A deep competitor analysis an agent can run and you can verify

Some competitor questions do not need a monitoring program. They need one careful answer, soon: before a pricing change, before entering a new segment, before a board meeting where someone will ask how you compare. An agent can do most of the legwork for that kind of deep dive. It reads pricing pages, documentation, terms and filings far faster than a person can. What it cannot do on its own is decide which question matters, which sources count, and when an answer is good enough to put in front of decision-makers. This guide walks through a one-off analysis from framing to delivery, with the agent doing the research and a person keeping it honest.

Frame the decision before the research

“Analyze our competitors” is not a task an agent can finish, because nothing tells it when to stop. Start from the decision the analysis will inform, then derive the questions from it.

decision:     whether to offer a lower-priced plan for small teams
competitors:  Competitor A, Competitor B, Competitor C
dimensions:   entry price and limits, features in the entry plan,
              self-serve or sales-led onboarding, data residency options,
              contract terms for small accounts
out of scope: enterprise plans, roadmap rumours, anything behind a login
deadline:     end of next week
done when:    every dimension has a cell per competitor that is either
              supported by a quoted source or marked as an open question

That last line is the acceptance criterion, and it matters more than it looks. It tells the agent that “not found” is an acceptable cell, so it has no reason to fill gaps with guesses, and it tells the reviewer exactly what to check. Writing it first is the habit described in write acceptance criteria before the agent starts.

Write the research plan, and approve it before it runs

A research plan turns the brief into steps an agent will follow and a person can read in a minute. It has three parts: the question, an ordered list of steps and an allowlist of sources.

1. For each competitor, read the current pricing page and plan comparison; record entry price, limits and included features, with quotes.
2. Read onboarding and signup pages to establish whether the entry plan is self-serve.
3. Read documentation and security or trust pages for data residency options.
4. Read the terms of service, and order forms where they are public, for minimum terms and cancellation.
5. Compile the comparison table; list every cell without a source as an open question.

Approving the plan before execution is the cheapest review you will ever do. It catches a wrong competitor list, a missing dimension or an allowlist with a dubious source in it before any time is spent. If the question changes halfway through, for example because leadership now wants enterprise plans covered as well, revise and approve the plan again rather than letting the agent expand the scope on its own; a changed question needs a new plan explains why.

Choose the sources the agent may use

Much of the quality of a competitor analysis is decided in the allowlist. A reasonable default for this kind of question, from most to least authoritative:

• Competitor-owned pages: pricing, plan comparisons, documentation, changelogs, terms, trust pages. Authoritative for what they offer and claim, not for how well it works.
• Official filings and registries: legal entities, locations, corporate structure.
• Your own material: win and loss notes, call summaries, support tickets that mention a competitor. Often the best source on how a competitor actually behaves in a deal, and often the one agents are never given.
• Independent reviews and community discussions: useful for patterns, weak for single facts. Require corroboration before a claim from here enters the table.
• Press coverage: secondary; prefer the primary source it reports on.

Exclude explicitly what you do not want: pages behind logins you have no right to use, scraped data aggregators of unknown origin, and anything that would require misrepresenting who you are. For a fuller treatment, see deciding which sources may answer a question.

Build the comparison table so every cell can be checked

The table is the deliverable people will actually read, so it has to hold up under scrutiny. Each cell carries a value and the evidence behind it, not just the value.

Three conventions keep a table like this honest. “Not published” is a finding with a source: the agent saw the page and the value is not there. “Open question” means no source was found, which is different, and both differ from a cell nobody checked. Every value links to the passage it rests on, so a reviewer can verify a cell without redoing the research. And inferences, such as “B is nudging small teams toward annual contracts”, go in a separate notes section, labelled as interpretation and tied to the cells they rely on.

Treat open questions as part of the result

An analysis that answers everything is suspicious. Real competitor research always leaves gaps: prices behind a sales call, terms that are not public, behaviour that only shows in a live deal. The list of open questions is not an admission of failure; it is the most useful input for the next step.

For each open question, record what was searched, why it stayed open and how it could be closed: ask a sales rep who met the competitor recently, check a public procurement record, buy a trial, or ask a customer who switched. Keeping questions without a source apart from supported findings lets decision-makers see exactly how much of the picture is evidence and how much is still unknown. Missing evidence is an open question goes deeper on this habit.

Running it as reviewable work

This is the kind of job GoCux is designed to hold. A work order carries the goal, the scope it may touch and the acceptance criteria above, and moves through four visible states: queued, running, awaiting decision and done. While it waits for a decision, nothing downstream begins. Each step the agent takes is written to the work order's record as it happens, so the result, its sources and the decisions made along the way sit in one place.

The research plan follows the same discipline. Creating it records a draft; it runs only after approval, and execution cannot reach beyond what was approved. If the allowlist is empty, the agent may consult no source at all, which is a safe default rather than an open door. Findings supported by a quotation are kept apart from questions where nothing turned up. Note what GoCux does not do: it does not search or crawl the open web itself. Competitor pages come in through an authorized browsing or export step and keep their original page reference and retrieval date once imported. The finished comparison can be produced in Artifact Studio as a versioned deliverable with evidence links, and sending it outside the workspace pauses for a person to approve, edit or reject.

Review before it reaches decision-makers

Even with a good plan, read the result before it is shared. A practical review takes less time than the research did:

• Pick a few cells per competitor and open the cited passage. Does it say what the cell claims, and is it current?
• Look for cells resting on a single secondary source, and either corroborate them or downgrade them to open questions.
• Check retrieval dates against the deadline; pricing pages change, and a stale capture is a wrong answer.
• Read the interpretation section separately from the table. Every inference should name the cells it relies on.
• Check the tone. Describe what competitors offer, factually, and leave out characterizations you would not want quoted back to you.

If the analysis turns out to be something you need every month, it has become a monitoring problem; the guide to automating competitive intelligence covers that shape of work.

Checklist for a one-off analysis

• The decision, competitors, dimensions and out-of-scope items are written down.
• The acceptance criterion says when each cell counts as done, including “open question”.
• The research plan lists ordered steps and an allowlist, and was approved before the agent ran.
• Scope changes went through a revised plan, not silent expansion.
• Every table cell links to a quoted passage with a retrieval date, or is marked as not published or open.
• Interpretations are separated from observations and cite the cells they rely on.
• Open questions list what was searched and how each could be closed.
• A person sampled the evidence and checked the tone before the report was shared.