Reusing Past Market Research: How Old Is Too Old?

Reusing past market research saves a field budget until the answer turns out stale. What every archive answer must show: source, field date, permission, cont...

A brand manager types a question into the new insights assistant on a Tuesday afternoon. Do families with small children still choose the larger pack for the weekly shop? The answer arrives in four seconds, fluent and confident, with a citation to a usage and attitudes study. Nobody in the meeting opens the citation. If they had, they would have seen a field date from three years ago, in a market where the category price has since moved twice. Reusing past market research is one of the best habits a research operation can build, and this is the exact moment where it goes wrong.

Key Takeaways

• An answer from the archive is only usable when it shows four things: the source, the field date, the permission to see it, and the context of the study.
• A 2026 benchmark from Meta Superintelligence Labs and UCLA found that the component reordering retrieved documents favours older, richer documents even when the fact has changed.
• A correct looking answer can hide whether the model followed the archive or its own memory. Accuracy alone does not show it.
• How long a finding stays valid depends on what kind of finding it is. Stable behaviour ages slowly, and price and media habits age fast.
• Checking the archive before commissioning fieldwork is a formal step with an owner, and it can end in reuse, partial reuse or new field.

What a living research archive is, and what it is for

A living research archive is the full body of a research operation's past work, kept findable and queryable after each project closes. That means reports, questionnaires, transcripts, decks, data files, dashboards and the record of what was decided with them. Living means somebody maintains it, and that the date of each finding stays attached to it.

The alternative has a name in the trade press. Greenbook's GRIT commentary puts it plainly: accumulation without access is an archive. The same piece describes questions funded again and again because no one can locate the original answer. Qual and quant sit in separate places. That is an opinion piece with no figures in it, and it matches what most insights managers would say about their own shared drive.

A field date is the date the data behind a finding was collected. It is different from the date the deck was saved, the date the report was approved, and the date somebody last touched the file. Only the field date tells you how old the answer is.

A 360 study is several methods pointed at the same business objective, read together. A usage survey, two focus groups and a round of shopper interviews on one category question form one body of evidence, rather than three files in three folders.

How an old study gets reused today

PERSONAL EXPERIENCE Most reuse I have seen follows the same route, and the route skips the one check that matters.

1. A new brief arrives, and somebody remembers that a similar question was studied before.
2. The search starts with whoever ran that study, by message, and then moves to a folder organized by client and year.
3. A deck is found, and the slide that looks closest to the new question is copied.
4. The slide goes into the new document with the chart intact and the footnote cropped.
5. The finding is presented as current, because nothing on the slide says otherwise.

Step four is where the date disappears. The footnote carried the field period and the base, and it was the first thing trimmed to make the slide fit.

Quirk's describes the underlying condition well: most organizations are rich in research and poor in reuse. The piece adds a point that stings. The more polished and tailored the deliverable, the less reusable it is, because decks are built to persuade a room rather than to be found and compared later. An assistant sitting on top of the archive does not fix that. It makes step three faster.

Why the search layer can prefer the old study

This is the part most teams do not expect, and it has been measured.

A re-ranker is the second pass in an archive search. The first pass collects candidate passages, and the re-ranker reads the question against each one and decides which reach the model. We described that arrangement in our piece on the market research knowledge base, so here the only question is what it does with time.

In April 2026, researchers at Meta Superintelligence Labs and UCLA published FRESCO, a benchmark built for exactly this. It pairs questions that ask for the recent fact with older revisions of the same Wikipedia pages. The authors report a consistent failure across existing re-rankers: a strong bias toward older, semantically rich documents, even when those documents are factually obsolete.

Read that across to a research archive. The old usage and attitudes study is long, detailed and written in the vocabulary of the question. The recent tracker wave is a short topline with three charts. On semantic richness alone, the old study wins the ranking, and the answer inherits its date.

The authors also report that tuning the instruction given to the re-ranker recovered up to 27 percent on the tasks where knowledge changes. That is their measurement on their benchmark, and it is not a promise about any archive. What carries over is the direction. Recency has to be asked for on purpose, because the default ranking does not reward it.

A correct answer can hide where it came from

UNIQUE INSIGHT The second problem is quieter. An answer can be right for the wrong reason, and the archive cannot tell you which.

A May 2026 paper, Does RAG Know When Retrieval Is Wrong?, asks whether a model followed the retrieved evidence, relied on what it already believed, or wrote a rationale after the fact. The authors argue that judging only whether the final answer is correct hides that question. Under their worst case probe, where misleading material was injected into retrieval, standard retrieval answered correctly only 15.0 percent of the time. They frame the whole issue as observability, which is the right word.

Retrieval augmented generation, usually shortened to RAG, is the arrangement where a model searches a document store first and answers from the passages it found. The paper that introduced it in 2020 already listed two open problems: giving provenance for the model's decisions, and updating its knowledge of the world. Six years on, a research archive meets both on every question.

