The Taphonomy of Customer Feedback: Which Complaints Survive to Reach You (2026)

Most customer feedback is destroyed before it ever reaches you, and the destruction is not random: it removes specific, predictable classes of evidence. Paleontologists have a name for the study of that destruction. It is called taphonomy, and it is the most useful idea product teams have never borrowed. Once you know which kinds of evidence your collection channels systematically destroy, you can stop treating your feedback pile as a sample of reality and start collecting the missing classes on purpose.

The practical payoff: your support tickets, NPS comments, and sales-call notes are not a small, faithful sample of what customers experience. They are a biased cast of the hardest, most durable parts of it. The soft parts, which are usually the majority, leave no trace at all unless you use a capture method designed to preserve them.

What taphonomy is, and why it is not just another bias

The term was introduced to paleontology in 1940 by the Soviet scientist Ivan Efremov, and it names the study of how organisms decay and become fossilized or preserved in the paleontological record. Taphonomic phenomena are conventionally grouped into two phases: biostratinomy, the events that occur between the death of the organism and its burial, and diagenesis, the events that occur after burial.

That two-phase split is the part worth stealing. It separates losses that happen before the evidence enters your record from losses that happen after it is already in there. Most teams only ever audit the second phase.

The governing principle is broader than any single mechanism: in taphonomy, any factor that affects the likelihood that an organism is preserved as a fossil is a potential source of bias. Note the structure of that claim. It is not that the record is small. It is that the filter has a shape, and the shape is knowable.

This is what separates taphonomic thinking from the more familiar framing in survivorship bias in customer research. Survivorship bias says: you are only hearing from the customers who stayed. True, and important. Taphonomy says something stronger and more actionable: the destruction process selectively removes particular classes of evidence, so you can predict exactly which insight types are absent and go collect those specific types directly. Survivorship bias tells you the sample is wrong. Taphonomy tells you which shelf to go and look on.

It is also distinct from the problem described in why your repository cannot show you what it lost. That is a diagenetic problem: material that made it into the record and then degraded or vanished without a log entry. Taphonomy asks the prior question, the biostratinomic one: what never made it into the record at all, and why not?

Hard parts and soft parts

The central taphonomic fact is that preservation is class-biased. Organisms that contain hard parts have a far greater chance of being represented in the fossil record than organisms consisting of soft tissue only.

The magnitude of that bias is not a rounding error. Soft-bodied organisms may form 30% to 100% of the biota, but most fossil assemblages preserve none of this unseen diversity. The missing class is routinely the majority.

The sharpest single measurement comes from the Burgess Shale in British Columbia, a 508-million-year-old deposit famous for the exceptional preservation of the soft parts of its fossils. Only about 14% of genera in the Burgess Shale had possessed biomineralized tissues in life. Read that the other way round: in an ordinary deposit, roughly 86% of that fauna would simply not exist as far as the record is concerned. Not underrepresented. Absent.

Your feedback record has hard parts and soft parts too, and the ratio is not kinder.

Every row on the left has a container built for it: a ticket field, a form, a CRM note. Every row on the right has none. The asymmetry is structural, not cultural, and no amount of encouraging customers to "share more feedback" changes which container exists.

The first phase: what dies before burial

Biostratinomy, for feedback, is everything that happens between a customer having an experience and that experience entering a system you can query.

The attrition at this stage is severe and it has been measured. The 2006 Retail Customer Dissatisfaction Study, run by the Jay H. Baker Retailing Initiative at Wharton with the Verde Group, found that only 6% of shoppers who experienced a problem contacted the company about it, while 31% went on to tell friends, family, or colleagues what happened. The experience was not forgotten. It was preserved, at volume, in a deposit you do not own.

Nothing about that 6% is a random draw from the 100%. Contacting a company selects hard on effort tolerance, on how articulate the customer is, on whether they believe complaining works, and on whether the problem was legible enough to describe in a ticket form. Those are precisely the traits that correlate with the kind of problem being reported. Sharp, nameable, single-cause failures survive. Diffuse, cumulative, hard-to-articulate friction does not.

The second phase: what degrades after burial

Diagenesis is the set of changes after the evidence is already in your record, and it is where paraphrase does its damage.

A customer says something specific and slightly awkward. A CSM writes a cleaner version of it in the CRM. A summary of that note reaches a planning document. A slide quotes the summary. At each step the utterance gets more quotable and less informative, and every step is invisible in the final artifact, which looks like a clean fact. This is why a relayed customer claim is not customer evidence, and why the verbatim matters more than the tidy version of it.

How paleontologists correct for it, and how you can

Paleontology does not solve this problem by wishing the record were better. It solves it two ways, and both translate.

The first is to study the filter itself. If you know the mechanism that destroys soft tissue, you know what is missing, and you can reason about the gap instead of ignoring it. Practically: write down which of your insight classes has no container, and stop treating its absence from dashboards as evidence of absence in the world.

