Prompts to Analyse Competitor Ads

An ad is the shortest document a competitor publishes. Six words of headline, maybe twenty of body. Everything else that makes it an ad sits outside the copy: the image or video, the page it points to, and the people it was shown to.

Paste the words on their own and you have handed over the smallest part of the thing. A model asked what the campaign is doing will answer anyway, because six words plus a request for interpretation is a writing task rather than a reading one.

The specific failure is worth naming, because it is easy to miss. Ask about visual language and you will get a paragraph about visual language. It will be plausible, it will use the right vocabulary, and there is nothing in it to check. Nothing in the answer says the model never saw an image.

An ad set is a list of experiments

Then there is the question of what the set even is. A pricing page is a decision. A filing is a statement under liability. An ad set is a collection of things somebody is currently trying, and a good proportion of it will be switched off within a month.

So reading an ad set as a competitor’s positioning is a category error. It is their positioning experiments. The interesting question is not what they say but which of the things they tried is still running, and that question needs dates rather than cleverness.

Their considered position is on their own pages, written once and defended, and reading a competitor’s positioning is a different job with a different source. An ad set is where they are still arguing with themselves about it.

This job normally comes with a measured run: the prompt, what came back, and the source it was checked against. Here there is none, and the reason is the same thing that makes the job hard.

The public ad libraries are interfaces rather than documents. They do not give an ad set a fixed address you could open in six months, they refuse automated requests, and the ads themselves disappear from the record once they stop running. A run printed here would rest on a screenshot, and a screenshot is not something a reader can check.

That is worth stating plainly rather than filling the gap. A result you cannot reproduce is not evidence, and manufacturing one on a page about unverifiable claims would be an odd way to make the argument. The same call was made for review analysis, where no public corpus qualified either.

Two prompts, and the second one runs first. Establishing what is missing before the reading is what stops the gaps being filled quietly.

Note the instruction to flag uncertainty inside the sentence rather than in a note at the end. A caveat at the bottom gets dropped the first time somebody copies a paragraph into a deck. A clause inside the sentence travels with it.

Three inputs, and the second one is a job you have to do by hand. Writing one clause per ad about what it shows takes a few minutes across a whole set and changes what every later answer is made of.

The ads you just read are the ones running today. That is not a limitation of the prompt, it is what the libraries hold. An ad stops, and within weeks there is usually no public record that it ever ran.

The cost is specific. What you lose is the history, and the history is where the signal lives. A claim a competitor ran for six months and then dropped tells you something. So does one they have quietly run all year. Neither is visible in a single visit, and neither can be recovered later by anybody who was not saving.

A jump in live variants is also one of the first visible things a competitor does with new money, which makes an ad library worth checking the week after a competitor announces a round rather than a quarter later.

There is also a floor no amount of watching lifts. Nobody outside publishes what an ad cost, who saw it, or whether it worked. Duration is the only performance signal on offer, and it is weak. Treat any confident claim about a competitor’s advertising performance as somebody guessing.

Estimates exist and they are all built the same way: model the traffic, model the cost, and multiply. What a competitor ad spend estimate can and cannot support is worth knowing before anybody quotes a figure from one in a meeting.

Flares keeps a dated record of how competitors describe themselves across their public surfaces. A claim that appears in an ad this month can then be read against when it first showed up elsewhere, which is often a page change nobody announced.