For an insights team this means the citation under the answer is necessary and still not enough. A citation shows that a document was retrieved. It does not show that the sentence above it came from that document, or that the document is still true.

The AAPOR task force report on responsible AI in survey research makes the same point from the survey side. When AI acts as a briefer, summarizing findings for stakeholders, the risk rises if the summary abstracts away uncertainty, caveats or methodological context. The report adds that fluent output invites over interpretation by non expert users, which describes the Tuesday meeting exactly.

Four things every answer from the archive has to show

UNIQUE INSIGHT The fix is a display rule, and it is short. Every answer carries four fields, visible to the reader, before anyone decides anything with it.

Source is the study, the document and the page or passage the answer came from. Field date is when the data was collected. Permission is whether the person reading, and everyone else in the conversation, is allowed to see that document. Context is the market, the sample, the method and the question as it was asked in the field.

Four things every answer from the archive shows

At the top, a dashed band stands for the answer on screen: a finding, fluent and cited. Two arrows lead down to four numbered cards, each naming one field the answer has to display and the question that field answers. Card one, source: which study, which document, which page or passage, answering where the sentence came from. Card two, field date: when the data was collected, not when the deck was saved, answering how old the finding is. Card three, permission: checked for everyone in the conversation, before retrieval, answering whether all of the people involved may see it. Card four, context: the market, the sample, the method and the question as it was asked, answering whether the finding fits the new question. A closing band says that an answer missing any one of the four goes back to the reader as a lead to check rather than a finding to use.

The citation alone covers the first card. The other three are the ones that get cropped on the way to the slide.

An answer missing any one of the four goes back to the reader as a lead to check rather than a finding to use. That rule costs a few seconds per question. It saves the meeting where a three year old number is presented as this quarter's.

The Quirk's piece arrives at a similar record from the knowledge management side. It proposes a canonical insight record carrying the claim, the evidence, the context, a confidence level and the provenance, including dates and method. The difference here is where the fields live. A record in a separate catalogue gets skipped. Fields printed inside every answer do not.

Permission belongs inside the answer, not after it

A shared archive puts studies from different clients, business units and confidentiality levels behind one search box. The permission check has to happen before the passage reaches the model, not after the answer is written.

Enterprise AI Must Enforce Participant-Aware Access Control, a 2025 position paper, states the principle cleanly. Any content used in training, retrieval or generation has to be explicitly authorized for every user involved in the interaction. The authors show that prompt sanitization, output filtering and system isolation are probabilistic defences. A filter that usually catches a confidential passage is a filter that sometimes does not.

Participant aware access control means the archive checks the permissions of everyone in the conversation, not only the person who typed. In a research setting that matters in a specific way. An analyst with full access shares a session with a client, or pastes an answer into a channel the client can read. The answer has now crossed a permission line that the retrieval step never saw.

The industry bodies are pointing the same way from the contract side. Guidance from MRS, ESOMAR and the Insights Association, reported in June 2026, notes that a de-identified dataset, an aggregated summary or a synthetic output can still carry the client's strategic thinking. A summary of an old study belongs to the same client as the study. An archive that forgets that has turned a confidentiality obligation into a search result.

How old is too old: shelf life by type of finding

UNIQUE INSIGHT There is no single number, and anyone who gives you one is guessing. What we use is a house rule by type of finding, with a question that decides it, and no year counts attached.

The table is our operating rule, not a measurement. The numbers you might want in those cells would have to come from your own trackers, category by category, and we have not seen a published source that gives them honestly.

Two habits make the table work. The first is to read the field date before the finding, every time. The second is to treat a reused finding that fails its question as a hypothesis for new fieldwork rather than as a dead file. The old study still tells you what to ask.

Desk research inside the house, before anyone commissions field

The expensive version of this problem is the study commissioned because nobody proved the answer already existed. The cheaper, quieter version is the study commissioned without anyone checking what the archive already knows, so the questionnaire repeats questions and misses the gap.

Internal desk research is the formal step of searching the organization's own archive for the new question before a brief becomes a field proposal. It is the same discipline as external desk research, pointed inward, with an owner and a written result.

The sequence we recommend, before any new fieldwork is approved:

1. Write the business question in one sentence, the way the decision maker would ask it.
2. Search the archive for every study that touched it, across methods, and list them with their field dates.
3. Check that everyone who will see the result is allowed to see each study on the list.
4. Test each finding against the shelf life question for its type.
5. Mark each part of the question as answered, partly answered, or open.
6. Commission fieldwork only for the open parts, and design it to update the partly answered ones.

