The AI

Once a diff is stored, RivalSweep classifies it into a category (pricing, product, integration, positioning, content, or noise) and an importance (low, medium or high), and writes a short title and summary of the change. On Team and Scale plans, high-importance events also get a short “possible impact” note — one or two sentences on what the change might mean competitively.

• It never decides whether a page is allowed to be crawled — that's governed separately by robots.txt, blocklists and the SSRF check, before any AI is involved.
• It never replaces the evidence. The stored before/after diff is the fact; the AI only summarizes and labels it.
• A possible impactnote is always presented as inference, never as observed fact — you'll see it labeled that way everywhere it appears, alongside a confidence figure.

AI classification isn't a Team/Scale-only feature — every plan, Starter included, has its changes read by an AI model while the workspace's monthly AI budget lasts:

• All plans get AI category and importance labeling, a factual summary, and a one-line possible-impact note for every detected change, metered against the plan's monthly AI budget.
• Team and Scale additionally get a second pass by a more capable model on high-importance events: a fuller possible-impact write-up on what the change might mean competitively.

If the AI provider is ever unavailable, or the workspace's monthly budget (plus any purchased credit) is used up, RivalSweep automatically falls back to a fast offline heuristic for that event. A crawl or a change never gets lost or delayed because of an AI outage.

Which models: classification runs on DeepSeek V4 Flash, and the deeper write-ups — the Team and Scale impact pass, the weekly analyst note, competitor discovery and Ask — run on Anthropic's Claude. Both are reached through OpenCode Zen, a US-hosted gateway, under zero-retention terms: your data is never used to train anyone's model. We may change models as better ones appear; the current list is always in our privacy policy.

If you dismiss the last three reviewed changes on a page in a row, RivalSweep takes the hint: the next classification on that page is automatically downgraded to noise, the same way a human reviewer would have treated it. It never applies to pricing changes — repetition alone is never reason enough to stop watching a price page — and it's entirely local to that one page in your workspace, never shared across competitors or customers.

“Not sure who to watch?” works two ways, and they cost differently.

• From your domain — we read your homepage and derive candidate phrases from it offline. While your AI budget lasts, a Claude pass over that same page text proposes sharper phrases and a product-category guess. If the budget is exhausted or the AI is unavailable, the offline heuristic keeps this path working exactly as before.
• From a description you write— there is no page to read here, so this path is the Claude pass, and it draws on your workspace's AI budget. If it can't run we tell you which of two things happened: your monthly allowance is used up — with a link to top up AI credit — or something failed on our side, which is ours to fix and never costs you anything. Either way we ask you for the search phrases directly rather than guessing at them, and the competitor lookup that follows works the same.

Every workspace has a monthly AI budget sized to its plan. You can watch usage against that budget from Settings at any time.

Automated change detection — crawling, diffing, and the offline heuristic classification — keeps running no matter what. What pauses when the monthly budget (and any purchased credit) is used up is the richer AI summarization and impact analysis; it resumes automatically at the start of the next month.

If you need more headroom before the month resets, the workspace owner can purchase an AI credit pack from Settings. Purchased credit never expires and is drawn down after the monthly budget is exhausted.