Every vendor that sells “intent data” is quietly doing one of three things: aggregating bid-stream data, licensing keyword searches, or running a pixel network. The quality of the signal is a direct function of which one. Most buyers never ask which.
I am George, founder of Leadpipe. Orbit is our pixel network. It sits across 5 million websites, observes real browsing behavior with consent, and feeds 60 billion behavioral signals into a person-level intent layer that refreshes every 24 hours. This post is about how a cross-site pixel network actually works, the principles behind it, the tradeoffs that come with the architecture, and why we built it instead of reselling someone else’s feed.
What cross-site intent data really is
Intent data, at its core, is the answer to one question. Which people are actively researching a topic right now, even on sites that are not yours?
To answer that, you need two things.
- Cross-site observation. You have to see a person on more than one website. Otherwise you only know what they did on yours.
- Identity resolution. You have to tie those observations to a real person. Otherwise you have a bag of anonymous sessions.
A pixel network gives you the first. An identity graph gives you the second. The combination is what makes person-level intent data possible.
BROWSER SESSION ORBIT PIXEL IDENTITY JOIN
───────────────── ────────────── ─────────────
visit site A ─┐ person: Sarah Chen
visit site B ─┼─> events ingested ─> clustered ─> topics: CRM,
visit site C ─┘ sales-automation
score: 78/100The pixel provides the observation. The graph turns observations into a named person with a topic profile. That is the whole game.
The verified scale
The numbers are not marketing language. They are the operational scale Orbit runs at:
| Metric | Current |
|---|---|
| Websites contributing to the network | 5M |
| Behavioral signals ingested | 60B |
| Verified person profiles in the underlying graph | 280M |
| Topic taxonomy | 20,000+ topics |
| Refresh cycle | 24 hours |
The 5M number is worth pausing on. A pixel network that sits on partnered sites with consent, at this scale, is rare. Most “intent” vendors are not running a network of this size. They are aggregating a bid-stream firehose or licensing third-party data.
What Orbit actually observes
The Orbit pixel is lightweight. It records what page was viewed, when, and enough context to place it in a topic taxonomy. It does not record the content of forms, keystrokes, or anything the visitor types.
| Signal | Captured | Not captured |
|---|---|---|
| Page URL | Yes (for topic classification) | Form content, DOM contents |
| Referrer | Yes | Content after login walls |
| Session timing | Yes | Mouse tracking, keystrokes |
| Anonymous browser identifiers | Yes (cookies, device IDs) | PII in cleartext |
| Geography (city/state/country) | Yes | Exact coordinates |
Every event is passed through a topic classifier that maps URL patterns and page semantics to entries in the 20,000+ topic taxonomy. The output is not “visited a page.” It is “showed interest in cloud cost optimization.” For the parsing engineering side, see parsing JavaScript-rendered intent signals.
The five-stage aggregation pipeline
Raw page views are not intent signals. Intent is a derivative. The pattern of repeated, topic-aligned visits by the same person across multiple sites over a short window.
PIXEL EVENT -> TOPIC CLASSIFY -> CLUSTER -> SCORE -> SERVE- Pixel event. Lightweight POST from the browser to the Orbit ingestion endpoint. Throughput-dominated, has to absorb spikes without dropping primary events.
- Topic classify. Each event is mapped to zero, one, or many topics. Pages that do not classify with confidence are dropped rather than guessed.
- Cluster. Events are grouped by anonymous browser session and, via the identity graph, attached to a person if the link can be verified.
- Score. Compute an intent score (0-100) per person per topic, using recency, frequency, and cross-site breadth.
- Serve. The person’s topic profile is exposed through the Intent API, the MCP server, and the dashboard, updated daily.
The expensive stage is the cluster step. Without the identity graph, you are stuck at session-level intent, which nobody can act on.
Why a pixel network beats bid-stream
The common alternative to a pixel network is bid-stream intent data. A vendor buys a firehose of ad auction requests from exchanges, extracts topic signals from the URLs in those requests, and sells the output. The honest comparison:
| Signal source | What you observe | Noise level | Identity join |
|---|---|---|---|
| Pixel network (Orbit) | Direct page views on partner sites | Low, filtered at the source | First-party, deterministic |
| Bid-stream | Ad requests, which may or may not render | High, includes bots and nonviewable impressions | Hard, often probabilistic |
| Keyword search licenses | Search queries | Medium, depends on the licensor | Hard, search engines strip identifiers |
A bid-stream signal tells you an ad slot existed on a page. It does not tell you the page was read, a person engaged, or anyone real was there. Pixel-network signal tells you a browser loaded a page and spent time on it. That is a cleaner truth.
