Plan a pilot box
Menu

Item-level or collection-level description for digitized archives

Reader question: When we digitize a collection, do we need to describe every item, or is a collection-level finding aid with linked images enough?

By Dastok Editorial TeamIssue: Published

No, you rarely need to describe every item. For most digitized archives, describe at the level people search: a collection-level or series-level finding aid with images linked to its folders, plus searchable text where the originals contain text. Add item-level records where the material demands them, chiefly photographs, maps and objects that carry no text of their own, and items your users cite one by one. Every digitized item still needs a small set of fields, such as an identifier, a date, a rights statement and a link to its place in the collection, so the files stay findable and importable.

That answer has become more nuanced in the last few years, because machines can now draft item-level fields from recognized text at low cost. Drafts are not description, though. This article explains the levels, the case for each, a rule by material type and expected use, the minimum fields, and how to use machine-drafted item records without misleading your users.

What “level of description” means

The level of description is the unit a catalog record describes: a whole collection, a series within it, a file or folder, or a single item. Archives usually describe from the top down, and many stop above the item.

The Society of American Archivists’ content standard, DACS (Describing Archives: A Content Standard, version 2022.0.4), puts it this way in its chapter on levels of description: “Archival material can be described at many different levels.” A finding aid may consist of a single level of description or many, and a multilevel finding aid “may provide information at successively narrower levels of arrangement (such as subseries, files, and even items) for some series while confining information to a single level of hierarchy for others.” Crucially, “DACS does not attempt to define the proper level of description for any set of archival materials. Archivists should follow the prescriptions of their institutions and apply their own judgment in making such determinations.”

So the standard leaves the decision to you. That is the right place for it, because the right level depends on the material and on what people will look for.

The levels in plain terms

Level What one record describes Typical example
Collection Everything from one source or creator The Okerlund family papers, 1893 to 1896
Series A group kept together by function or form Correspondence; financial records
File or folder The contents of one folder Letters, Ida Okerlund to Marit Okerlund, 1893
Item One document, photograph, map or object One letter dated April 12, 1893

In a digitized collection, images attach at the bottom of whatever hierarchy you build. If the lowest level is the folder, a folder’s images are viewed together, in order, the way a researcher in the reading room would turn the pages.

Levels and the systems you use

Your systems shape what is practical. Archival management systems are built around hierarchy: a collection record with series, folders and sometimes items beneath it, exported as a finding aid. Many online collection platforms, by contrast, are built around single objects: one record, one or more images. When the two meet, a folder of forty letter images may have to become either one object with forty pages or forty separate objects. Neither is wrong, but the choice affects how users browse and how search results look, so decide it with the platform in front of you rather than in the abstract. Before committing, try importing one folder each way and look at the result as a user would.

The case for collection-level description

Collection-level and series-level description is the archival default for good reasons. Archival records are usually used in context: a letter means more beside the letters around it, and a ledger page means more inside its volume.

Context is information

An archival record’s meaning often depends on what sits next to it. A letter that mentions “the trouble” makes sense only beside the letters before and after it, and a ledger entry for flour on credit means more when you can see the weeks of entries around it. Folder-level and series-level description preserve that context by design: the images stay in their original order and the record describes the group. Item-level records can lose it unless they carry a link back to the folder and series, which is why “part of” is one of the minimum fields below.

More Product, Less Process

The strongest argument comes from “More Product, Less Process” (MPLP), a 2005 article by Mark Greene and Dennis Meissner. The SAA’s Dictionary of Archives Terminology defines MPLP as “minimal processing to arrange and describe archival series and collections to reduce or avoid backlogs,” and summarizes its principles: that “use is the end of all archival effort”; that backlogs hinder use and undermine the confidence of resource allocators and donors; and that processing decisions should consider “the mission, audience, and resources of the present” rather than the traditions of the past. The dictionary entry also notes that minimal processing “might shift somewhat more work onto reference staff.”

Applied to digitization, the MPLP logic says: get the collection online, described well enough to be found and used, rather than describe a few items exhaustively while the rest waits.

Full text changes the calculation

For text-heavy material, recognized text does much of the work item description used to do. If every page of a letter series is searchable, a researcher looking for “footbridge” or “Wren Gulch” will find the right page without anyone having written an item record for it. The text is not description, but it makes a folder-level record far more useful than it would be over unsearchable images. That only holds if recognition is good enough on your pages, which you should measure; our text recognition page explains how.

