What primary data is, how to collect it, and how to analyze it once AI enters the workflow. Persistent IDs, locked codebooks, and red-flag automation pattern.
Primary data is information you collect yourself, firsthand, for your own question: surveys you field, interviews you run, observations you record, assessments you administer. You choose the questions, the people, and the timing, which is why primary data can prove change in the specific people you serve, something no reused dataset can do.
This page covers examples of primary data by method, its main sources, and the collection practices that decide whether firsthand data survives review. If you are choosing between collecting and reusing data, that decision lives on the primary vs secondary data comparison. If you want the reuse side, with public sources and validation, start with secondary data.
Key takeaways
Common examples of primary data include survey responses from your participants, interview and focus group transcripts, observation notes and checklists, test and assessment scores, and intake records your program captures at enrollment. The common thread is origin: your organization collected it, directly from people or events, for your own question.
Sector examples make the definition concrete. A workforce program's baseline confidence survey and its 90-day employment follow-up are primary data. A school's reading assessments and classroom observation rubrics are primary data. A clinic's patient-reported outcome measures are primary data. A foundation's interviews with grantees, and the open-ended answers in a grantee report form, are primary data too, and usually the least-used evidence the foundation holds.
Primary data comes in both quantitative and qualitative forms, and the strongest examples pair them. Quantitative primary data: scale ratings, test scores, counts, and structured intake fields, anything that arrives as a number you generated. Qualitative primary data: open-ended survey answers, transcripts, field notes, and photographs, anything that arrives as language or observation. A 90-day follow-up that captures an employment status and, in the next field, the participant's own account of what helped is one instrument producing both forms about the same person, which is exactly the pairing analysis needs.
Four characteristics distinguish primary data in any field. It is original: first collection, not reuse. It is purpose-specific: instruments were designed for the question being asked. It is current: as fresh as the last wave you ran. And it is controlled: you know the sampling, the wording, and the conditions, because you set them. Those four are also the source of its burden, since every one of them is a responsibility no agency has already carried for you.
Two classroom variants are worth settling. Is a survey primary data? Yes, when you field it; the same survey becomes secondary data for anyone who reuses your file later. Are your own program's intake records primary data? Yes at collection time; reused three years later for a different question, they function as internal secondary data. Origin is about who collected it and for what purpose, not what the file looks like.
The five main sources of primary data are surveys, interviews, focus groups, observation, and assessments or measurements. Surveys scale to hundreds of people and quantify change; interviews explain mechanisms in a person's own words; focus groups surface group norms and disagreement; observation captures what people do rather than what they report; assessments measure skill or condition directly.
Each method has a failure mode the others cover. Surveys miss the why; interviews do not generalize; observation is expensive per data point; assessments capture ability but not circumstance. That is why strong programs pair one quantitative and one qualitative source on the same people rather than perfecting a single instrument. The full choosing logic, method by method, lives on data collection methods; the paired design is what mixed-mode data collection operationalizes.
Execution decides more than selection. Whatever the method, the same collection disciplines apply: pilot the instrument on a handful of respondents before fielding, because confusing wording is cheap to fix on Tuesday and unfixable in December; schedule waves against the program calendar, so the baseline lands before the intervention rather than two weeks into it; and state the sampling plan in writing, including who will not be reached, so the eventual percentages carry honest denominators. Teams that skip these steps do not collect worse data so much as data whose weaknesses are undocumented, which is what reviewers actually punish.
The one rule that spans all five sources: collect clean at the source. Every response lands on a persistent participant record at the moment of capture, with the field definitions locked before wave one. Sopact's shorthand for the result is the Outcome Thread: intake, exit, and every follow-up connected on one record, so "did confidence change?" is a query, not a matching project.
Plenty of teams collect primary data that cannot survive a hard question. The difference between a pile of responses and evidence is five properties, all set at collection time. One: a persistent ID assigned at first contact, so waves connect without name-matching. Two: a data dictionary that locks each question's wording and scale, so wave two measures what wave one measured. Three: paired quantitative and qualitative capture on the same record, so every number has the words that explain it.
Four: documented sampling, who was asked, who answered, who is missing, so a 72 percent outcome states its denominator. Five: an audit trail from any reported figure back to the responses behind it. Teams that match spreadsheets on names and emails instead of IDs lose 20 to 30 percent of their follow-up links, and every lost link is a participant whose outcome cannot be claimed.
AI raises the stakes on all five. Generative models make two characteristic errors with primary data: numeric hallucination, where a model summarizing thousands of rows invents plausible totals, and codebook drift, where open-ended answers get coded against subtly different categories in each session, so waves stop being comparable. Both are collection-architecture problems. A locked dictionary and computed statistics leave AI doing what it is good at, reading language at scale, with numbers the system calculated deterministically.
A composite illustration shows the properties working. TechBridge, a workforce training program with 80 participants across Chicago, collects primary data at three points: an intake survey with baseline skills, confidence, and zip code; an exit assessment; and a 90-day follow-up on employment and wage. Every response lands on the participant's persistent record, and each open-ended answer sits beside the scores it explains.
Because geography was captured cleanly at intake, the team then joins context from the Chicago Data Portal and Census ACS income tables to ask an equity question the survey alone cannot answer: are we reaching, and placing, participants from the highest-need community areas? The answer, disaggregated by neighborhood income quartile, shows placement rates holding in the lowest quartile, a claim that is only possible because the primary side carried persistent IDs and clean geography. The reuse half of this analysis, with source validation, is worked through on the secondary data page.
Primary data analysis follows the same order of operations whatever the sector. Quantitative fields are computed first: change scores per participant, rates with stated denominators, disaggregation by the groups the program cares about. Qualitative fields are coded against a stable codebook, so themes can be counted and tracked across waves rather than quoted ad hoc. Then the two are read together, per person, which is where the findings that change programs actually live: the participants whose scores improved but whose words describe a barrier about to undo the gain.