The other industry standard is Bombora-style co-op aggregation, which depends on a publisher consortium contributing data into a shared pool. That works at the company level but does not produce person-level intent at scale. For the head-to-head, see Orbit vs Bombora.
The identity join
Without identity resolution, Orbit would be a clever ad network. With it, Orbit becomes person-level intent data.
The join works like this. Every page view from the Orbit pixel is tied to a browser session. That session carries anonymous identifiers (a cookie, a device fingerprint, an IP range). The identity graph maps those identifiers to a verified person when the evidence is deterministic.
If the evidence is weak, the intent stays session-level and never gets exposed as “Sarah Chen is researching CRM software.” We would rather under-report than mislabel. That policy is the same one that produces the 8.7/10 accuracy result on the visitor identification side.
Scoring and decay
Intent without decay is just history. Someone researching CRM software eight weeks ago is not necessarily researching today. Scoring has to weight recency heavily.
The intent score blends:
- Recency. Today’s visits count more than last week’s.
- Frequency. Three visits across three days beat three visits in one hour.
- Cross-site breadth. Visiting three different sites on a topic is stronger than visiting one site three times.
- Topic affinity. Visits to canonical category pages count more than tangential pages.
The output is a 0-100 score per topic per person, refreshed every 24 hours. The manifesto on daily truth is the principle behind the refresh cadence. Static lists decay the moment they are exported, so the graph has to be updated daily.
Tradeoffs we made
Partner-curated sites over open web
Orbit sits on partnered websites, not on the open web. That limits the breadth of sites we observe, but it gives us consented, first-party signal instead of scraped approximations. The compliance posture is materially better, and the data quality is more consistent.
Topic taxonomy over free-form
We classify into a finite taxonomy of 20,000+ topics rather than extracting free-form topic tags per page. A curated taxonomy is easier to reason about, easier to filter, and easier to make reliable. The cost is that very new or niche topics have to be added manually as they emerge.
Person-level over account-level
The category standard is account-level intent. “Acme Corp is researching your category.” Orbit serves person-level. “Sarah Chen at Acme Corp is researching your category.” Person-level is harder to produce and harder to price, but it is what makes outbound actually work. See our intent data only shows companies for why account-level alone leaves you stuck.
Daily refresh over streaming
Intent scores refresh daily, not continuously. A streaming score would look impressive on a dashboard, but sales teams cannot act on every flicker. Daily is the cadence at which the signal is actually useful.
Built network over licensed feed
Orbit is built, not licensed. The same logic as the underlying identity graph. Licensed feeds give you parity with everyone else who licensed them. A built network compounds.
Why we built it instead of reselling
Three reasons that did not change between year one and year five.
The signal source determines the ceiling. A reseller of bid-stream or co-op data inherits the upstream quality cap. Building the network meant we set our own cap, which is higher than what the licenses we evaluated were producing.
Person-level requires the identity graph. Most third-party intent feeds are account-level by design because their identity model does not extend to persons. Without our own graph, person-level intent would have been impossible regardless of the pixel choice.
Compliance posture. A first-party pixel network with consented partner sites has a different compliance shape than an aggregated co-op feed. The DPA and subprocessor list look cleaner because the data lineage is shorter and we control more of it.
What I would prioritize today
- Hot topics as first-class. A handful of topics spike unpredictably (product launches, news cycles, outages). Those deserve a faster refresh than 24 hours. Sub-daily for the spike-prone categories, daily for the steady ones.
- Cross-network reconciliation. When a person’s browser is seen on Orbit and on a customer’s pixel, we reconcile them. A cleaner data model for “Orbit saw this person off-site” plus “your site saw this person on-site” would make the end-to-end journey easier to query.
- Explainable intent. A 78/100 score is opaque. An explanation like “3 visits to category pages across 2 competitor sites in the last 5 days” is actionable. The underlying data exists. Exposing it cleanly is a product decision worth prioritizing.
What this means for customers
Orbit is how we answer “who is researching your category” before a visitor ever lands on your site. The output flows into Orbit audiences, competitive intelligence, LinkedIn ads targeting, and the daily watchlists sales teams use to run outbound.
The practical implication: you stop running outbound against strangers and start running it against the small slice of your ICP who are actively shopping. That is what the pixel network buys you, and it is why we built it instead of reselling someone else’s.