The second is to find, or create, a conservation deposit. These are the deposits known for the exceptional preservation of fossilized organisms or traces, and their value is precise: they preserve lightly sclerotized and soft-bodied organisms or traces of organisms that are not otherwise preserved in the usual shelly and bony fossil record, and so they offer more complete records of ancient biodiversity and behavior. The Burgess Shale is one. The Chengjiang deposits in Yunnan, at 518 million years, are another.

You cannot wait 500 million years for favorable conditions. You can, however, build the conditions. A conservation deposit for customer evidence is any capture method that records the soft parts: the hesitation, the workaround, the abandoned sentence, the tradeoff reasoned aloud. A conversation does that. A form does not. This is the entire reason a moderated interview outperforms a survey on the same question, and it is the reason an unmoderated survey at ten times the sample size does not close the gap: it is a concentration deposit, dense in hard parts, empty of everything else.

How Koji handles this

Koji is built to be the conservation deposit in your stack rather than another form that collects hard parts.

• AI-moderated conversations, not forms. Koji's AI interviewer asks a real follow-up when an answer is thin, vague, or interesting, which is the only mechanism that gets the unprompted reasoning on the record at all. A static survey cannot probe; it captures the answer the respondent was willing to type first.
• Voice and text in the same study. Voice interviews preserve hesitation, self-correction, and the qualifier before the claim. Koji runs voice and text side by side, so you are not trading modality for reach.
• Verbatim transcripts kept under the analysis. Every theme in a Koji report stays linked to the transcript segment it came from, so the diagenetic chain from utterance to slide is inspectable instead of assumed.
• In-vivo theme tags. Koji's analysis can tag a theme in the participant's own framing rather than forcing it into a house label, which is how a soft, awkwardly-phrased signal survives coding instead of being rounded to the nearest known category.
• Structured questions where hard parts are genuinely what you want. Koji supports six question types: open_ended, scale, single_choice, multiple_choice, ranking, and yes_no. The discipline is to use scale and choice questions for the durable, countable things and open_ended with AI follow-up for the soft tissue, in one session, rather than running two disconnected studies. See structured questions in AI interviews.
• A quality gate on the deposit. Koji scores each conversation from 1 to 5 on relevance, depth, and coverage, and only conversations scoring 3 or above consume a credit. A thin interview is a poorly preserved specimen, and you are not billed for it.

The practical workflow: run the countable questions and the conversational probe in the same Koji study, then treat the resulting transcripts as your primary deposit and the ticket queue as what it actually is, a bone bed of hard parts, useful for counting and useless for anatomy.

Frequently asked questions

Is taphonomy just a fancy word for survivorship bias?

No, and the difference is operational. Survivorship bias says your sample over-represents survivors. Taphonomy says the filter destroys identifiable classes of evidence in a predictable way, which means you can name the missing class and go collect it deliberately. Survivorship bias diagnoses; taphonomy prescribes. The two also sit at different stages, and it is worth reading survivorship bias in customer research alongside this.

What counts as a soft part in customer feedback?

Anything with no container built for it: an invented workaround the user no longer notices, a tradeoff resolved silently, the job a feature request was a guess at, hesitation before a confident claim, and the reasoning that connects a complaint to its cause. If your systems have no field for it, it is soft tissue, and it is being destroyed by default.

Does collecting more feedback fix the bias?

No. More volume through the same channel deepens a concentration deposit: you get more hard parts, more precisely counted, with the same classes missing. Fixing a class-biased filter requires a different capture method, not a larger sample through the existing one. This is why running a bigger survey rarely resolves a question a handful of conversations would answer.

How do I know which classes are missing if they leave no trace?

Study the filter rather than the record. List your collection channels, and for each one write down what it structurally cannot capture: a dropdown cannot capture a reason not on the list, a ticket form cannot capture a problem the user does not consider reportable, an NPS comment box cannot capture a tradeoff. The union of those blind spots is your missing-class list, and it is stable enough to plan against.

Where does the 6% figure come from?

From the 2006 Retail Customer Dissatisfaction Study by the Jay H. Baker Retailing Initiative at Wharton and the Verde Group, which found that only 6% of shoppers who had a problem contacted the company, while 31% told friends, family, or colleagues. It is a retail study, so treat the exact number as illustrative rather than as your own conversion rate, and treat the direction as robust.

How does Koji change the preservation ratio in practice?

By replacing the form with a conversation at the point of capture. Koji's AI interviewer probes thin answers, runs in voice or text, keeps the verbatim transcript under every theme, and can tag themes in the participant's own words. That combination preserves the classes of evidence that a survey destroys, which is the only intervention that changes the ratio rather than the volume.

Related Resources

• Survivorship Bias in Customer Research - the sampling half of the problem, and the right companion to this piece
• Structured Questions in AI Interviews - when to use each of the six question types, and when to let the AI probe instead
• Nobody Logged the Deletion - the diagenetic failure mode, after evidence is already in your record
• Why the Loudest Complaint Hides the Real One - what happens when a preserved signal drowns out a fainter one
• Hearsay in Product Research - how paraphrase degrades a quote on its way to a slide
• In-Product Research Recruiting - reaching the customers who would never open a ticket