Sales-call competitive intelligence

Competitive leads and product marketers want one thing from recorded sales conversations: evidence about who else the buyer is talking to and what stops deals. The output of this job is a list of moments. Each row names the rival, the call, the timestamp, who spoke, and the quote. Nothing more. A row without a quote and a locator is an opinion with a deal number attached.

The battlecard is a separate artifact with a separate owner. A quote from one call is not a competitor's price, product limit, or roadmap. It is what one buyer believed on one Tuesday. Treat every extracted claim as a report of a claim.

Gong and Klue sit in different parts of this. One class of tool holds the recordings, transcripts, and the deal join. The other holds the competitive program and the cards reps read. Neither one turns a transcript line into a verified fact about a rival. Public-side change tracking belongs with continuous diffing of competitor sites and pricing, which watches what the rival publishes rather than what a buyer repeats.

Name every rival with the strings that actually appear in speech. Legal name, product names, nicknames reps use, common mistranscriptions, and the acronym. Speech-to-text mangles brand names constantly, and a list built only on the marketing spelling misses every mention the transcript spelled its own way.

Then write the trap list. Rival names that are ordinary words, cities, or common surnames will fire on unrelated sentences. If a competitor is called Meridian, "meridian" in a discussion of time zones is not a mention. Keep those strings on a review path instead of dropping or auto-accepting them.

The objection taxonomy needs the same care. Define a small number of categories a person can apply consistently: price, missing capability, integration risk, security review, incumbent switching cost, timing and budget. Loose categories mean the model sorts by tone, and "hesitation" becomes a bucket that never changes anything.

Two more inputs are non-optional. Recording consent and scope, so calls that should not be mined are excluded before ingestion. And the CRM outcome join, so a call can be tied to a real closed-won or closed-lost record.

When the alias list is missing or the taxonomy is undefined, run nothing. A quarter of calls scanned against a half-written rival list produces a mention count that people will quote for months.

Three classifications sit on every mention, and collapsing them is the most common way this becomes misleading.

Who raised it. The rep saying "unlike Meridian, we handle multi-site" is not competitive pressure. The buyer saying "we already demoed Meridian last month" is. Count them in separate columns and never in one total. Programs that skip this step report a rival as surging when the surge was a new talk track.

What was actually compared. A rival named as the incumbent is a different situation from a rival in an active evaluation, and both differ from a rival mentioned as a reference point the buyer once read about. Classify the context from the surrounding turns, and leave it unclassified when the transcript does not support a choice.

Whether the deal resolved. Attach an outcome only when the CRM record is closed and the close date is real. Open deals belong in the mention list with the outcome field blank. Win and loss reasons typed by a rep in a rush are weak, and they are the field most likely to be back-filled to match whatever story the team already tells.

Objections follow the same rule. Extract the instance, quote it, tag one category, and leave it as an instance. An objection instance is not an objection rate, and a handful of them is not a pattern.

This walkthrough is hypothetical, with invented company names and numbers. A company selling route-planning software to regional food distributors has 61 recorded first and second calls in a quarter. Two rivals matter: Meridian, an established suite, and Ostrom, a cheaper new entrant.

The scan returns 34 mention rows across 22 calls. Split by speaker, 19 rows came from reps and 15 from buyers. The rep rows cluster around one sentence introduced in a January enablement session, which explains most of the rise in raw mention counts.

Of the 15 buyer mentions, Meridian appears in 11. Nine of those describe it as the incumbent system the operations team already runs. Two describe an active evaluation. Ostrom appears in 4, all in deals under a certain size, and in three of them the buyer names price before naming the product.

Objection extraction on the same calls returns 47 instances. Integration risk is the largest category, and 12 of those instances mention the same warehouse management system by name. That is the specific finding worth carrying forward, and it is checkable: twelve quotes, twelve timestamps, twelve accounts.

Deal outcomes are attached to 23 of the 61 calls, because the rest are still open. Five of the closed-lost records mention Meridian. Five is not a win rate against Meridian, and the report should say so on the row.

What the competitive lead does with this: pulls the twelve integration quotes, checks whether the connector actually works, and takes the answer to product. What nobody does: publish "Meridian is winning" from five closed deals.

The mention list is easy to over-read, and the errors travel fast because they sound specific.

Rep-planted mentions inflate everything. If enablement tells reps to name a rival, mention volume climbs and nothing about the market changed. Always report buyer-raised counts on their own line.

Small samples turn into rates. Five losses, three of which mention a competitor, is not a competitive loss pattern. Show the denominator on every claim. If it is under a dozen deals, describe the finding as instances.

Stale CRM fields corrupt the join. Loss reason picklists are filled in a hurry, sometimes weeks later, sometimes to match a forecast conversation. When the field disagrees with the transcript, trust neither and mark the row as conflicting.

Buyer claims become facts about the rival. A prospect saying the other vendor includes an audit log for free is a claim to verify against the rival's own published material, not a line for a card. This is where the mining job stops and where research briefs built from cited public sources start.

The last failure is scope creep into coaching. This system reads what buyers said about rivals. Using the same extraction to score whether a rep handled the objection well changes the incentive on the recording, and reps who feel graded stop having candid conversations.

The competitive owner is the gate. A mention row becomes a card line only after a person opens the call, hears the sentence in context, and checks the underlying claim against something outside the transcript. Two quotes and a rival's own documentation beat forty extracted rows nobody listened to.

Blanks stay blank. If a competitor's behavior in a category was never discussed on a call, the card does not get a sentence about it. Filling gaps from general knowledge about that vendor is how a card acquires a claim the company cannot support in front of a customer.

Keep the direction of evidence straight. Mining shows what buyers said in a sales conversation, where the buyer had reasons to shape the story. What customers say unprompted in public sits in reviews and forum threads clustered into themes, and those two sources disagree often enough to be worth comparing.

Counter-messaging is its own decision with its own review. Before a new line reaches live calls, pre-test it the way LLM personas checking positioning against ICP segments describes, and route the final wording through the same claim approval path as any other copy.

Publish the sample size with the finding, every time. The honest version of this program says what it heard, from how many buyers, in how many deals, and what remains unverified. That version survives the quarter when a rival changes its pricing page and half the mention list goes out of date.