Before commissioning new fieldwork

A decision flow in four steps, each with an exit to the right. Step one asks whether the archive holds an answer, which means listing every study that touched the question, across methods, with field dates. If nothing is on file, new fieldwork is justified and the proposal says so. Step two asks whether everyone involved may see it, with permission checked for each person who will read the result. If not for everyone, the case is escalated to the data owner before the study is used. Step three asks whether the finding is current for its type, using the shelf life question for behaviour, attitude, price or media. If something relevant changed, the finding is kept as a hypothesis and that part is refreshed in field. Step four asks whether the context matches: same market, a comparable sample, the question asked the same way. If it matches only partly, the matching part is reused and only the open part goes to field. A closing band says that four yes answers mean the finding is reused with the source and the field date printed on the slide.

Every exit to the right is a legitimate result. The one outcome this check rules out is paying twice for an answer nobody looked for.

The written result of step five goes into the proposal. It shows the client that the new money is going only where the archive has nothing current to say. That line lands well with procurement, and it is true.

Several methods on one objective: the 360 study

UNIQUE INSIGHT An archive organized by project hides a good deal of its value. An archive organized by objective shows it.

Most categories have been studied more than once, by more than one method, over several years. A segmentation, a round of in-home interviews and a pack test may all have been answering pieces of the same question: why this household chooses this format. Read one at a time, each looks partial. Read together, with their field dates side by side, they form a 360 study that no single project could have paid for.

The discipline is the same one we described for keeping the evidence chain in qualitative coding. Each claim stays attached to the study, the method and the date it came from, even when the synthesis crosses all three. A 360 read that loses those attachments is a new opinion wearing old evidence.

The date matters more here than anywhere. Combining a recent quantitative wave with qualitative work from years earlier is fine when both are labelled. It is misleading when the synthesis reads as one moment in time.

When the person who ran the study leaves

PERSONAL EXPERIENCE The archive's real test is not a search. It is the week after somebody senior leaves, and a client asks a question only that person could have answered.

Quirk's names this too: when a key researcher leaves, the unwritten context goes with them. The context is exactly the fourth field in the display rule. Why the sample excluded a region. Why a question was worded that way. What the client was really worried about that year.

The practical answer is dull and it works. Context gets written into the archive at project close, by the person who ran the study, as a short note that travels with every answer drawn from it. The same applies to a change of area, a reorganization or a long holiday. Knowledge that lives only in someone's memory has a field date too, and it expires on their last day.

Where Cassi.ai comes in

@Cassi.ai is a software engineering company specialized in the pains of market research, innovation and insights, founded by Cassiano Albuquerque. The problem in this article is one clients describe to us in their own words. There is no library with the house's files indexed, and there is no control over what the team is doing with them.

Deskor is a Cassi.ai product that connects company knowledge, AI agents and workflows, so the team builds proposals, decks, reports and dashboards from what the company already knows.

The Deskor is our answer to that. It indexes the house's own archive by passage, answers with the document and page each claim came from, and separates what a client sees from what an administrator sees. The wider argument for why a general chat assistant leaves those gaps open is in why ChatGPT is not a research platform, and the rest sits in the cassiai.com portfolio.

The four field display rule and the shelf life question are how we set up the work with each team. They are a discipline agreed at the start of the project, written into how answers are shown and how new fieldwork is approved. The model does not decide whether an answer is still current. The research director does, with the date in front of her.

FAQ

How old can market research be and still be used?

It depends on the type of finding rather than on a fixed number of years. Stable behaviour ages slowly, attitudes move in jumps after events, and price perception and media habits age fast. Read the field date first, then ask whether anything relevant changed since then.

Should we reuse an old study or commission new fieldwork?

Run internal desk research first. List every study that touched the question with its field date, check permissions, test each finding against its shelf life, and mark each part answered, partly answered or open. Commission fieldwork for the open parts and use it to update the partial ones.

Why does an archive assistant sometimes return an outdated answer?

Partly because of how the search ranks documents. The FRESCO benchmark found that re-rankers favour older, semantically richer documents even when the fact has changed. A long old study can outrank a short recent one unless recency is asked for on purpose.

Who should be able to see past research in a shared archive?

Everyone involved in the conversation has to be authorized for every document used, not only the person asking. Research on participant aware access control argues that output filters are probabilistic, so the check belongs before retrieval rather than after the answer is written.

What is a 360 research study?

It is several methods aimed at the same business objective and read together, such as a survey, focus groups and shopper interviews on one category question. Each claim keeps its study, method and field date attached, so the synthesis does not present evidence from different years as one moment.

The fastest archive and the most useful archive are rarely the same system. The useful one prints the field date where nobody can crop it.

If your team is connecting an assistant to past research this year, take the last five answers it gave and check which of the four fields each one showed. The empty cells are the work.

About the author

Cassiano Albuquerque is the founder of @Cassi.ai. He has spent twenty years in market research operations, from fieldwork and tabulation to the reports that reach steering committees, and now builds AI systems for research teams. To discuss an archive that shows the field date and the permission inside every answer, contact him on LinkedIn or see the cassiai.com portfolio.

Published by Cassi.ai. Read the full article at https://www.cassiai.com/blog/research-archive-old-answers.