The Assemblage and the Type Fossil: Which Quotes Actually Carry Information (2026)

Your most frequently mentioned feedback is usually your least informative, and the reason is not psychological but logical: a signal that appears in every period cannot tell you which period you are in. Geologists have used this principle for two centuries to date rock. They call the rare, high-information fossils index fossils, and they ignore the abundant ones. Product teams do the reverse, ranking themes by mention count, and then wonder why the top of the list never changes.

The short version: frequency and information are different quantities, and for most decisions they are inversely related. A theme's value comes from how narrowly it occurs, not how often.

The index fossil, and the four criteria

An index fossil, also called a type or guide fossil, is the fossilized remains or traces of particular plants or animals that are characteristic of a particular span of geologic time or environment. Find one in a layer and you know, within limits, when that layer was laid down.

Not every fossil can do this work, and biostratigraphy is explicit about the properties required. A good index fossil has a limited vertical time range, wide geographic distribution, and rapid evolutionary trends. It must be short-lived as a species, so that the period of time during which it could be incorporated in the sediment is relatively narrow. And it should be easy to preserve and easy to identify.

Read those criteria as a filter and something striking falls out. The most useful fossils are the ones that existed briefly. Longevity is a disqualification. A creature that thrived unchanged for 200 million years is abundant, easy to find, and worthless for dating, because its presence is consistent with almost any answer.

The named examples follow the rule. Trilobites are particularly useful for sediments of Cambrian age. Ammonites and inoceramids are particularly useful for correlating environmental events around the world during the super-greenhouse of the Late Cretaceous. Each marks a bounded interval, which is exactly why each is diagnostic.

Applying the four criteria to feedback

Run your themes through the same filter and the ranking inverts.

The second criterion is what keeps this honest. Narrow occurrence alone is not enough, or every one-off complaint would be a finding. A diagnostic signal is narrow in time and broad in space: it shows up in one window, across many independent accounts. A signal narrow in both is an anecdote. A signal broad in both is background.

So what does your permanent top-three theme actually tell you? "Onboarding is confusing", "the search is not good enough", and "we want better reporting" appear in every corpus, from every segment, in every quarter, for every B2B product ever built. They are the trilobites that never went extinct: abundant, easy to identify, and carrying no information about which period you are in or what changed. They are real. Customers genuinely mean them. They simply cannot discriminate between the world where your last three releases worked and the world where they did not.

Meanwhile the theme mentioned by nine people across six segments only since March is the ammonite. It has a bounded range and wide distribution. It is telling you something happened in March.

Why mention-count ranking destroys the signal

Sorting a theme list by frequency is a specific analytical error, and it is worth naming the mechanism.

Frequency ranking treats every mention as an equal vote, which implicitly assumes each mention carries equal information. It does not. A mention of a perennial theme carries almost none, because you could have predicted it before collecting any data. A mention of a time-bounded theme carries a great deal, because it would not have appeared last quarter.

Worse, the two classes are not merely different in value: they compete for the same slot. Perennial themes are perennial because they are broadly experienced, so they will always out-count a newly-emerged theme in raw volume. The ranking is therefore structurally guaranteed to bury the diagnostic signal underneath the background one, permanently, no matter how good your tagging is. This is a failure of the sort order, not of the coding.

The fix is not to weight the counts. It is to compute a different quantity: for each theme, its range. Which periods does it appear in, and which segments? Then read the themes whose range is narrow in time and wide in segment first, and treat everything with unbounded range as environmental background that needs a roadmap decision once, not a re-litigation every quarter.

This is a different concern from whether a quote needs its context to mean anything, which is the subject of why a tagged quote is not evidence. That question is about a single item's relationship to its surroundings. This one is about which items carry discriminating power at all.

The assemblage is still the unit

None of this means a lone diagnostic quote settles a question. An assemblage is a set of artifacts or ecofacts found together, from the same place and time, and in archaeology two or more excavated objects that are thought to be related are said to be in association, which is how the date of a particular context gets established.

The two ideas fit together cleanly. The assemblage is the unit of analysis; the index fossil is what dates it. You read the whole co-occurring set to understand a context, and you use the few diagnostic members to place it in time. Practically: report the full pattern of what a segment said, and lead with the one or two time-bounded signals that tell you when and why the pattern shifted.