What collection-level description costs you

Collection-level description has real limits:

  • Images without text are invisible to text search. A box of photographs described only at the folder level cannot be searched for the people, places or events in them.
  • Aggregators and some systems expect item records. If you plan to contribute to a statewide or national portal, check whether it accepts folder-level objects.
  • Citations and reuse. Researchers often want to cite or reproduce one item, which is easier when it has its own identifier and landing page.
  • Reference load. As the MPLP literature itself notes, thin description can move work to reference staff, who answer the questions the catalog does not.

The case for item-level description

Item-level description is the norm for many museum collections and for visual materials in libraries and archives. It is worth the cost when each item is likely to be sought, shown or reused on its own.

Item records pay off for:

  • Photographs, negatives and slides. They usually carry no searchable text, so the record is the only way to find them by subject, person, place or date.
  • Maps, plans and drawings. Users search by place and date, and each sheet is often used alone.
  • Objects and artworks. A collection management system needs one record per object, with dimensions, materials and provenance.
  • Items that are famous or heavily requested. The letter everyone asks for deserves its own record and a stable link.
  • Material bound for an aggregator that requires item-level records with specific fields.

The museum object record

Museums usually work at the item level for a structural reason: the object is the unit of accession, loan, insurance, conservation and display. Each object has an accession or catalog number and one record that collects its description, dimensions, materials, provenance and location. Digitization adds images to that existing record rather than creating a new layer of description. For museums the decision is less about level and more about depth: how many fields to complete for each object, and which images (overall views, details, marks, inscriptions) each record should carry. The minimum fields below still apply, and the identifier is simply the object number.

The cost is time. Item-level cataloging can take minutes to an hour per item depending on depth, research and authority work, which is why it is usually applied selectively. Our digitization budget calculator separates cataloging time per item from capture time, so you can see how much a choice of level changes the total.

Item fields for photographs

Photographs need a slightly richer minimum than documents, because the record is often the only way in. A practical item record for a photograph includes:

  • A devised title that says what the image shows in plain words, such as “Two girls in matching dresses, studio portrait.”
  • Date, exact or approximate, with the basis noted when it is inferred, for example from the mount style, the photographer’s active years or clothing.
  • Names of people where known, with the source of the identification (a pencil note on the back, a donor, a published caption).
  • Place shown or the studio location, with the source.
  • Photographer or studio, from a stamp, blind stamp or imprint on the mount.
  • Format and process, such as cabinet card, glass plate negative or gelatin silver print, which also helps with dating.
  • Inscriptions, transcribed exactly, front and back.

Capturing the back of every print and mount is what makes most of these fields possible. A fifteen-second capture of a pencil note can save an hour of research later.

A rule by material and expected use

The decision is easiest when you take each series in turn and ask two questions: does the material carry its own searchable text, and will people seek individual items? The table gives a starting point.

MaterialSearchable text after recognition?Sought item by item?Suggested level
Correspondence seriesYes, if handwriting recognition is adequateSometimesFolder level, full text, item records for notable letters
Ledgers and registersPartly; tables need careRarely; pages within volumesVolume level, page-level text and page numbers
Minutes, reports, typescriptsYesRarelySeries or folder level, full text
Newspapers and clippingsYes, with layout workBy date and issueIssue or clipping level, full text
Photographs, negatives, slidesNo, except captions and backsOftenItem level, with a minimal record for every image
Maps and drawingsLittleOftenItem level (sheet), with place and date
Museum objectsNoYesItem level (object)
Suggested starting level of description by material and expected use. Editorial guidance from the Dastok Editorial Team, based on DACS Chapter 1 (levels are a local judgment) and the MPLP principle that use drives processing decisions; adjust for your users. Checked November 2026.

Mixed approaches are normal

Most real collections end up mixed: folder-level description for the bulk, item records for photographs and a few notable documents, and full text over everything that has text. DACS explicitly allows a finding aid to go deeper for some series than for others. The point is to decide deliberately per series and write the decision down, so that the next phase of digitization follows the same rule.

How users search digitized collections

The right level depends on how your users look for things, and different audiences search differently. Genealogists usually search by name, often with spelling variants, and want to land on the page that mentions an ancestor. Local historians search by place, building or event and browse what surrounds a hit. Academic researchers often start from a finding aid and read whole series. Students and casual visitors search a single word and judge results by thumbnails. A collection with full text over folders serves the name searcher and the browser well; item records with good titles serve the casual visitor and the photograph searcher. If you know which audience matters most for a series, the level of description follows.

