How it works
If we cannot show you where a claim came from, you never see the claim
Most competitive intelligence tools summarise. A model reads what changed and writes a sentence about it, and that sentence arrives in your inbox with the same confidence whether it is well founded or invented. The failure is quiet, which is the problem: a plausible wrong claim is worse than no claim, because someone repeats it in a deal.
Vaarta is built so that cannot happen. Every synthesised implication has to cite the specific verified signals it was drawn from. If the citations cannot be resolved to real signals, the claim is discarded before delivery. Not flagged, not shown with a warning. Discarded.
Before anything is written: the verification gate
Raw collection is a fire hose. Alerts and mention trackers forward everything that names a company, unverified and unclassified. Vaarta puts every raw candidate through four stages, and a candidate that fails any of them does not become a signal.
Entity resolution
Is this actually your competitor, or a company that shares a word with them? Generic-word names collide constantly, and a name match alone is not evidence. Signals that cannot be attributed to the right company with confidence are dropped here, before any language model sees them.
Relevance
Does it change anything? Most mentions of a company are not events. A candidate with no fetched content and no positive disambiguator does not proceed on the strength of a model’s opinion.
Classification
Every surviving signal is assigned one of 17 event types. One of those is a catch-all, and it is deliberately the hardest to earn: a signal routed there ships only if it clears a materiality bar the specific types are not held to. Anything vague enough to need the catch-all, and unimportant enough to fail that bar, is dropped rather than filed.
Summary
Three lines and an explicit reason it matters, written against the verified source rather than a headline. The source URL travels with the signal so you can always open the thing itself.
Then the citation rule
Verified signals are facts about what happened. The layer above them is analysis: what a set of signals implies, and what you might do about it. That layer is where a language model can be most useful and most dangerous, so it is the layer with the hardest rule.
Each implication must reference the signals it rests on. Those references are checked against the signals that actually exist for your workspace. An implication that cites nothing, or cites something that cannot be found, does not get stored and therefore never reaches a briefing, a digest or the API.
This has thrown away real work
On one analysis run the model returned its citations in a format the validator did not accept. All 34 implications in that run failed the check and were dropped wholesale rather than published ungrounded. The run cost real tokens and produced nothing deliverable. That is the correct outcome, and it is the whole point: the system would rather say nothing than say something it cannot stand behind.
Worth being precise about what failed, because the distinction is the argument. Those claims were not fabrications. They were drawn from real signals, and the citations pointed at real evidence. They were dropped because the validator could not mechanically resolve the citation format, and a citation it cannot resolve is a citation it cannot verify. We fixed the citation parser afterwards. We did not touch the rule that discarded the output.
What this looks like on the page
The rule is only visible if the product shows its work. Every implication is rendered with the verified signals it was drawn from, linked to their sources, so the walk from conclusion to evidence is one click rather than a support request.
Ships
Entry pricing moved from $29 to $49 in eight weeks, which is consistent with a deliberate move upmarket rather than a one-off test.
Grounded in 2 signals
- Entry tier raised from $29 to $49 per month · competitor.com
- Starter plan removed from the pricing page · competitor.com
Both signals exist in your workspace and both carry the source URL they were verified against. Every figure in the claim, $29 and $49, appears in the verified record it was drawn from. Nothing in the sentence is unsupported, so the sentence ships.
Dropped
Entry pricing moved from $29 to $49, and roughly 40% of their customers churned as a result.
The first half is the same well-founded claim, citing the same two signals. The second half contains a figure, 40%, that appears nowhere in your verified record for that competitor. The check is mechanical: every number in a claim has to already appear in the verified signals the analysis pass was given, and this one does not, so the whole claim fails and is discarded. It is not shown to you with a caveat. It is not stored at all.
Being exact about the boundary, because it is the point of this page: that check is deterministic and catches invented figures. It is not a truth oracle. A claim can carry no numbers at all and still be a thin reading of the evidence, which is why the analysis layer is also instructed never to predict, and why every claim ships with its sources attached for you to judge.
You can check it yourself
Grounding is only worth something if you can verify it rather than trust it. Every analysis served by the API carries the ids of the signals underneath it, and each of those ids resolves against the signals endpoint. You can walk from a conclusion back to the evidence without asking us.
# an implication, with the signals it was drawn from curl https://api.vaartahq.com/v1/analysis \ -H "Authorization: Bearer vt_live_..." # then resolve any id in basis_signal_ids curl https://api.vaartahq.com/v1/signals \ -H "Authorization: Bearer vt_live_..."
The same record is reachable over the REST API and a hosted MCP server, so an AI assistant can trace the same path.
What this does not claim
- It does not make the summaries perfect. A grounded claim can still be a weak reading of a real event. Grounding guarantees the evidence exists and is linked, not that the interpretation is the best one available.
- It does not mean we catch everything. A gate tuned to drop unverifiable signals will sometimes drop a real one. We would rather miss an event than invent one, and that trade is deliberate.
- It does not apply to sources we cannot reach. Some things are simply not public, and no amount of verification conjures them.
See what survives the gate for your competitors
Track one competitor free, no card required.