From here the subject is marketing, and it is not like the first six days. Marketing has no fixed definitions — ask five successful marketers what positioning means and you get five different answers, each with results behind it. So take the core of what follows, not the letter. These lessons also lean towards product marketing, because that is what you will be doing.
Day 2 said context controls quality. Day 5 had you write a context block so you stopped retyping it. Neither said where that context comes from when the subject is a real product in a real market. It comes from research — and how research gets done changed completely in the last two years.
- Research is where the real-world context comes from.
- Six kinds of research, needed at different stages.
- What research used to cost, and what capped it.
- The shift: agents gather, you judge.
- Where a fact came from decides whether you can act on it.
- Where AI research still goes wrong.
- This is one flow, not the flow.
- Scout — what is actually inside a research agent.
- What this lesson did not cover, and how to go deeper.
1. Research is where the real-world context comes from.
Context comes from several places. A model carries general knowledge. A company has product documentation, analytics, and whatever its team already knows.
What none of those give you is the current state of a real market — what a competitor charges today, what a customer wrote last week, which alternative people actually switched to. That part only comes from going and looking, and it is the part that decides whether your work is specific or generic.
Everyone has the same models now. Everyone can produce a competent post in ninety seconds. The average output of the internet is free and infinite.
So the only thing separating your work from everyone else's is what you know that they do not.
| Input | Comes from | Can a model produce it? |
|---|---|---|
| Fluent writing | The model | Yes, instantly |
| General knowledge of a market | The model | Roughly, and often stale |
| What a customer wrote last Tuesday | Going and looking | No |
| What a competitor charges today | Reading their pricing page | No |
Four of you found this from the other side on Day 4 — the swap test. If a competitor could publish your line word for word, it fails. Research is where unswappable material comes from.
AI removed the cost of producing content and left the cost of knowing something. Research is now the expensive half.
2. Six kinds of research, needed at different stages.
"Do some research" is not an instruction. These are six distinct jobs, and you need different ones at different points in a product's life.
| Kind | Answers | When it matters |
|---|---|---|
| Market | Category, size, direction. What is emerging, what is dying. | Before you build, or before entering a new market |
| Customer | Who has the problem, in their own words. Pains, triggers, objections, vocabulary. | Before positioning |
| Competitor | What else solves this, how they position and price, what customers praise and complain about. | Before positioning |
| Marketing | What content, channels and creators are working in this space. | Before you pick channels |
| GTM | How comparable products are distributed, priced and packaged. | Before launch |
| Opportunity | The synthesis — gaps, hypotheses worth testing, what changed. | Continuously |
Only two of these are your job this week, and it is worth being precise about why.
Market research asks whether a thing is worth building. LotsSocial is built and live, so that question is behind us. Going back to re-answer it now would be interesting and would change nothing you are about to do.
What is genuinely unsettled is who it is for and what it is up against — customer and competitor. Those are the two inputs tomorrow's positioning work cannot be done without.
Research is not a phase you complete. It is six different questions, and the useful skill is knowing which one you are actually being asked.
The other three arrive later: marketing intelligence when we choose channels, GTM before anything launches, and opportunity research never stops — which is why the serious version of it runs on a schedule rather than as a one-off.
3. What research used to cost, and what capped it.
The old shape of this job: open thirty tabs, read competitor sites, scroll review pages, read forum threads, copy quotes into a document, and try to hold it all in your head.
It works. I did it for years. Two things capped it.
- Time. Two hundred reviews is a day. So people read twenty and generalised from those.
- Reach. You read what you could find. Whole sources — forums you did not know existed, a competitor's app reviews, a trend that moved last week — never got looked at.
The result was that research got skipped, or done thinly and treated as though it were thorough. That is still the most common failure in marketing, and it is worth naming plainly: a confident document built on twenty reviews and a hunch.
4. The shift: agents gather, you judge.
Day 6 defined an agent: you give a goal, it chooses its own steps and tools. Research is one of the best fits for that shape, because gathering is exactly the part that is mechanical, slow, and easy to describe.
What actually changed is not that the AI knows more. It is that it can now reach the sources directly — official platform APIs, review data, forum and social search, structured extraction from live pages, and a real browser for pages that block everything else.
