424 objection labels, three shared, and fifteen families underneath them

Search for the most common sales objections and the results disagree about the count and about the order, while none of them says where its own list came from.

The list sits upstream of every number built on it. Change one row and the counts underneath move, though nothing changed in a single conversation. We went looking for the size of that effect in our own data and came out with two findings pointing opposite ways: the wording of a list moves the counts a long way, and underneath the wording there really is a shared set of objections.

One figure in wide circulation says 58% of discovery calls contain at least one objection; it is credited to a conversation intelligence vendor whose own pages do not carry it, and the site that spread it furthest now flags the citation as dead, so we left it out.

The published rankings are vocabularies before they are findings

The most transparent objection dataset in the search results belongs to a contact center software vendor, and it illustrates the problem rather than settling it. MaxContact published counts from an analysis of more than 800,000 objections: No Immediate Need at 151,000, Not Interested at 150,000, Lack of Time at 94,000, Too Expensive at roughly 69,000, and a skepticism category at 50,364. Across all 803,000, agents overturned 39%. Those are real counts with a stated sample size, which is more than almost anything else on the first page of results gives you.

The category names are the giveaway. No Immediate Need, Not Interested, Lack of Time, Contractual Obligations. Those read like the wrap-up menu an agent picks from at the end of an outbound call, and outbound dialing is one of several things this vendor sells alongside inbound voice, quality assurance and workforce management. The ranking is a vocabulary before it is a finding, and the industry reads it as the ranking of a market.

The consequence is narrow. If your team works inbound rather than outbound, Lack of Time will sit far lower in your counts and that tells you nothing about your buyers. If the conversations you review are service calls that turn commercial, Not Interested may not appear at all, because nobody in that queue is cold. Two rankings built on two vocabularies are not comparable, and neither one is wrong. If a published ranking has ever been used to set a target for your team, what needs comparing first is the two lists of rows and not the two percentages.

Whether a buyer objected depends on whether the form has a row for silence

In our slice the share of calls where no objection was recorded runs from 0% to 62.9%. The two teams at the top have a row on their form that a reviewer can pick when the buyer raised nothing at all. The six at the bottom, spread across four other companies, do not.

One caveat before you read anything off that table. Our threshold for publishing from client recordings is five companies and a hundred calls, and this slice clears it exactly at the line: five companies, not six. No company names, no industry tied to a company, no quotes. A slice on the boundary supports a claim about mechanism, not about magnitude.

The mechanism is in the last column, and part of it belongs to us. The six teams at the bottom are not working a market that objects to everything. Their form asks reviewers, and the model that does most of the tagging, which objections the buyer raised, and it gives them a list with no exit, so when nothing fits they take the nearest row. That is our tagger's behavior: somebody on our side decided it should return the closest available label rather than nothing, and that decision produces the bottom row of the table.

We have to name the confound before anyone reads causation off it. Both teams with the row belong to one company and run one kind of conversation, first consultations, so company and conversation type move together with the presence of the row, and these data cannot separate them. The only comparison holding the list constant is A1 against A2, and there two teams on the same form at the same company still sit 8.3 points apart. A row for silence moves the number a long way. It does not pin it down, and two teams of one company sitting apart on one form is the same shape we hit on the next-step question.

Survey methodologists have been arguing about this since the 1960s. Krosnick and twelve co-authors compared forms that offered an explicit no-opinion option against forms that did not and published The Impact of "No Opinion" Response Options on Data Quality in Public Opinion Quarterly in 2002. They concluded that the option does not improve data quality and may preclude measurement of some meaningful opinions, so they advise against offering it. Our advice runs the other way, and the difference is in what the option covers: in a survey, "no opinion" can hide an attitude the person actually holds, while on a tagging item "no objections raised" describes a real state of the call.

Three passes at the same list gave three different counts

We counted this vocabulary three times and got 446, then 440, then 424, each pass lower than the last and none of them adding a single call. That drift is most of what the slice has to say, and it is not a story about carelessness.

The first correction was case. The database does not fold Cyrillic to lower case, so two labels differing only in the capital letter at the front arrived as two labels; correcting case by hand took 446 to 440. The second was separators: one company carries the same label in two spellings, hyphenated and spaced, worth 192 and 53 mentions. Treating hyphens and underscores as spaces took 440 to 424. A production bug we wrote up two days ago rhymes with this and is not the same thing, because that one moved a client's own number over time while these three live only in the analysis behind this article.