Three quick cases

  • A minute book series from a town council. Typed or neatly written, used to trace decisions over time, rarely cited one page at a time. Describe at the volume level with page-level text. Item records add little.
  • A box of unidentified family snapshots. No text beyond a few pencil notes, valued for faces and places. Give every image a minimal item record with an approximate date and whatever the backs say, and plan to deepen records as people are identified.
  • A set of hand-drawn survey maps. Each sheet covers a different area and is used alone. Describe each sheet as an item with place, date, scale where stated and surveyor, and link the set together at the series level.

Questions that move the decision

  1. Who are your users, and how do they search? Genealogists search names; historians often browse series; local residents search places and buildings.
  2. What will the platform display? Some systems show a folder of images well; others expect one record per image.
  3. Is the material going to an aggregator? Check its required fields before you choose.
  4. How good is the recognized text? If the character error rate on a hand is high, folder-level description over poor text may leave items unfindable, and item-level fields for names and dates become more valuable.
  5. What can you afford now, and what can wait? A folder-level first pass can be deepened later; the reverse rarely happens.

The minimum fields for every digitized item

Even when you describe at the folder level, every digitized item (or every image, for photographs) needs a minimal record so that files can be identified, moved between systems and connected back to their context. Think of these as housekeeping fields, not description.

Field Purpose Dublin Core element
Identifier Ties the file to its record and to its box, folder and item identifier
Title (devised if needed) A human-readable label such as “Letter, Ida Okerlund to Marit Okerlund” title
Date Exact or approximate, in a consistent format date
Part of The collection, series and folder the item belongs to relation
Type and format Text, image, physical object; file format type, format
Rights A standardized rights statement rights
Source Where the original is held source

The Dublin Core element names come from DCMI’s Metadata Element Set, which defines fifteen elements: contributor, coverage, creator, date, description, format, identifier, language, publisher, relation, rights, source, subject, title and type. For rights, use one of the twelve statements published by RightsStatements.org, so that the value is machine-readable. Our metadata mapping worksheet helps map your own field names to these elements, and our comparison of Dublin Core, MARC and EAD explains when each standard fits.

Dates that sort and filter

Dates deserve a convention of their own because they drive sorting, filtering and timelines on almost every platform. Record exact dates year first (1893-04-12) so they sort correctly. For approximate or uncertain dates, pick one machine-readable convention, such as the Library of Congress’s Extended Date/Time Format, which marks an approximate year with a tilde and an uncertain one with a question mark, and keep a human-readable display form beside it (“about 1895”). Record date ranges for folders and series as a start and an end. Whatever you choose, apply it the same way across every series and every phase of the project; mixed conventions are one of the most common reasons an online collection’s date filter returns odd results.

Identifiers and file names

The identifier is the field that holds everything together. Use the same identifier, or a recorded link between identifier and file name stem, in the catalog record, the file names and the preservation package. A file named for its box, folder, item and page sorts correctly and can be traced to its original; the file naming builder produces names of that kind. If you change systems later, these fields are what allow the import to reconnect records and images, as our guide to importing digitized records shows.

Making folder-level records work online

A folder-level record that works in a reading room can fail online, because a remote user has no archivist to ask. If you choose collection, series or folder description for a digitized series, make each record carry enough to stand on its own.

What DACS requires at minimum

DACS sets out the elements a description must include at minimum. The required elements in the standard are reference code, name and location of repository, title, date, extent, name of creator (if known), scope and content, conditions governing access, and languages and scripts of the material. For a digitized folder, those translate into practical questions a remote reader will have: what is in this folder, who made it, when, how much of it there is, whether it is complete, and whether they may use the images.

Write scope notes for searchers

The scope and content note is where folder-level description earns its keep online. Write it for someone who arrived from a search engine, not for a colleague. Name the people, places, events and kinds of documents that appear, in the words users search with. A note that reads “Family correspondence” helps nobody; one that reads “Letters from Ida Okerlund in Wren Gulch to her sister Marit, 1893 to 1896, describing a spring flood that destroyed the schoolhouse footbridge, a strike at the smelter, a Fourth of July band from Colorado Springs and their mother’s illness” gives search engines and readers something to match. That note is fictional, but its form is the point: specific nouns, dates and places.

Keep the order and say so

Images linked to a folder should appear in the order of the originals, and the record should say what that order is (chronological, as found, or rearranged). Note gaps, such as missing pages or items withheld for privacy, so that a reader does not assume the folder is complete when it is not. If an item has been pulled for item-level description, leave a pointer in the folder record so the reader can follow it.

