Competitive intelligence from sales calls, sorted three ways

Search the phrase and the first page gives you battlecard templates. The sections are consistent: where the rival wins, where it loses, trap questions to plant, proof points, a comparison table. The research instruction underneath is equally consistent, and it points outward. Read their site. Open a trial. Pull their review profiles. Ask customer success why people leave. Every one of those sources describes the competitor, and none of them records what your buyer was doing in the three seconds before the name arrived.

The card gets reached for constantly, because the field is small. The 2026 B2B Buying Disconnect report, published 15 July 2026 on a survey of 1,862 technology buyers and 444 vendors fielded in January 2026, found that 83% of buyers shortlisted three products or fewer. A list that short means the other name is almost always in the room. The question is not whether reps will need the card. It is whether the card knows which of three conversations it just walked into.

A card built from public material answers a question nobody asked

A battlecard assembled from a rival's website, reviews and collateral describes the rival accurately and is still the wrong document for the moment it gets opened. What a rep needs is a response to one sentence a specific buyer just said. The sentence is not on the card, because the card was written before anyone heard it.

One of the better-known guides in this genre argues against feature comparison tables, on the grounds that a matrix invites the buyer to treat the options as interchangeable and pushes the conversation toward cost. That is a real improvement over the standard template, and it leaves the deeper problem untouched: the card is still organized around the rival rather than around the moment. Reorganizing the rows does not help if every row answers the same implied question.

The implied question is always the same one. Why should I pick you over them. It is a perfectly good question and buyers do ask it. They also say a dozen other things with the same proper noun in them, and roughly half of those are not asking anything at all. Whether your own conversations mostly ask that question or mostly do something else is an hour of sorting.

One name, three situations, and they do not share an answer

The same name arrives in at least five distinguishable shapes, and the response differs for each. Three of them are the common ones: status, question, objection. The split matters because two of the three are not sales moments. They are information the buyer handed you about their situation, and answering them with a differentiation talk track spends the conversation on the wrong thing.

Rows one and five are the expensive confusions. A buyer reporting an existing arrangement has told you about a contract, a switching cost and an internal owner, and none of that is on the card. A buyer who has already eliminated the rival has told you that the card is now about the wrong company, and running it anyway means spending five minutes explaining why something the buyer already rejected is worse than it looks.

What this costs is hard to see afterward, because the call carries on normally. A conversation gets one competitor exchange, maybe two, and one of them has now gone to a question nobody asked. The buyer also registers that the name triggered something prepared, and the rest of the call gets phrased more carefully on their side than it would have been.

A row on the card is only worth arguing about if somebody can check it. In Spellit ClientTruth every conclusion ends at a verbatim buyer line with a pointer to the place in the recording it came from. A disputed row then gets settled by opening it, which takes half a minute, rather than by two people recalling different meetings.

Telling the three apart is a labeling problem with a measured difficulty

Deciding which act a sentence performs is a known research task with published numbers, and the numbers say it is not trivial. Classifying an utterance by function rather than by topic, statement against question against agreement, is what dialogue act tagging has worked on for decades, and the headline accuracies are sobering for anyone who assumed a rep could do it reliably while also selling.

Stolcke and colleagues trained and evaluated on 1,155 hand-labeled telephone conversations from the Switchboard corpus and reported the results in Computational Linguistics in 2000. Automatic labeling reached 71% accuracy from clean word transcripts and 65% from recognized speech, against a chance baseline of 35%. Human coders reached 84%. That last figure is the one worth sitting with: careful people, working from recordings of ordinary telephone talk, with a written manual in front of them, agreed with the reference about five times in six.

Two things follow for a battlecard. The first is that the act is not recoverable from the topic, which is exactly what the gap between 35% and 84% represents. Knowing a call contained a rival's name is close to the chance baseline of useful. The second is that a call-level tag destroys the distinction before anybody reads it. One recording routinely carries all three moves, sometimes in the same minute, and a checkbox marked competitor mentioned flattens them into a single row.

