What Research Automation Actually Replaces
Outcome: you leave this lesson with a written list of the research questions worth automating for your own motion, and the order to run them in.
- Surface
- App and MCP server
- Level
- Beginner
- Uses
- No actions — planning lesson
- Credits
- 0
- Prerequisite
- None
The work being replaced
Open a prospect’s website. Read the homepage. Check the pricing page. Look at careers for hiring signals. Skim recent posts. Decide whether they are worth contacting. Write one sentence proving you did all that.
Six minutes per account, done well. At 200 accounts that is twenty hours, which is why in practice it gets done for the first fifteen and skipped for the rest — and the skipped ones get a template.
Research automation is worth it exactly where that pattern shows up: a repeatable question, asked of every row, whose answer changes what you do next.
The test
Before automating any research question, check all three:
Is the answer on the public web?
If it only exists inside someone’s CRM or in a conversation, no amount of scraping finds it. Automation cannot manufacture information.
Does the answer change your behavior?
If you would send the same message either way, the answer is trivia. Skip it — it costs credits and buys nothing.
Is the question the same for every row?
“What is their pricing model” generalizes. “Why did their CTO leave in 2023” does not. Per-row bespoke questions stay manual.
Anything failing one of the three stays a human job. Most teams find three to five questions that pass, and those three to five cover the majority of the value.
Scrape or ask an agent
Both are in this course, and they fail differently.
| Scraping | Research agent | |
|---|---|---|
| You give it | A URL and the fields to pull | A question |
| Good for | Facts in a predictable place — contact details, pricing tiers, job listings | Open questions where the answer could be anywhere |
| Fails by | Returning nothing when the layout changes | Returning something confident and wrong |
| Cost | Lower | Higher |
| Check by | Spot-checking empties | Spot-checking answers |
The failure modes matter more than the capabilities. A scraper that breaks is loud — the column goes empty and you notice. An agent that guesses is quiet — the column fills with plausible text and nobody checks. Lesson 06 covers prompts that make agent failures loud too.
Rule of thumb: if you can name the page the answer lives on, scrape it. If you would have to search for it, ask an agent.
Why research runs before enrichment
Research is cheaper per row than contact enrichment, and it disqualifies rows.
Run in the wrong order, you buy verified emails for 500 companies and then discover 300 of them do not fit. Run in the right order, research cuts the list to 200 and you buy 200 emails.
Find
Get candidate rows in. Cheapest stage.
Research
Ask the disqualifying question. Filter on the answer.
Enrich
Buy contact data only for survivors.
Outreach
Generate copy only for rows you will actually send to.
The savings compound with list size, which is why this ordering is repeated in every course in this library.
Do this now
Write down, for your own motion:
- The three to five questions that pass the test above
- For each one — scrape or agent?
- Which single question disqualifies the most accounts
That last one is your filter, and it goes first in every table you build. Lessons 03 through 07 build the columns; this list decides what they should contain.
Check your work
- Can you name a research question you were about to automate that fails the “changes your behavior” test?
- For each question on your list, do you know whether the answer lives on a page you can name?
- Which question, if answered “no”, means you stop working the account?
Where this breaks
The expensive mistake in this course is running research on the full list before checking the output shape on a sample. Research columns are the second most costly thing in a table after generated copy. Run any new research column on five rows, read all five answers yourself, and only then widen it — a prompt that returns a paragraph where you wanted a yes or no will do that 500 times just as happily as five.
Further automation
Once a research question proves itself, it becomes a reusable column in a Template, and — if you run it from an AI client rather than a table — a skill file you can install from GTM agents .
Next lesson
02 — Scraping basics and limits, which covers what a scraper can reach, what it cannot, and the obligations that come with pointing one at somebody else’s site.
Reference for this lesson: Web Scrapers and AI Agents.