Page-level navigation

Folder-level records benefit from page numbers or image sequence numbers that a reader can cite, and from recognized text tied to each page so search hits open the right image. A long folder without page navigation forces readers to scroll through dozens of images; with page-level text and numbering, the folder behaves almost like a set of item records without the cataloging cost.

The middle path: item records drafted by machines

Machine learning has changed the cost of item-level description, but not its quality bar. Language models can read recognized text and suggest names, places, dates and subjects for each item; image models can suggest what a photograph shows. That makes item-level records possible where they were once unaffordable, as long as people review what matters.

What the evidence says

The Library of Congress tested this directly. In its Exploring Computational Description experiments, catalogers reviewed machine-generated metadata, and the Library concluded that “because library catalog records must contain highly reliable and accurate metadata, no current AI tools are currently good enough to run fully automatically for this task.” It reported that some fields, “identifiers, author and title,” could be extracted “at around 85% accuracy, while others like subject headings, genre, and dates were only generated accurately at or under 50% of the time.”

Those findings come from published books, not handwritten letters, but they point the same way for archives: extraction of names and identifiers from good text is useful; subject headings and dates need a person.

How to use machine drafts responsibly

  1. Draft only what the text supports. Names, places and dates found in recognized text are good candidates. Interpretive fields such as subjects and summaries need closer review.
  2. Label every machine-made field until it is reviewed. Researchers should be able to tell a reviewed date from a suggested one. Our guide to labeling AI-generated metadata shows field-level patterns, and our AI policy describes how we apply them.
  3. Review by priority. Review every field that controls access (dates, names, rights) and sample the rest. Record the sampling, so the catalog’s reliability is known.
  4. Keep the folder-level description authoritative. Item drafts sit beneath a reviewed collection and series description, which provides the context machines cannot.
  5. Measure in a pilot. Count how many suggested fields survive review in a sample before you plan the whole collection around them.

A review workflow small teams can sustain

A review process that depends on one cataloger reading every suggestion in every record will stall. A sustainable pattern splits fields by risk:

  1. Fields that control access are always reviewed. Dates, names of people and rights statements decide who finds an item and whether it can be shown, so a person checks every one.
  2. Descriptive fields are sampled. Subjects and summaries are reviewed on a random sample of each batch, and the batch is accepted or sent back based on the error rate found.
  3. Low-confidence suggestions are routed to people. Where the tool reports its own confidence, send the least confident suggestions for review first.
  4. Corrections are recorded. Keep a short log of what reviewers changed. It shows where the machine is weak and whether a different prompt, model or vocabulary would help.
  5. Labels change when review happens. A field moves from “machine suggestion” to “reviewed” only when a person has checked it, and the record shows which.

This keeps the catalog honest about what has been checked while letting machine drafts cover the long tail of items nobody has time to describe by hand.

What it does not replace

Machine drafts do not replace authority work (deciding that “W. Okerlund” and “Walter Okerlund” are the same person), interpretation of photographs with no text, or the judgment about what a collection is about. For photographs with unknown sitters and places, our article on cataloging unidentified photographs covers what can and cannot be inferred.

A worked example: one fictional box

Example (fictional). The Aster County Historical Society digitizes Box 3: 41 letters (96 pages), 18 cabinet cards, a 212-page store ledger and 9 clippings and a telegram, 1893 to 1896. It decides the level of description series by series:

  • Letters: folder-level records under the series “Correspondence, 1893 to 1896,” with every page’s recognized text searchable. Machine-drafted item records list names, places and dates for each letter; a cataloger reviews names and dates in every record and samples subjects. The letter about the washed-out footbridge, already requested twice, gets a fully reviewed item record.
  • Cabinet cards: item-level records for all 18, because they carry almost no text. Each record has a devised title, an approximate date (“about 1895,” reasoned from the mount style and labeled as such), the studio name from the mount and the pencil names from the back.
  • Ledger: one volume-level record, with page-level images, page numbers and recognized text. No item records for individual entries.
  • Clippings and the telegram: item-level records, because each is short, dated and likely to be cited alone.

Every image, whatever its level, has an identifier such as ACHS-3-012, a rights statement and a link to its folder. The society can see the whole box online, search the letters’ text, browse the cabinet cards by person and date, and know which fields a person has checked. Our example collection shows records like these.

The example shows the general pattern: description follows the material and the use, and a single collection carries several levels at once.

