AdSpark AI: A Search Campaign from a URL

Someone who has never run a search campaign faces the same wall every time: a blank keyword list. Everything downstream is judgement they can learn, but the first step needs volume they do not have. A class can be taught the structure of a campaign in an hour and still not be able to start one.

The industry answer is to open Keyword Planner, which needs an account, a billing method and a tolerance for its interface. For a teaching session that is three obstacles before the lesson begins.

An application that starts from the only thing a participant reliably has: the website.

You paste a URL and choose the objective — awareness, traffic, lead generation or sales. The application reads the site, stores the text it found and writes a short summary of what the firm does. From that it generates a keyword list.

The part that matters is what happens next. Every keyword in the table has its own button. Press it and the application writes ad copy for that keyword — headlines, display URLs, descriptions — and a landing page checklist for that keyword, against that objective. Nothing is generated for the list as a whole, because a campaign is not planned at the level of a list.

It is the same job as the Google Ads Agent System, approached from the opposite end. That one is a team of agents producing a bulk upload file for someone who already runs campaigns. This one is a single screen for someone who has never run one.

Twenty-four people have used it across twenty-seven projects, on real brands rather than exercises: a pet food company, a travel operator, a pet retailer, a fintech, a B2B manufacturer, a cosmetics brand. One project runs the app against my own IIM Calcutta programme page.

All four objectives have been used. Every project completed.

The number worth reporting is the shape of what came out. The application generated 649 keywords and wrote ad copy for 41 of them. That is around twenty-four keywords produced per project and copy requested for one or two. People want the list. Almost nobody works the list.

No live campaign results are recorded. Nothing here was spent against, so there is no click-through rate, no cost per lead, and no claim that the copy performs.

The bug I could not fix was a bug I should have deleted. Clicking a keyword’s AI assist kept returning the first keyword’s results. The first fix passed the keyword more carefully. The second made generation automatic on selection. The third fixed the selection logic again. None worked, because the problem was that a single “selected keyword” could be stale at the moment the request went out. The version that worked abandoned selection entirely and put a button on every row. The state that was wrong no longer existed.

Validate the shape at the boundary. An early failure had the model returning keywords correctly but wrapped in an object where the application expected an array. The content was right and the application was broken. Anything crossing from a model into a program needs checking for shape before it is used, not after it fails.

Generation is cheap and attention is not. Six hundred and forty-nine keywords cost almost nothing to produce. Forty-one of them got a human’s attention. If I built this again I would generate far fewer and make the application argue for them, because the scarce resource in the room was never the keywords.