Published
We now train our own models to read the public web and turn it into signals. Here is what that unlocks.
“We moved our agents to Hermes.”
Resolved to the AI lab, not the scarf.
A signal is only as good as the audience behind it. These are our targets for this quarter.
We build every layer: the collection, the models, and the graph that links them.
A general AI model is too slow and too costly to run on everything, so most signal pipelines read a sample. Ours does not have to. Our collection system gathers the raw pages, and models we trained on our own data read every one.
The first model pulls out the companies, products and people. A second decides which real company each mention belongs to. Because the models are ours, they are fast enough to run on all of it, and every new source makes the next one better.
We read public pages ourselves and keep every one.
Our models find every company, product and person in the text.
Each mention is linked to the right company, with a confidence score.
Mentions become typed events and weekly trends per company.
Our last Reddit feed watched 500 communities. The new one reads the whole site and gives you a weekly trend for every company in it.
Weekly mentions of the AI lab grew from 1,800 to 3,006 in three weeks. Then one day brought 14 posts about switching to its Hermes models from DeepSeek. The normal number is zero.
Our biggest project right now: one identity for each person and company, across every platform.
A handle on X is not a signal until you know who is behind it. We are matching millions of profiles on X, YouTube and other social platforms to the business person who owns each one: LinkedIn URL, work email, job title and company.
Custom models do the matching, trained on links we verified by hand. Every link carries a confidence score. Company accounts get the same treatment, so a channel or a page resolves to one company record and not to a name that looks close.
A handle becomes a person with a job.
Channels and pages tied to one company.
We do not guess. Every link is scored.
The signal you have asked for most. Every record tells you who reacted, how, and to which post.
The most common buying signal there is.
See who engages with a rival's posts.
In your bucket, and as an API signal type.
Send a competitor and your accounts. Get back every LinkedIn interaction between them, scored.
Interactions Search is a new endpoint in the Signal API. You name up to ten competitors and the accounts or job titles you care about. We find each competitor's sellers and leaders in our graph and read their public posts and comments.
Every interaction is matched to the person on the other side. You get one row for each, with both people, both companies and the post. When several point at the same account, we roll them into one pattern and score it from 0 to 100.
{ "competitors": [ { "domain": "servicenow.com" } ], "competitor_contacts": { "departments": ["Sales"], "seniority": ["Director", "Vp", "Cxo"] }, "audience": { "accounts": { "domains": ["acme.com"] }, "seniority": ["Vp", "Cxo"] }, "signal_types": [ "linkedin-post-contact", "linkedin-comments-contact", "linkedin-engagement" ], "since": "2026-11-01T00:00:00Z", "max_age_days": 14, "max_credits": 200}{ "interactions": [ { "description": "VP commented on Acme CIO's post", "interaction_type": "comment_on_post", "actor": { "title": "VP, Enterprise Sales" }, "target": { "title": "CIO", "match": { "confidence": 0.92 } }, "sentiment": { "label": "positive" } }, { "description": "VP of IT celebrated a vendor post", "interaction_type": "reaction_on_post", "reaction_type": "celebrate" } ], "summary": { "threats": [ { "competitor": "servicenow.com", "account": "acme.com", "threat": { "score": 74, "tier": "high" }, "interaction_count": 5 } ] }, "coverage": { "servicenow.com": { "contacts_scanned": 388, "truncated": false } }, "billing": { "interactions_matched": 1340, "credits_charged": 200 }}A sentence a rep can read.
Know how sure we are.
Many signals, one score.
Shown on every response.
Fresher files, more records, and one company record behind all of them.
Most of this release went into the feeds you already receive. Patents and new business formation moved to weekly, so you see events within days. We also repaired earnings calls.
Hiring volume grew by a fifth. And company data is moving to one database, which ends the mismatched company records some of you reported. It is live in the API today and reaches every feed next.
Patent signals moved from monthly to weekly and split into two feeds: one by company, one by inventor. History is backfilled to July 15.
We found and fixed a bug that skipped most earnings calls since March. Every folder from 2026-03-31 is reissued. If you load all folders, re-pull from that date.
Every company now comes from one database of 18 million companies, with HQ location, founded year and confidence scores. It is live in the API today and reaches every feed next, old files included.
Hiring signals grew 20% since September. New business formation moved to weekly, so a new company reaches you within days.
patents-company and patents-contact, delivered weekly and backfilled to July 15. docs ↗patentChallenge and portfolioVelocity. docs ↗unknown. changelog ↗/v1/logs shows what each call charged. LinkedIn reactions, Reddit trends and the Interactions API open to a small group first. Tell us what you want to see.