Competitive intelligence is mostly someone's Tuesday afternoon.
In most companies it works like this. Someone remembers to check three competitors' pricing pages. They paste the changes into a spreadsheet that was last accurate in March. Then a deal is lost to a competitor who moved six weeks ago, and the postmortem concludes that someone should really keep an eye on that.
The tools meant to fix this mostly watch pages for any change at all. That sounds right until you run it: a rotated testimonial, a reflowed layout and a copyright year all fire, while an actual price change buried inside a rebuilt component slips past. People mute the alerts within a month.
Compesight is built on the opposite bet. The hard part was never detecting change. It is being right about which changes matter, and being able to prove where each one came from.
Six decisions we don't relitigate.
These are constraints we've accepted costs for, not values on a wall.
Show the work or say nothing
Every finding links to the snapshot it came from. Every battlecard section names the finding that changed it. A competitive claim nobody can verify is a rumour, and rumours lose arguments in product meetings.
Silence is never success
The easiest bug to ship in this category is a failed crawl that renders as "no changes". It's invisible, it's reassuring, and it's wrong. A source that can't be checked says so, in the data, the API and the UI.
Noise is the real competitor
Nobody churns because a tool missed something. They churn because it cried wolf for six weeks first. We'd rather miss a marginal change than train you to dismiss the notification.
Stay a guest on the open web
We read public pages, honour robots.txt, rate-limit ourselves and back off on failure. There's no override flag, because the first customer who wants one would get our crawler banned for everyone.
Tenancy is structural, not a filter
Crawls are isolated per workspace. Two customers watching the same pricing page get separate schedules and separate history. It costs us duplicate traffic and we pay it, because cross-tenant leakage stops being a bug we have to avoid and becomes a shape that cannot exist.
The model is a component, not the product
An LLM classifies changes and drafts battlecards. Its output is validated against a schema before it touches the database, and it is never asked what it already knows about a company. If monitoring didn't capture it, we don't claim it.
The shortcuts we've turned down.
Most of what makes this category untrustworthy is available and tempting.
Scrape behind logins
Competitors' gated content isn't public, and taking it would make our customers' legal teams right to worry.
Ignore robots.txt, ever
Not per-site, not per-customer, not as a hidden flag. A banned crawler breaks the product for everyone at once.
Let the model fill gaps
Q&A answers from captured findings or declines. It never answers from the model's general knowledge of a company and presents it as monitoring.
Share crawl data across customers
It would cut our costs substantially. It would also mean one query away from a cross-tenant leak, permanently.
Where we are
Compesight is early and deliberately narrow. The pipeline that crawls, extracts, diffs, classifies and routes is running across ten source types, with battlecards, cited Q&A and alert routing on top of it. We'd rather do that well for a small number of teams than add a second product before the first one is trusted.
If you're evaluating us, the most useful thing you can do is point us at a competitor you know well and check whether what we surface matches what you already know. That test is why the free plan exists.