is about, not just what the image contains. Whether that moves the SEO needle or just makes descriptions more useful for screen readers is an open question, and probably the wrong one. The better question is why most teams still treat this kind of work as a manual chore when it clearly does not have to be.