So the work splits:
| Gathering | Judging | |
|---|---|---|
| What it is | Finding, fetching, extracting, clustering | What matters, what is trustworthy, what it means |
| Who does it | The agent | You |
| Scales? | Yes, enormously | No, and it should not |
This is the same division from Day 4 — AI reviews first, human reviews second — applied to research. The agent does the reading. You decide what is true and what to do about it.
And the honest warning: an agent gathering two hundred sources you never check is worse than reading twenty yourself. Volume without judgement is how you end up confidently wrong at scale, which is exactly what happened to my blog.
5. Where a fact came from decides whether you can act on it.
This is the most useful idea in today's lesson, and it is what separates a research agent from a chatbot with a search box.
Not all sources are equal. A price read off a search snippet and the same price read off the pricing page are different claims, even when the number matches. So sources get ranked, and you work down the list.
| Tier | What it is | |
|---|---|---|
| 1 | The product's own API | The real record, not a description of it. |
| 2 | The platform's official API | Reddit, Search Console, Analytics, licensed data providers. Where a number you would act on should come from. |
| 3 | Structured web data | Scraping and schema-shaped extraction. Real pages, read directly. |
| 4 | Search | Good for finding where something is. Poor for what it says. |
| 5 | Browser automation | For pages that need a click, a login, or that block everything else. Slow, so it is last. |
Three rules go with it, and they are the whole discipline:
Never present a lower tier as if it were a higher one. Say which one you actually did.
A missing tier is a finding. If the official API is not available, say so and continue with the honest partial answer. Do not quietly substitute a search and present it as equivalent.
Report a tool failure, do not diagnose it. "The page could not be read" is accurate. Inventing the reason sends someone hunting a bug that may not exist.
Apply this to your own work even with no agent involved. Did I read the pricing page, or did I read someone's blog post about their pricing? Both are research. Only one is a fact.
6. Where AI research still goes wrong.
Four failures. The first is the one that ruins everything downstream.
It answers from memory and calls it research
Ask about a market and a model will produce a fluent, organised answer without looking anything up. It reads exactly like research and contains none.
The test: did it actually call a tool this run? If not, nothing was researched. And a date stamp on remembered information — "as of September 2026" — is not a helpful detail, it is a false claim.
It paraphrases the quote
A tidied quote is not evidence.
Real: "I'd already written the post, it was fine. Then I spent half an hour cropping it for LinkedIn and I just gave up."
Tidied: "Users find multi-platform adaptation time-consuming."
The first contains a headline. The second contains nothing.
It names a theme it cannot quote
If it reports a pattern and cannot point at the lines underneath, the pattern is invented. Discard it. Ending your instruction with "if you cannot quote it, do not list it" catches most of this.
It gathers until it runs out of room
A run that keeps searching and never writes delivers nothing. Three well-sourced findings beat twelve gathered and none delivered — and that is a real failure mode, not a hypothetical one.
7. This is one flow, not the flow.
Everything above — the six kinds, the source tiers, agent gathers and human judges — is how we do it at LotsTech. It is not an industry standard and I am not presenting it as one.
There are many other flows. Some teams run everything through one deep-research tool. Some keep it entirely manual and get excellent results because they talk to more customers than anyone else. Some build their own pipelines. All of these work.
So take the shape, not the specifics. And when you find something better — a source we are not using, a step that is wasted, a failure we have not hit — say so. You are the first people outside this company to run this, and I would rather it improved than be defended.
8. Scout — what is actually inside a research agent.
Scout is our research agent. You will use it today, so here is what it is made of rather than a description of what it does.
It covers the six beats from section 2, and it has one hard boundary: it observes, it never acts. No posting, no replying, no outreach. Its output is a finding, not a move — which is what makes it safe to run unattended.
It carries 44 tools, and they map onto the tiers exactly:
| Tier | Tools it holds |
|---|---|
| 1 — own products | Read-only access to our own context notes, so it does not contradict what the team already decided |
| 2 — official APIs | Reddit search and comment retrieval, Hacker News, Google Trends and search-demand data |
| 3 — structured web | Page scraping to markdown, multi-page field extraction, brand lookup, news search, and change monitors |
| 3 — social and forums | X, LinkedIn, TikTok, YouTube, Facebook, Instagram, forum search, news, places and review data |
| 4–5 — search and browser | Web search, and a real browser for pages that block everything else |
Two details worth noticing, because they are the difference between a demo and something usable.