The biozone concept completes the picture. A biozone is the interval marked by overlapping fossil ranges, from the appearance of one species to the appearance of the species defining the next zone. The equivalent is genuinely useful: your product's history divides into intervals defined by which diagnostic themes were live. Naming those intervals gives you a chronology of your own product that a flat theme count cannot produce.

How Koji handles this

Identifying diagnostic themes requires two things most stacks lack: the same questions asked across many periods, and theme data with dates and segments attached.

• Consistent protocol across time. Koji's AI interviewer asks the same core questions in every wave, so a theme's appearance and disappearance reflect the customers rather than a change in who was moderating or how they phrased the question. Range is only measurable against a fixed instrument.
• Dated, segmented theme tags. Koji tags themes on every interview and keeps each one bound to its transcript and its interview date, which is what makes a range calculation possible instead of a guess.
• In-vivo tagging preserves distinctiveness. Koji can tag a theme in the participant's own framing rather than collapsing it into a house label. That matters directly here: forcing a sharp, novel complaint into a generic bucket like "usability" is how a diagnostic signal gets reclassified as background.
• Structured questions bound the range precisely. Koji's six question types, open_ended, scale, single_choice, multiple_choice, ranking, and yes_no, let you attach a countable measure to each wave, so you can see the exact window in which a scale score moved rather than inferring it from prose. See structured questions in AI interviews.
• Real-time reports show emergence as it happens. Because Koji analyzes each conversation as it completes, a newly-emerging theme is visible while the study is still running, which is the only moment at which a narrow-range signal is actually actionable.
• Programmatic range analysis. Koji's MCP tools, including koji_get_study_data and koji_export_data, let you pull themes with their dates and segments and compute range directly, rather than eyeballing a frequency chart.

Used this way, Koji stops being a machine that produces a ranked list of complaints and becomes one that tells you which complaints are dating your product.

Frequently asked questions

Why is my most-mentioned theme not my most important one?

Because a theme mentioned in every period and every segment cannot discriminate between states of the world. It is consistent with your product being excellent and with it being broken, so its presence tells you nothing about which is true now. High frequency with unbounded range is the signature of background, not of priority.

Does this mean I should ignore perennial complaints?

No, it means you should decide about them once rather than rediscovering them every quarter. Perennial themes are real constraints and often deserve a standing roadmap commitment. What they do not deserve is the top slot in a quarterly research readout, where they crowd out the signals that actually changed.

What makes a theme diagnostic?

Four properties, borrowed directly from index fossils: a bounded time range, a wide distribution across segments, enough mentions inside that window to be more than noise, and a description clear enough that two analysts would tag it the same way. Narrow in time and broad in segment is the combination that carries information.

Is a single narrowly-occurring quote a finding?

Not on its own. A signal narrow in both time and distribution is an anecdote, and the second criterion exists precisely to catch this: an index fossil needs wide geographic distribution as well as a short range. One vivid complaint from one account in one month fails the test until other accounts show the same thing in the same window.

How is this different from the archival bond idea?

The archival bond concerns a single record's dependence on its context, covered in why a tagged quote is not evidence. This piece concerns which records have discriminating power in the first place. A quote can be perfectly contextualized and still be a trilobite, present in every layer and useful for dating none of them.

How do I compute a theme's range in practice?

For every theme, list the periods it appears in and the segments it appears across, then sort by narrow-in-time and wide-in-segment. If your themes carry interview dates and participant attributes, this is a single query. Koji's exports and MCP tools expose theme, date, and segment together so the calculation needs no manual coding pass.

Related Resources

• Structured Questions in AI Interviews - the six question types, and how consistent measures let you bound a theme's range
• Why a Tagged Quote Is Not Evidence - the companion problem of a record's context
• Customer Quotes - extracting and tagging verbatims so the diagnostic ones stay findable
• The Complete Guide to Thematic Analysis - producing the theme set this method then ranks by range
• How to Build a Qualitative Research Codebook - keeping theme definitions distinctive enough to be identifiable
• Crossdating Customer Stories - using diagnostic signals to place an account against known patterns