Competitor deep dive — one rival, every dimension

The question this skill answers

What has this competitor actually done, what did the platform measure, and what does it change for us?

One rival, all six dimensions filtered to them, plus the Strategic Briefing's researched read on that same rival. The output is a dossier a founder, PMM or sales leader acts on — not a landscape, and not something anyone reads mid-call.

The platform already did the research

The briefing's competitors section carries a per-rival read, and its deep- sections cover funding, hiring, product launches, reliability, customer voice and the developer ecosystem for every monitored competitor — probed across more sources than a search returns, kept with history to difference against, and shipped with their own stated limits.

Read those sections. Do not re-derive them. Searching the same ground gives a thinner answer from fewer sources with nothing to compare it against. Web research here fills a gap the platform does not cover, and it is the exception, not a step.

Before you start

Read references/reading-the-data.md. Four of its rules carry most of a dossier: null is not zero, counts never rates, per engine never pooled, and model prose is the model's. The fifth is dates — every figure here came out of a run, and a dossier that does not say when it was measured is a snapshot posing as a trend.

Steps

1. Resolve the competitor. mcp__competlab__list_projects → mcp__competlab__list_competitors to match the name or domain (the customer's own row is marked isOwn: true) → mcp__competlab__get_competitor for the pages actually monitored for this rival → mcp__competlab__get_project for when each dimension last ran.

A page nobody monitors is not a page that does not exist, and a dimension that has never run is not a finding about the competitor. If the rival is not monitored at all, say so and offer to add them: anything assembled without monitoring data is unverified web research and has to be labelled as that.

2. Read the briefing's read on them. mcp__competlab__get_briefing with sections: ["competitors"] plus whichever deep- sections this dossier needs — deep-funding-capital, deep-hiring-gtm, deep-product-launches, deep-reliability-status, deep-customer-voice, deep-ai-ecosystem, deep-agent-readiness, deep-landscape. The response's contains array says what the edition actually holds.

Check meta.status first. On running or failed the item is null but an earlier edition is usually still readable — mcp__competlab__get_briefing_history, then mcp__competlab__get_briefing_edition. A null item never means the project has no briefing.

3. The four monitored surfaces, filtered to this rival.

• Positioning — mcp__competlab__get_positioning_dashboard for the headline, what it calls itself and

who it says it is for, as its page words them (this rival's entry in item.companies[], its claims verbatim); mcp__competlab__get_positioning_history for whether the message has moved, and when. competlab-positioning says how to read both.

• Pricing — mcp__competlab__get_pricing_dashboard: this rival's entry in item.companies[] (its

plans as a US buyer sees them, in the page's own words and currency), its rows on item.setups[] (the typical buyer setups, cheapest first — report its rank of of, never a rank of your own), and its lines in item.changes[], each quoted as the check wrote it, from X to Y with the two dates and its level in the customer's words. mcp__competlab__get_pricing_history for the checks before. A company whose page could not be read shows its last read's facts, dated shownAsReadAt — say the date, and never read it as a company with no prices. No average and no price position: the dossier states what they charge and what changed.

• Content — mcp__competlab__get_content_dashboard for volume, categories and gaps;

mcp__competlab__get_content_changelog for the URLs added and removed since the previous run. What a competitor started publishing is often the earliest visible signal of what they are building.

• Tech & trust — mcp__competlab__get_tech_trust_dashboard. A measured 0 trust signals is a

finding. A null is a scan that could not read the site. Two different sentences — never one that could be either.

4. AI Visibility — membership first, never the score. mcp__competlab__get_ai_visibility_dashboard. Read summary.promptMarket before anything else: unless its state is rivals_named_in_most_answers, say the prompts may not describe this project's market, and do not lead with the map.

Then summary.marketMap — is this rival among the companies the models name in this category, and where does the customer sit relative to them? Presence with its range and its count, ordered by presence and by nothing else. Two brands whose ranges overlap are tied. Do not order them, and do not turn the overlap into a story.

The map arrives one page at a time, and every page carries the customer's row and every tracked competitor's. A monitored rival missing from it was named in no answer — say not named in any answer, never a place and never not measured. mcp__competlab__get_ai_visibility_trend gives its count either way, 0 of N included, and its now is the latest map.

