The Prompt to Analyse Competitor Mentions in Sales Calls

A search index answers a question about four thousand calls. A model answers a question about the twenty you put in front of it, and the two outputs are written in the same voice. Nothing in a summary says which one you are reading.

That matters because of how the twenty were chosen. Almost always by searching the archive for a competitor’s name, which means the set arrives pre-selected for the conclusion. An output that says a competitor is raised more than any other, about calls gathered by searching for that competitor, is a sentence with no information in it, and it reads exactly like a finding.

The fix is not a cleverer instruction. It is putting the query in the brief, so the model can qualify its own answer: twenty calls, from a search on these terms, over this date range, out of this many the search returned. One sentence, and the output stops claiming something it was never in a position to know.

Three prompts, and the third one comes first in practice, which is why they sit above the input contract rather than below it. Building the spelling list is a five-minute job, and it is the only place in this library where one prompt manufactures an input another one needs.

Notice what the second prompt is forbidden from doing. It may not judge whether a claim is accurate, and it may not draft a response. Both are useful and both belong later, with a person and a source in the room, because a model asked to sort evidence and rebut it in the same pass will quietly start sorting the evidence in a way that supports the rebuttal.

Three inputs, and only the first is a document. The other two are what make the first one interpretable, and both are routinely left in somebody’s head: the spelling list the prompt above produces, and the query that decided which calls you are holding.

Paste whole calls rather than the lines containing the competitor. What was being discussed in the two minutes before a name is raised is most of what the mention means: a competitor named while a buyer talks about budget and a competitor named while they talk about a missing capability are different events, and an extract of the naming line alone has thrown that away before the model sees it.

The buyer sentence below is constructed rather than lifted from a real recording, and it is the shape a transcript search returns every day. What is measured here is what the model did with it, and the packaging it gets checked against is a live page anyone can open.

The correction is not the useful part

It would be easy to read the second output as fact-checking the buyer, and that is the least valuable thing in it. The buyer’s belief is a fact about the deal regardless of whether the page agrees: somebody compared your price against a package they think exists, and that comparison is what priced you, so correcting them changes nothing about how the deal went.

What changes is the next conversation. A rep holding “they include it in Starter” will tell the buyer they are wrong, which is an argument nobody wins in front of a prospect. A rep holding “the buyer believes it is in the entry tier, the page lists it two tiers up” asks what they were shown, and the answer to that is frequently a bundle, a discount or a version of the page that has since changed.

Two sentences, same competitor, same six words of substance. A seller says “unlike them, we handle this natively”. A buyer says “they handle this natively”. One is your talk track and one is a rival’s argument reaching a buyer, and an export that has flattened the speaker labels puts both in the same count.

The direction of the error is always the same, which is what makes it worth naming. Sellers raise competitors more often than buyers do, because raising one first is a standard technique. So an unlabelled archive systematically overstates how much buyers are thinking about your competitors, and the number that gets quoted in a quarterly review is closer to a measure of your own positioning than of theirs.

Count accounts, and say so in the output

The other correction is arithmetic. One buyer returning to a concern four times across a negotiation is one buyer; four unrelated buyers raising it once each is a pattern. Both produce the number four, and only the second is a reason to change anything. Requiring the prompt to report distinct accounts, and to say that is what it counted, is the difference between a chart somebody argues with productively and one they dismiss.

Where this lands is a rep’s preparation rather than a report. A claim recorded verbatim, with the deal and the date beside it, belongs on the sheet a rep reads before a call, and the same claims collected across a quarter are what a win/loss synthesis is trying to explain after the fact. Calls catch it while the deal is still open, which is the only real advantage this source has over every other one.

Everything on this page depends on a set of transcripts somebody assembled, and that assembly happens on whatever day somebody remembered to do it. A competitor appearing in a segment for the first time is the event worth knowing about, and it is invisible to a quarterly sweep because a quarterly sweep only reports what is already common.

Most call platforms will alert on a term, which is the cheap half of the answer and worth setting up for every competitor you can name. The half it does not solve is the competitor you cannot name yet, arriving under a spelling nobody has added, in a segment nobody is watching. An alert list is a list of the rivals you already know about.

The other half of the problem is on the outside. A buyer’s account of a competitor is only checkable against what that competitor currently publishes, and by the time somebody opens the pricing page to check, the page may have moved. A claim and a dated page belong together or neither is much use.

Flares keeps competitor pricing, packaging and messaging current and dated, so the claim a buyer repeats on a call can be read against what that competitor actually published that week. Hearing the claim in the first place still depends on your own recordings, and on somebody searching them for the right spelling.