Deepening description over time

Choosing folder-level description now does not close the door on item records later. Many institutions treat description as iterative: a first pass that makes everything findable, then deeper records where use shows they are needed. Practical triggers for deepening include:

  • Reference requests. When staff answer a question about a specific item, they add or improve its record at the same time, so the next person finds it without asking.
  • Exhibitions and anniversaries. Items chosen for a display or a commemoration get full records because they will be seen and cited.
  • Researcher contributions. Researchers and descendants often identify people in photographs or transcribe difficult passages; record their contributions with their source, and review them before they become catalog data.
  • New text recognition. Recognition models improve. Re-running recognition on a difficult hand a few years later may make folder-level description over full text adequate where it was not before.

Each of these adds value without a separate cataloging project. The condition is that the identifiers and the folder structure were consistent from the start, so new records attach to the right images.

Planning the description work

Description is often the largest staff cost in a digitization project, so the level you choose should be planned and budgeted, not left to whoever catalogs first.

Steps

  1. Inventory by series. Use a collection inventory worksheet to record each series’ format, extent and condition.
  2. Choose a level per series using the material-and-use questions above, and write the decision down with its reason.
  3. Define minimum fields for every item and the fuller field set for item-level series, mapped to your system and to Dublin Core where you publish.
  4. Measure description time in a pilot, separately for folder records, reviewed item records and machine-drafted item records.
  5. Budget and schedule from the measured times, and decide which series get deeper description now and which can be deepened later.

Estimating time

Time per record varies widely with depth and research. Measure it rather than borrowing someone else’s number. Time a sample of 20 records of each kind in the pilot, including review of machine suggestions, then multiply. The digitization budget calculator takes cataloging minutes per item as an input, so you can compare scenarios, such as folder-level for all letters against item-level for all letters, before committing.

Where outside help fits

A collection assessment can produce the series-by-series plan and measured times before any capture starts, and AI cataloging produces machine-drafted item fields with every suggestion labeled for your catalogers to review. Either way, your staff make the decisions about level and review, because they know the users.

Write the decision down

Record each decision in a short description plan that travels with the project. For each series, note the level chosen, the reason, the minimum and full field sets, the vocabulary used for names and subjects, whether machine suggestions are used and how they are reviewed, and the measured time per record. A plan of this kind fits on a page or two. It keeps description consistent when different people catalog different boxes, it explains to a funder or board why the project made the choices it did, and it lets the next phase start from a tested rule instead of a new debate.

Example (fictional). A second fictional institution, the Red Mesa County Museum, holds about 4,000 photographic prints and 300 objects from a closed railroad depot. Its plan sets item-level records for every object, using its existing accession numbers, and minimal item records for every photograph, with devised titles and approximate dates, drafted by a model from the captions on print backs and reviewed for names and dates. Photographs with no caption at all get a folder-level record for now and a note inviting identifications from visitors. The plan budgets two minutes per photograph record and twenty minutes per object, figures the museum measured in its pilot, and schedules object records first because they feed a planned exhibition.

Common mistakes and how to avoid them

  • Describing everything at item level by default. It slows the project and leaves most of the collection waiting. Choose per series.
  • Describing nothing below the collection for photographs. Images without text cannot be found. Give every photograph at least a minimal record.
  • No identifiers linking files to records. Without them, a system migration breaks the link between images and description.
  • Unlabeled machine fields. Users cannot tell suggestions from reviewed data, and errors spread into citations.
  • Inconsistent dates. Mixing “ca. 1895,” “1895?” and “about 1895” makes sorting and filtering unreliable; choose one convention and record approximate dates the same way every time.
  • Deciding once for the whole institution. Different series and formats need different levels; the rule should be by material and use.
  • Losing the order. Linking images to a folder but presenting them out of the original order destroys the context folder-level description exists to keep.
  • Treating approximate dates as exact. A date inferred from a mount style or clothing should say so, both for honesty and so that later research can revise it.

What to do next

Pick one series you plan to digitize and answer five questions: Does it carry its own text? Will people seek single items? What does your platform expect? What do your users search for? How long does a record take at each level? Write the answers down, choose a level, and test it in a pilot before applying it to the whole collection.

If you are starting from scratch, our overview of what digitizing a historical collection involves puts description in the context of the whole project. If you want to see the levels on real material, a pilot box run through every stage, including description, is described on our engagements page.

Working through this with a real collection?

If this article describes a decision you are facing, tell us about the collection. A pilot box answers most questions with your own material.

Plan a pilot box

Studio and contacts

24 South Weber Street, Suite 225
Colorado Springs, CO 80903
(719) 244-8450sales@dastok.com