Which Prompts Actually Work for Competitor Research?

Every prompt in this guide is built from the same four parts, and a prompt missing any of them tends to produce something unverifiable.

1. The material, in the message. Reviews, post captions, ad copy, a pricing page pasted as text. A prompt that names a competitor and asks what they are doing is asking the model to recall, which is where invented specifics come from.
2. A named output shape. Themes with counts, a table with fixed columns, a ranked list of five. An unnamed shape produces an essay, and an essay cannot be traced back to the input.
3. Counts, wherever a count is possible. A theme with 14 reviews behind it is checkable in a minute. A theme described as common is not checkable at all.
4. A residue instruction. Ask it to list separately anything it could not categorise or support from the material. That list is where the risk went, and it is much shorter than the document.

Paste the review text first, including the star rating and the date on each one. Then use one of these.

• Complaint themes with counts. Prompt: Here are reviews of one company, each with a star rating and a date. Group the complaints into themes, give the number of reviews in each theme, and quote one verbatim line per theme. List separately any review you could not categorise. Fails at: telling you which theme matters commercially, because it does not know what your product does well.
• Switching reasons. Prompt: From these reviews, extract every passage where the reviewer describes what they used before or what they moved to, and group the passages by the reason given for switching. Quote each passage. Fails at: volume estimates, since only a self-selected minority of reviewers mention switching at all.
• Promise against experience. Prompt: Column one is this company's marketing copy. Column two is their reviews. List every claim in column one that column two contradicts, quoting the review line for each. Fails at: claims nobody reviewed, which is most of them, so absence of a contradiction proves nothing.

Paste captions with their dates, and the engagement figures if you have them.

• Format and angle table. Prompt: Each line below is a post with a date. Produce a table with one row per month, columns for format counts, and a second table listing the three most repeated arguments with the count of posts making each. Fails at: identifying paid amplification, because nothing in a caption marks a boost.
• What changed between two periods. Prompt: Group one is posts from January to March. Group two is April to June. List every theme present in one group and absent from the other, and every theme whose share of posts moved. Fails at: explaining why, which needs a source outside the posts.
• Outlier explanation. Prompt: Here are posts with their like counts. List the five furthest above the median for this account, and for each one name the specific thing in the copy or format that the other posts lack. Fails at: proving causation, so treat every answer as a hypothesis to test on your own account.

Copy the visible text of the page into the prompt. A URL alone leaves the model either guessing or fetching a page whose JavaScript-rendered pricing table it cannot see.

• Beliefs behind the claims. Prompt: Below is a competitor's homepage as text. For each headline claim, write the buyer objection it is designed to answer, in the buyer voice. Then list the objections no claim on the page addresses. Fails at: knowing whether the objections are real, which comes from your own sales calls.
• Pricing structure comparison. Prompt: Here are two pricing pages as text. Ignore the absolute prices. Compare how each one meters value, which usage patterns each one penalises, and where each one puts its upgrade pressure. Fails at: current accuracy, so re-check every number against the live page before quoting it.
• Version diff. Prompt: Version A was captured on the first date, version B on the second. List every change line by line, mark each as an addition, a removal or a rewrite, and say which buyer belief the change targets. Fails at: changes below the fold or inside a collapsed section that your capture missed.

Paste the ad copy with first-seen dates from the ad library. Screenshots work too if your assistant reads images.

• Offer and promise extraction. Prompt: Each block below is one ad with a first-seen date. For each, state the offer, the promise, the named audience and the call to action in four short lines. Then list every ad where any of the four is absent. Fails at: performance, since a long-running ad is not necessarily a winning one.
• Creative pattern across a set. Prompt: From these ads, list the message variants being tested, group the ads under each variant with counts, and name the variant that has run longest. Fails at: budget split, because the number of creatives says nothing about the money behind each one.

These five requests are unfixable by prompting, and the failure is quiet in every case.

• What does this competitor charge? Training data holds older prices and a live fetch may miss a pricing table the browser assembles after load.
• Which ads are they running now? Ad libraries are not in training data, and the live fetchers documented by model vendors retrieve a page on request rather than sweeping a library.
• What changed since last week? Nothing stored last week, so there is no earlier version to compare against.
• What is their traffic, spend or conversion rate? None of it is public, so any figure is modelled at best and fabricated at worst.
• Who are my competitors? A plausible list arrives instantly, and it will mix real rivals with companies that merely sound adjacent.