The call worth knowing: mcp__competlab__get_ai_visibility_check_detail with includeAnswers: true and brand=<their domain> returns every answer of one check filtered to that one brand, and leaves the summary out. Take the latest check's id from mcp__competlab__get_ai_visibility_history with limit: 1. It is the cheapest way to answer where do they beat us, and where are they invisible — on a large market it still runs to tens of thousands of characters (references/platform.md § Answer sizes). An answer that comes back with an empty brands list is an answer the model gave without naming them — a real finding, and a different one from a question that produced no answer at all.

The trend's rows are the customer, every tracked competitor and up to 3 untracked companies: a reading now (pooling its checksAnalysed checks, never the latest check alone), a reading at the start of the window, and whether the two are separable. Without `detail: "series"` it is two readings, not a time series. Call the difference a rise or a fall only when presenceChangeSeparable is true; where the intervals overlap, it is two readings, not a movement.

5. AI Sources — the pages behind those answers. mcp__competlab__get_ai_sources_dashboard. Perplexity and Google AI Overviews only: they are the two engines that hand back the pages they retrieved. Per engine, never pooled. Retrieved, never cited — the engines do not disclose which pages they leaned on.

In a rival's dossier the rows that matter in summary.coreHosts are the ones this competitor owns (ownership: "competitor_owned") — hosts more than one engine reads for this market that belong to them and cannot be pitched. Count them, name them, and say plainly that they are not approachable. The customer's own work list belongs to competlab-ai-sources, not to this dossier.

6. Fill a gap, if one is left. WebSearch only for what no dimension and no briefing section covers. What it returns is a claim, not a fact: verify it with mcp__competlab__fetch_url (cleanHtml: true) before it reaches the dossier, and drop what does not verify rather than softening it into a hedge.

7. SWOT, on evidence. Every item names the measurement behind it and the run or edition it came from. An item that cannot be sourced that way does not go in — "strong brand" is not a finding, and neither is a number nobody can trace. The evidence has one order of strength: what the platform measured — timestamped, and comparable against the run before it — carries more than what a search returned, and a search result stays a claim until the fetch in step 6 confirms it. Opportunities are what the customer can act on. Threats are tied to something already observed, not to what a rival might conceivably do.

Output

# {Competitor} — dossier
*{Project}. Dimension runs: {dimension} {date}, … · Briefing edition {n}, {date}.*

## Where they sit
{3–5 sentences: who they are, what they are betting on, where they are exposed. The paragraph
someone reads before deciding anything.}

## Positioning · Pricing · Content · Tech & trust
{a short block each — what the run says, what changed since the one before it, what is unmeasured}

## AI Visibility
{Membership: is this rival among the companies the models name, and where is the customer relative to
them. Counts with their universe and their range. Ties marked as ties. Per engine.}

## AI Sources
{Per engine: which companies the answers named, and which core hosts this rival owns.}

## What the briefing already researched
{Funding, hiring, launches, reliability, customer voice, ecosystem — from the sections, naming the
edition. Not re-derived.}

## SWOT
{Strengths · Weaknesses · Opportunities for us · Threats from them — each item carrying its evidence.}

## What we did not measure
{Dimensions that have never run, scans that could not read the site, engines not asked, briefing
sections this edition does not contain.}

## Recommended response
{3–5 prioritised actions, each tied to something above.}

What NOT to do

• Do not re-research funding, hiring, launches, reliability, customer voice or the developer ecosystem

with a search. The briefing did it, across more sources, with history.

• Do not lead with the AI Visibility Score, a mention rate, a position or a citation count.
• Do not read a trajectory or a rate of change off the trend tool. It holds two readings, and their

difference is a rise or a fall only when presenceChangeSeparable is true.

• Do not order two brands whose ranges overlap, or call an overlapping difference a rise or a fall.
• Do not pool engines, and do not report a figure without the universe it came out of.
• Do not treat a null as a zero, or a page that could not be read as a page the brand is absent from.
• Do not present model prose as fact — "Claude described them as…", never "they are…".
• Do not put an unverified web claim into a dossier someone will quote.
• Do not speculate about private metrics — revenue, burn, churn — unless they were publicly reported.
• Do not write three thousand words where fifteen hundred covers it.

Decision questions

End with 3–5 questions whose answers would change the recommendation, tied to what this dossier actually found. State both branches where you can:

"Their homepage moved toward enterprise buyers two runs ago and their pricing has not followed yet. If it does not, the opening is in the segment they are leaving. If it follows next quarter, that opening closes. Is this worth watching monthly, or acting on now?"