A default worth starting from: label the sentence, not the conversation. The unit is the utterance, and a call with four mentions contributes four labeled items. It is the rule that lets an objection taxonomy survive contact with real transcripts, and it is why a buyer question is worth sorting by what happened to it rather than by its subject.

We shipped the phrase competitor mentions before anyone wrote down what a mention was

Our own landing copy for the customer-voice product listed competitor mentions next to pains, objections and feature requests, as though all four were the same kind of thing with the same machinery behind them. Three of them had screens. The fourth did not exist. In the product specification it sits as an open point with a one-line note beside it saying there is no data for this.

The sentence got written because it reads like a capability. Nobody had to decide what counted, so nobody noticed that no one had. Does a buyer naming their incumbent count? A buyer repeating what a colleague heard at a conference? A rep who brings the name up first and the buyer who agrees without elaborating? Each of those answers produces a different number from the same month of recordings, and until one of them is written down there is nothing to count at all. The line survived two documents because a phrase that sounds like a feature does not look like an open question. Nobody has to define a mention until somebody counts them, and a period of your own recordings forces the definition.

The repair is small and it is the whole job. Before a row goes on the card, write the sentence that would qualify for it and the sentence that would not. If the second one cannot be written, the row is a phrase rather than a finding, and it will be argued about every quarter by people who each privately mean something different by it. Deciding how a pain gets worded before anybody counts it runs into exactly this, and skipping the definition there costs a quarter of arguing too.

Counting conversations and counting mentions give you two different cards

One month of recordings supports two rankings, both correct, that put different rows on top. One counts conversations where the thing came up. The other counts how often it came up. Neither is a mistake and they are not interchangeable, and a card built on the wrong one points a team at the wrong row.

We ran into this on our own objections screen, and it is worth saying plainly that nothing was broken. The weekly column chart counts calls: a conversation gets its first objection label and nothing else, so a call carrying three distinct objections appears exactly once, filed under whichever one landed first. The catalog sitting next to it carries two shares for every label, one over conversations and one over mentions. Two blocks, one page, and they rank the same list differently on purpose. Read side by side for the first time they look like an arithmetic error. They are not. Nobody had said which question the ranking was supposed to answer.

In Spellit ClientTruth a share is reported against the base the question belongs to, and the base is printed next to the number rather than assumed. That sounds like formatting until the first time two people quote the same screen at each other with different numbers. The battlecard version of this rule is one line at the top of the card saying what the counts underneath are counts of. A card without that line gets rebuilt from scratch every six months, because nobody can reconstruct what the old one measured.

There is a second-order effect worth naming. A card ranked by mentions over-represents one talkative deal, and a card ranked by conversations under-represents the rival that gets raised four times in every call it appears in. Printing both numbers removes both distortions. Arguing about which unit is the real one removes neither.

What to do next

Open the last twenty recordings where a rival's name was said out loud by the buyer. In a firm where the meeting itself is the deliverable, do this on the engagement calls rather than the sales ones, because that is where the comparison actually gets made. Do not read them for content. For each separate mention write one word in a column: status, question or objection. Twenty conversations take under an hour and usually produce more mentions than expected, because a single call carries several.

Then count the three groups and work on the largest one only. If status dominates, the response is a short list of questions about the existing arrangement, its dates and its internal owner, and it belongs in the discovery sheet rather than on the card. If questions dominate, the response is one repeatable sentence of difference, and it also belongs on a page of your site, because a comparison question asked on a call is usually one the site failed to settle earlier. If objections dominate, the response is the named cost the comparison leaves out.

Whichever group wins, write the qualifying and disqualifying sentence for every row you add, and print the base next to every share. A card with those two things on it survives the next person who inherits it. A card without them gets rewritten. The analysis that supports all of this is the same material that drives the actions a team takes from scored conversations. We can sort a period of your own calls into the three and tell you which group turned out largest.