It watches, not just answers. Change monitors track a competitor's pricing page on a schedule and report the diff — so the question becomes "what moved" rather than "what does this say".
It reads the team's existing notes first, read-only. Researching without them produces a brief that contradicts decisions already made. And when a note is stale or unsupported, it is told to say so — "the ICP note claims X; I found no evidence for it and two cases against" is one of the most useful things it can report.
Try it: lotsagent.com/agents/scout
It has never been used by anyone outside this company. Some of it will not work as intended. Tell me what breaks — that is genuinely part of today's task.
9. What this lesson did not cover, and how to go deeper.
Covered: research as the scarce input, six kinds and when each is needed, the manual method and its ceiling, the gather/judge split, source tiers, four AI failure modes, and one real agent.
Not covered:
- Interviews — talking to real people, without leading them. Rob Fitzpatrick's The Mom Test is short and will improve your next conversation immediately.
- Survey design — question order, scales, and how surveys produce confident nonsense.
- Sampling and bias — who you did not hear from, which is usually the group that matters.
- Quantitative research — analytics, cohorts, funnels. Answers "how many", not "why".
- Search demand data — the words people type, at scale. That is the search lesson.
- Ethics and consent — what you may record, store and quote publicly.
Where to go. Read The Mom Test. Then do three real conversations, which will teach you more than any tool in this lesson.
Today’s task — research LotsSocial twice (90 min)
Real work from today. You are researching LotsSocial, a live LotsTech product, from the outside — as a stranger would. Everything you submit is read by me, and work I approve gets used on the product with your name on it. You may say publicly that you did it.
Once the lessons finish you will work inside the product. For now, outside and reporting in.
You are doing the two kinds of research from section 2 — competitor and customer — first by hand, then with an agent. The comparison is the lesson.
1. Understand the product — 15 min, no AI
Read lots.social properly. Write down, in your own words: what it does, who it appears to be for, how it is priced, and what it leads with.
Where the site does not tell you something, write "unclear from the site". That is a finding, and it is the same gap a real buyer hits.
2. Work out what it is up against — 15 min, no AI
I am not giving you a competitor list. Finding it is the work.
Name at least four alternatives, and say how you found each one. Include the ones that are not products — doing it by hand, posting straight into each app, paying someone, using a general AI assistant, doing nothing at all.
Those non-product alternatives are usually the real competition, and they never appear on a comparison page.
3. Find out what customers actually say — 25 min, no AI
Go where people talk about the alternatives you just found.
- Reviews — G2, Capterra, or the app stores. Three-star reviews first. Five stars are enthusiasm, one star is usually a support incident. Three stars explain what actually works and what does not.
- Communities — search the problem, not the product. "Social media takes too long", "posting for clients", "scheduling tool too expensive". Post nothing.
Collect five verbatim quotes — their exact words, with links. Not your summary of what they meant.
4. The agent pass — 15 min
Use Scout at lotsagent.com/agents/scout. If you would rather use ChatGPT or Claude with search, do that and say which.
Give it the same three jobs you just did by hand — the product, the alternatives, what customers say — and ask for a quote and a link behind every claim.
End your instruction with: "If you cannot quote it, do not list it."
5. Audit what it gave you — 10 min
This is the part that teaches. Check honestly:
- Did every link resolve? Open them.
- Were the quotes real, or tidied into cleaner English?
- Did it name any theme it could not quote?
- Did it say where each fact came from — the page itself, or a search result?
- Did it state anything as current that it never actually fetched?
- Did it find an alternative you missed? Did you find one it missed?
6. Write the report — 10 min
One page, four short sections:
- What I found — product, alternatives, and the five quotes, with sources.
- Manual vs agent — what each was better at. Be specific.
- What the agent got wrong — every failure, with its exact output.
- What I would trust it with — and what I would still do myself.
Submit the report here as a document or a message. You need it tomorrow.
Then post one line
In #daily-log: one alternative you found that is not a software product — and one thing the agent got wrong.
Both halves. The first is where positioning actually gets decided, and most people never look for it.
Tomorrow — Day 8: ICP and positioning. You will have evidence. Tomorrow you use it to make the two decisions everything else depends on — who this is for, and why they would choose it over everything else they could do.