Getting started with CSV matching
A step-by-step walkthrough of your first CSV matching job in Match Data Studio: upload your files, set up the match, run a sample, and download the results.
Matching records across two CSV datasets is one of the most common data tasks — and one of the most frustrating when the data is messy. Names are misspelled, IDs differ between systems, addresses are abbreviated differently.
This guide walks through your first matching job in Match Data Studio from start to finish.
There are two ways to set up a match. When you create a project, Match Data Studio asks you to choose a workflow. Pipeline is the guided, step-by-step flow this guide follows. Canvas is a visual editor where you connect the steps yourself; it takes one or more CSV files, so you are not limited to two, and it adds steps for filtering rows, finding similar items, AI checks, human review and grouping. Both work. Each step below says what changes on the canvas, and the features page lists every step.
Step 1: Prepare your CSVs
You don’t need to do much preprocessing. Match Data Studio works directly with raw CSV files. A few things that help:
- Make sure your files have headers in the first row
- Remove any completely empty rows or columns
- If you have a field that’s a likely unique identifier (like an email or product code), keep it in — it helps the AI configure matching rules
The files can be any size, but for your first run, a dataset of a few hundred rows will give you fast feedback.
Step 2: Create a project and upload
Sign in, click New project, give it a name, and upload your files. The uploader shows progress and confirms row counts on completion.
After uploading, Match Data Studio analyzes the columns and sample rows from both files to understand your data.
On the canvas: choose Canvas, then select one or more CSV files. You can add more files later from the canvas. Each file becomes a CSV Source step, and several files can be stacked into one dataset.
Step 3: Configure with the AI assistant
This is where the magic happens. The AI assistant will ask you a few questions:
- What are you trying to match? (e.g., “customer records from our CRM against a purchased contact list”)
- Which fields should be compared?
- What makes two records a definite match vs. a likely match?
Based on your answers, the assistant configures an eight-stage matching pipeline:
- Data preparation — normalization and cleaning
- Text completions — filling sparse fields (optional)
- Vector embeddings — semantic similarity
- Cosine similarity — pairwise comparison with thresholds
- Numeric matching — for dates, IDs, prices
- String matching — fuzzy algorithms for names and codes
- LLM confirmation — for borderline pairs
- Export — column selection for output
You can review and edit any stage before running.
On the canvas: you build the same ideas as connected steps. Use Transform to clean columns, Row Filter to drop rows that cannot match, AI Extraction and Embedding to prepare the data, Similar Items or Similarity Score to find candidate pairs, an AI check to judge the borderline ones, and Human Review to ask a person about the unsure ones. Click a step to see its settings. Nothing runs until you start a run.
Step 4: Run a sample first
Before processing your full dataset, run a sample (5 credits). This processes up to 100 rows from each file and shows you matched pairs with confidence scores.
Check the results:
- Are the matches correct?
- Is the threshold too tight (missing real matches) or too loose (including false positives)?
Adjust the configuration and run another sample if needed.
On the canvas: choose Run Sample. A sample reads the first 50 rows of each CSV Source by default; you can set a different number (10 to 5,000) or take random rows instead. A card shows the most the run can use before you confirm, and credits that a run does not use are returned.
Step 5: Run the full dataset
Once the sample looks right, click Run full. Cost depends on the number of rows — you’ll see a credit estimate before confirming.
The pipeline runs in the background. You can close the tab and come back — progress is tracked in real time.
On the canvas: Run Full shows the same card: the most the run can use and your balance. After an earlier run of the same setup, it also shows the amount you can expect. You can stop a run at any time; the credits used by finished steps are kept and the rest are returned.
Step 6: Download results
When the run completes, download the results CSV. It contains:
- All matched pairs
- Fields from both datasets side by side
- Similarity scores per field
- An overall confidence score
From here, you can import into your CRM, database, or spreadsheet for review.
On the canvas: the output comes from Export steps. You choose the columns and their order, and each Export step writes its own CSV.
That’s a complete first run. Once it works, the canvas is where you go next: more files, more steps, and people in the loop for the pairs a model should not decide alone.
Questions? Contact us or browse the blog for more guides.
Keep reading
- Entity resolution explained — understand the theory behind linking records across datasets
- Data cleaning before matching — the prep steps that make or break match quality
- Fuzzy matching algorithms explained — Levenshtein, Jaro-Winkler, and when to use each
- How to find and remove duplicates in a CSV file
- How CSV tools corrupt IDs before you ever match them
Ready to match your data?
Create a free account, upload your CSV files, and configure matching in plain English. 100 free credits — no credit card required.
Free account No credit card Match your files in minutes