Each pass joined labels that were already literally the same words, which is the weakest kind of joining there is. Nothing has been done about two companies writing "price too high" and "cost objection", and nothing can be done about a corpus running in three languages across two alphabets, where labels in different languages cannot match as strings however they are normalized. So 424 is an upper bound on how many distinct meanings are in play, 3 shared labels is a lower bound on the overlap, and both bounds are loose.

A separate mistake sits in a different quantity. The first version of the clean-call count treated a blank tag field and an explicit "no objections were raised" label as one thing, which is correct for the headline share and is still how the final numbers count it. What it failed to do was keep a column saying which of the two it was. Without that column the table of clean-call shares carries no explanation at all and reads as six teams with impossible buyers, because a blank and an explicit "none" are two different states that look identical in a report. We keep a running list of the places a tagging item quietly loses information, and that missing column sits near the top of it.

At the level of the family, a universal list does exist

Group the 424 labels by what they mean and the picture turns over. Ninety-two labels cover 90.9% of all mentions, 3,642 of 4,006, and those ninety-two sort into fifteen semantic families. Ten of the fifteen appear at more than one company. Two appear at all five.

That grouping is a judgment and not a measurement. One person did it by hand, and it is reproducible only in the sense that the families are printed above for anybody to disagree with.

The distinction between a family and a row decides who a published top five is useful to, and the answer is nobody who intends to act on it, because people work with the row. The review card, the dashboard column and the Monday agenda all carry the phrasing one company wrote for itself. "Already with a supplier" transfers between companies; "already has a lawyer" and already_with_agent belong to that same family and share nothing a string comparison can see. A published ranking holds up while it stays at the family level. Carry it down to your own rows and it stops describing your buyers.

Our own number fails the test this article hands out

A benchmark that does not publish the list of labels it was counted on is describing a form, which is the test this article has been applying to other people's numbers. Our 424 fails it. Those labels came off forms configured inside Spellit by the eight teams in this slice, we are not going to publish 424 labels written by clients, and the fifteen families are the only part of this work an outsider can check.

The harder admission is the bottom row of the clean-call table. A reviewer facing a list with no exit takes the nearest row because our tagger is built to return the nearest row, and that is a design decision somebody made rather than a law about lists. It generates the most quotable contrast in this article. We are changing it so a tagging item with no matching label can come back empty, and until that ships, any objection rate we show from a team without a row for silence reads close to 100% for reasons that have nothing to do with their buyers.

None of this disqualifies a vendor number. What disqualifies one is not saying which form it came off, and that is a thing any of us can fix and most of us have not.

The loss that costs the most carries no label on any list

Objection lists can only count what somebody said out loud. Dixon and McKenna's study of 2.5 million recorded sales conversations, published as The JOLT Effect in 2022, finds that between 40% and 60% of deals are lost to customer indecision rather than to a competitor. Nobody raises indecision as an objection. There is no sentence a buyer says that a reviewer would tag with it.

The most expensive category of loss in B2B selling is therefore absent from every objection ranking ever published, ours included. A top five is a list of the five things buyers are willing to say, which is a different list from the five things costing you deals. Our own data cannot close that gap either: the slice carries no reliable field for the final outcome of a deal, so we cannot put labels next to wins and losses at all.

What objection counts do tell you is what your buyers are willing to name, which is what you want when preparing a new rep for the calls they will actually get. An objection report answers "what did they say", and nobody should ask it "why did we lose", though in most companies both questions come up in the same meeting.

What to do next

Open the list your team picks objections from. It may be a dropdown in a CRM, a column in a spreadsheet, a page in a quality handbook, or the prompt somebody wrote for the model that tags your calls. Count the rows. Then check one thing: is there a row for a call where the buyer raised nothing at all? If there is not, the share of calls your reports show as carrying an objection is a fact about that list, and it will drop the day somebody adds the row.

Then take the twenty most recent calls tagged with your most common label and read the tagged moment in each transcript. The question is whether the buyer said the thing or whether the nearest available row got picked. In our slice the six teams without a row for silence came out with an objection on 97.5% of calls or more, and it looked entirely plausible in a report.

Last, do the grouping yourself. Write the rows out, sort them into families by meaning, and count the families. If fifteen rows collapse into four families, the extra rows are splitting one objection across several counts, and every number built on them is thinner than it looks. Which of our fifteen families would your own rows refuse to fit into?