an ai market research bot that shows its sources

the routine

give your researcher its own browser and a written question. before scheduling, send the list of permitted sources in the same chat and ask for one trial brief.

every other monday at 10, revisit this question using the sources list i gave you: what problems do small architecture studios describe when sharing project files with clients?

use your own browser. for each source record the page title, url, publication date if visible, and date you read it. bring back a short passage supporting each finding. separate a company’s claim, a person’s reported experience and your inference.

group repeated problems, but do not turn mentions into a market-size estimate. include contrary evidence and a list of pages you could not read. finish with questions i would need to ask customers. do not sign up, contact anyone or buy reports.

when

every other monday at 10:00, after a trial run with your chosen sources

needs

a bot with its own browser, a research question and a source list; no connected app required

what comes back looks like this (an example, not a real inbox)

research note for mira: client file handovers

finding: version confusion deserves an interview question.
evidence: a forum post by an architect describes sending a client the wrong drawing revision. passage: "we sent the previous drawing set." source: page title, url, read 12 november 2026.
interpretation: one reported experience, not evidence of how common the problem is.
contrary evidence: a studio blog post says a shared naming rule fixed this for them.
unknown: willingness to pay; neither source discusses budget.
next question: what happened the last time a client opened the wrong version?

choose a question you can investigate

“research the construction market” leaves too much room for a smooth answer without evidence. choose a buyer, a problem and a decision. tell the bot whether you are deciding whom to interview, what to build or which message to test.

start with sources you trust enough to inspect yourself. ask it to distinguish the original source from an article repeating someone else’s claim. a cluster of articles quoting one report is still one underlying piece of evidence.

keep a trail from claim to page

the browser is separate from yours. use Screen to watch and Take control when a page needs a sign-in or a two-step code. hand control back afterwards. that helps with access; it does not establish that the page is accurate.¹

request short supporting passages, not copied articles. a link without the relevant passage makes review harder. record missing publication dates as missing. for recurring changes to a known set of businesses, the competitor monitoring routine is a narrower job.

what the report cannot establish

public pages cannot establish private revenue, customer satisfaction across a whole market or demand for an idea nobody has tested. a company’s own claim belongs in the report as a claim. do not present an inaccessible report as evidence just because its title sounds useful.

use lead research to build a list for your next interviews. use web scraping when you already know the pages and only need specified fields. neither turns missing evidence into a finding.

the guard rails

• require readable evidence for each finding; drop claims supported only by a page title
• keep facts, reported experiences and inferences in separate lines
• stop at access restrictions and report the gap rather than working around them
• review contrary evidence before using a brief to make a spending decision