The lesson is not the tooling. It is that the analysis was decided at collection time. A version of TechBridge that ran anonymous surveys could compute averages and nothing else: no waves connected, no geography to join, no words beside the numbers.
In the survey-platform era, collection tools optimized for launching forms, and every form created its own island of respondents. Intake lived in one project, exit in another, follow-up in a third, and connecting them meant fragile matching in spreadsheets. Qualitative answers were exported to a document nobody had time to code. The data was primary, in the textbook sense, and still could not answer a longitudinal question.
The record-centric era inverts the model: the participant record is the unit, and every instrument writes to it. The evaluation test for any collection platform: ask to see one participant's intake answer and 90-day follow-up answer on one screen, with the change computed and the participant's own words beside it. Platforms that detour to a dashboard of aggregate charts are form-centric underneath, whatever the marketing says. The record-centric architecture is the same one described on the stakeholder intelligence pillar, and it is what makes the follow-up designs in longitudinal data collection feasible for small teams.
The costliest primary-data mistakes are invisible until it is too late to re-ask: a confusing question discovered at analysis time, a follow-up wave that quietly lost a third of the cohort. That is why firsthand collection belongs inside the Loop, Sopact's method for continuous impact intelligence: collect clean at the source, analyze the moment data arrives, improve while you can still act. A response read the day it lands can fix the instrument for everyone who has not answered yet.
The Loop is also what keeps waves comparable. Same wording, same scale, same denominator rule, every wave, is a discipline with its own chapter in Loop reliability; the trail from any reported figure back to the exact responses behind it is covered in Loop traceability.
One method, three moves that never stop
Then the cycle runs again, a little sharper each time. Read the method: the Loop methodology →
The fastest way to test your own collection is to run it through the five properties. Each prompt below is written to paste into Sopact Sense's Assistant, or to reason through with your team; the arrow above each one links the Academy walkthrough with the expected output and tips.
Academy walkthrough → Clean open-ended survey responses
Here is a sample of our primary data: [PASTE OR ATTACH]. Audit it against the five properties of evidence: persistent IDs, locked definitions, paired quant and qual, documented sampling, audit trail. For each gap, tell me whether the fix belongs in the instrument, the collection process, or the record structure.
Academy walkthrough → How to build a data dictionary
Build a data dictionary for this instrument: [PASTE SURVEY QUESTIONS]. For each field: exact wording, scale, wave schedule, denominator rule, and what would break comparability between waves. Flag questions where wording drift across waves would silently change what we measure.
Academy walkthrough → Analyze pre, mid, and post survey data
Design the primary collection for this program: [PROGRAM DESCRIPTION]. Recommend the waves (intake, exit, 90-day), the 3 outcome questions to hold constant across all waves, one open-ended question per wave that explains the numbers, and the sampling documentation we need so every reported percentage states its denominator.
Academy walkthrough → Connect quantitative and qualitative survey data
Using this cohort's records: [PASTE OR ATTACH], pair each participant's outcome change with their open-ended answers. Show me the participants whose numbers and words disagree, and what the disagreement suggests we ask in the next wave.
Each walkthrough is practical and short: what to do, the prompt to run, the output to expect, and the tips that make it reliable.
Watch: collecting primary data clean at the source, so analysis starts the day responses arrive.
Primary data is information you collect yourself, firsthand, for your own question: surveys, interviews, focus groups, observations, assessments. In Sopact's framing, its value depends on collection architecture: firsthand data on a persistent Outcome Thread can prove change; the same data scattered across disconnected forms cannot.
Examples of primary data include participant survey responses, interview and focus group transcripts, observation notes, test and assessment scores, and intake records captured at enrollment. Sopact's worked example is a workforce program's intake, exit, and 90-day follow-up, all landing on the same participant record.
The main sources of primary data are surveys, interviews, focus groups, observation, and assessments or measurements. Sopact's guidance is to pair one quantitative and one qualitative source on the same people, so numbers and explanations arrive together on the Outcome Thread.
A survey you design and field is primary data. The same survey file reused later, by your team or anyone else, functions as secondary data because the new question differs from the collection purpose. Sopact treats origin, who collected it and why, as the deciding test, not the file format.
Advantages: it fits your exact question, covers your actual participants, and supports change claims. Disadvantages: cost, time, respondent burden, and every quality risk sits with you. Sopact's clean-at-the-source practice addresses the quality half at collection time, when it is still cheap to fix.
Primary data you collect firsthand for your own purpose; secondary data someone else collected and you reuse. The full comparison, with a decision framework and examples, is on Sopact's primary vs secondary data guide; the short answer is that outcomes need primary data and context comes cheapest from secondary.
Through instruments you run: fielded surveys, structured interviews, observation protocols, assessments. Sopact's collection rule is that every instrument writes to a persistent participant record at capture, with definitions locked in a data dictionary before wave one, so waves stay comparable and follow-ups never require name-matching.
AI reads open-ended primary data at scale, but two failure modes recur: numeric hallucination when a model invents totals, and codebook drift when categories shift between sessions. Sopact Sense computes statistics deterministically and holds codebooks constant, so AI does language and the system does arithmetic.
No. A complete, clean cohort of 60 participants with baseline and follow-up on the same records beats thousands of anonymous responses, because change is measurable per person. Sopact's Outcome Thread is what makes small-cohort evidence defensible: every claim traces to named waves with a stated denominator.
Next: weigh collecting against reusing on primary vs secondary data, see the public-data side in secondary data, or choose instruments method by method in data collection methods.
