We found our competitors' AI prompt strategies hiding in Search Console

I was in Search Console for something ordinary. A page had lost traffic and I wanted to see which search terms went with it. Sorting by impressions, I found this sitting in the list like any other search:

Nobody types that into Google. It names three of our competitors, specifies a review site as the evidence, and issues an instruction about the shape of the answer. It is not a search. It is a test, written by somebody who had already decided what they wanted to find out.

One odd row proves nothing on its own. Search Console has become a noisy place: Google answers many searches with generated results now, and can split a single request into several, so a row is not always a person. What settles it is the behaviour around the row. Exact instructions repeated across hundreds of variants. Zero clicks at scale. Near-total desktop use. And a schedule that began on a specific date and then did not miss a day.

A useful tracking prompt needs more than a topic. Somebody has to choose the buyer, market, product job, competitor set, evidence source and form of the answer. Those choices are the research design.

In the row below, all of them are visible. It simulates a senior marketing buyer at a mid-sized construction company in North America. It tests three named products on AI visibility and executive reporting, asks G2 for evidence, then requires a decisive best-to-worst answer.

Seven repeated prompts compared Trakkr with Profound and Otterly across voice search, keyword tracking, local SEO, global SEO and multilingual features. They produced 721 impressions and no clicks. Several of those categories do not describe what Trakkr does at all.

We could leave that as a joke at somebody else’s expense, except the same mistake is available to everyone, including us. A measurement system will happily produce a stable, confident, weekly score against the wrong category, the wrong rivals or a buyer who does not exist. Nothing in the number tells you that. You only find it by reading the prompts.

So before you act on any AI visibility data, audit the prompt bank first. The jobs, personas, regions, competitors and sources are not settings around the model. They are the model.

We compared the structures in Search Console with first-party documentation and case studies. Three matches stand out, and each one identifies a product design, not an operator: an exact match to public documentation cannot prove which customer, agency or internal team submitted a query.

The repeated “what do users say ... on G2?” structure fits the same picture. G2 and Profound have a public partnership through which G2 ingests Profound visibility data and top prompts, which makes a review-site evidence constraint commercially coherent.

A visibility programme defines the market before it measures it. The chart below counts brand mentions inside the forced-ranking family: not rankings or market share, but a map of which products somebody believes belong in the same answer.

The evaluation criteria recur just as clearly. Executive reporting, citation analysis, brand mentions, sentiment, source tracking and enterprise fit run across the prompt families: a rough map of the capabilities buyers are being taught to compare.

Translation adds another layer. The same structure ran in five languages, and the German version logged 90.2% of its impressions in the United States and about 1% in Germany. Query language and execution market are separate settings.

When the prompts ask for a comparison, Google reaches for reviews. The result needs precise wording: Search Console records the page Google associated with the query, not a page anybody visited, opened or read. It tells us what Google considered relevant to these comparison-shaped searches, not what happened inside a model’s retrieval process.

For any brand, the practical question is whether Google surfaces the pages that best explain its category, evidence and differences. A clear review, comparison or alternatives page can answer that need. It should be accurate and useful in its own right, not written merely to target automation.

Google’s dedicated generative AI performance report separates AI-feature clicks and impressions by page, country, device and date, but it does not expose the query dimension. The ordinary Performance report is still where these prompt clues appear.

Start with narrow syntax filters. Then check whether the rows repeat on a schedule, cluster on desktop, appear in unexpected countries and collect impressions without clicks. Inspect the page dimension as visibility, not as a visit. Finally, read the prompts as a research brief: what do they assume about the buyer, market, competitors, sources and desired answer?