Every report runs the same pipeline: resolve the companies, gather evidence from three independent source types, score how much evidence was actually collected, then generate the analysis from that evidence under strict validation. This page walks through each stage.
Stage 1: Company resolution
Each input is resolved to a verified company name and domain before any research happens, and the resolution method determines how confident the pipeline is:
- You entered a URL or domain: the homepage is fetched and the company name is read from the site's own metadata. This is the high-confidence path.
- You entered a bare name: a web search for the official site runs first, and the top non-aggregator result becomes the domain. This works for well-known companies but is less reliable, which is why the tool recommends URLs.
- Resolution fails: the company is kept by name only, page research for it is skipped, and the research confidence score drops accordingly.
Stage 2: Page research
For every resolved company, up to four pages are scraped in parallel: the homepage, the about page, the pricing page, and the features page. Pages that do not exist or block scraping are simply recorded as failed; the pipeline never guesses their contents. From each successful page it extracts:
- the page title, meta description, and up to 12 headings,
- up to 6 key points: sentences that carry positioning signal (claims about who the product is for, what it automates, who trusts it),
- the main headline and marketing claims, and a coarse messaging tone (simplicity-focused, power/enterprise, speed-focused, AI-forward, or professional),
- a pricing model classification read from the pricing page (freemium, per-seat, enterprise, free tier, listed on site, or not listed),
- ICP signals (enterprise, SMB, startups, agencies, developers), social proof claims ("trusted by 2,000+ teams"), and technology hints detected from the page source.
Stage 3: Market signals
In parallel with page scraping, four categories of web searches run for every company, restricted to results from the past year: funding and acquisitions, hiring and growth, product launches and partnerships, and pricing and positioning coverage. Up to five results per category are kept. These signals feed the analysis and become the Sources list attached to your report, so you can check the underlying articles yourself.
Stage 4: Review sentiment (optional)
If you provided a review URL, the page is scraped and mined for sentiment: up to 8 positive mentions, 8 complaints, and 5 churn signals (phrases like "switched to", "cancelled", "no longer use"). If the review site blocks direct scraping, a search fallback tries to recover review content from the same domain. If nothing could be read, the report says the review data was not scraped and the review sentiment section is labeled low confidence instead of being silently invented.
The research confidence score
Every report carries a 0 to 100 research confidence score computed from what the research phase actually managed to collect:
| Signal | Weight | What earns it |
|---|---|---|
| Page coverage | up to 40 points | Share of the possible company pages that were successfully scraped. |
| Company resolution | up to 30 points | Average resolution confidence across all companies. |
| Review data | 15 points | Awarded when review sentiment was actually scraped. |
| Competitor count | up to 15 points | 5 points per competitor analyzed, up to three. |
A score in the 80s means the analysis is grounded in nearly complete page data; a score in the 40s means several companies resisted research and more of the analysis leans on the model's category knowledge, clearly labeled as such.
Stage 5: Analysis and validation
The full research payload (pages, signals, reviews, and your context note) is passed to a single large model call with rules that shape the output you see: every insight must tie back to specific evidence from the research data, fabricating data points or URLs is forbidden, and thin research must lower the stated confidence rather than be papered over. The scraped content itself is fenced off as untrusted data, so text on a competitor's page cannot redirect the analysis.
The output is then validated against a strict schema covering all eleven sections, down to the allowed values of every status and confidence field. If the output is malformed or incomplete, the model is re-prompted once with the exact validation error. If it fails again, the affected sections are marked failed and the report is delivered as partial or failed, with a retry available; see Sharing, retries, and limits.
What the tool does not do
- It does not log into anything: only publicly reachable pages are read, and requests to private or internal addresses are blocked outright.
- It does not present guesses as findings. Missing pricing pages produce estimates labeled as estimates; missing review data produces low-confidence sections that say so.
- It does not reuse one customer's research for another: every run researches live, and cached results are only ever returned for your own identical inputs within 24 hours.