How to map clickstream to purchase intent without leaking PII
Turn clickstream into privacy-safe purchase intent in three checks. First, strip personal data from page URLs and referrers at intake. Second, map multi-page paths to standard topic categories. Third, filter automated traffic using dwell time and scroll depth. Test all three on a sample file before any scoring begins.
How to use this article
Read the checklist here, then use the linked hub and product pages for procurement citations.
A clickstream feed can show what a company or a cohort is researching before anyone fills out a form. It can also carry personally identifiable information (PII), such as email addresses, inside web addresses, plus thousands of clicks that no person made. This guide covers the order of operations: clean the file first, read paths second, score last.
What is clickstream data, and why does it need cleaning?
Clickstream data is a log of page visits, kept by device or by session. Each event records which page was viewed, which page sent the visitor there and when it happened. A session ID (a code that groups one visit) is usually pseudonymous, and some feeds also carry a mobile advertising ID (MAID), a device code issued by the phone's operating system. Pseudonymous means the code is not a name or an email address. Article 4 of the GDPR, the European Union's General Data Protection Regulation, defines pseudonymisation as processing that keeps the extra identifying information separate, so the code alone cannot point to a person.
A pseudonymous code does not make the rest of the row safe. The page address (URL, short for uniform resource locator) and the referrer address can both carry personal data. Some visits come from scripts, not people. Clean both problems before a scoring model sees the file, because a model trained on dirty rows learns the dirt.
How do you strip PII from URLs without losing buying signals?
PII leaks into logs through query strings, the part of a web address after the question mark. The MITRE catalog of software weaknesses lists this pattern as CWE-598, use of an HTTP request with a sensitive query string. The Open Worldwide Application Security Project (OWASP) describes the same exposure in its page on information exposure through query strings. It names the referrer header and web logs as places the values end up.
The web's core protocol specification says the same thing in plain terms. As of November 16, 2026, RFC 9110 still states that addresses are meant to be shared, not secured, and that servers, proxies and browsers often log them where third parties can see them. Anything sensitive or personally identifiable placed in an address is a risk.
Scan a sample for three patterns. Look for the at sign, for its encoded form (percent sign, then 40), and for parameter names such as token or auth. Do not tell a vendor to delete everything after the question mark. Site-search terms often sit there, and they show what a visitor wants to buy. The goal is to remove identifiers and keep the intent.
The referrer deserves the same scrutiny. The W3C (World Wide Web Consortium) Referrer Policy exists because the address of the previous page can reveal who a visitor is, and it lets a site trim that address down to its origin. Ask whether the feed keeps full referrer paths, trimmed ones or both, and what was removed before delivery.
How do referrer paths show purchase intent?
A single pageview says little. A path says more. Order the events in a session by time, then note where the visit started and how deep it went. Use the whole sequence, not the last click.
Compare two sessions. In the first, a visitor arrives from a software comparison site, opens a product page, then moves to pricing and documentation. That pattern looks like active evaluation. In the second, a visitor arrives from an ad, reads one page and leaves. Weight the first well above the second.
Map pages to standard topic categories at the company-domain or cohort level, not per person. A shared category scheme lets you compare sessions across sites that name their pages differently. It also keeps the output away from individual browsing histories, which is the safer shape for a buyer who has to explain the data to a privacy team.
How do you filter bot traffic using dwell time and scroll depth?
Bots inflate counts. A scraper can generate thousands of page requests and make a quiet site look like a hot account. Engagement signals help, because scripts rarely behave like readers. Dwell time is the number of seconds a visitor stays on a page. Scroll depth is how far down the page the visitor went.
Start by dropping sessions that show zero dwell time in rapid loops. Weight the remaining sessions by dwell time and scroll depth. Flag sessions with a very high request rate and no scrolling. We do not give a single numeric cutoff here, because a safe threshold depends on the site and the panel, and no public standard sets one. Ask the vendor to state its rule and to show how many sessions it removed.
Also check the panel's device mix. A panel built from browser extensions, virtual private network apps or home routers skews toward different users, and one device type can look more bot-like than another for reasons unrelated to bots. The clickstream panel guide explains how those skews arise and which use cases survive them.
Which checks should every buyer run on a sample file?
Run the checks below against a trial sample before signing. Each one has a pass condition you can test in an afternoon. Request the sample through your normal sample file process and keep the results with the contract file.
Pipeline checks for a clickstream trial sample, as of November 2026. Source: MITRE CWE-598, OWASP query string guidance, RFC 9110, W3C Referrer Policy, GSDSI editorial analysis.
Check
What to inspect in the sample
Pass condition
URL scrubbing
Page and referrer addresses, scanned for the at sign, its encoded form and token-style parameters
No emails or tokens remain, and search-term parameters still appear
Entry context
Referrer fill rate and the most common referring domains
Referrers are present for most sessions and look like real sites
Topic mapping
Category tags on a hand-picked set of pages you know well
Tags match what each page is about
Bot filtering
Dwell time, scroll depth and request rate per session
Zero-dwell loops are gone and the vendor states its rule
Panel mix
Device and browser split in the file
The mix is disclosed and fits your use case
For the topic check, choose a hundred pages from your own market and audit the tags by hand. A vendor that has tested its tagging will welcome the audit. A vendor that has not will tell you the tags are proprietary. For scoring logic once the file is clean, see the B2B intent scoring model guide.
Can you link clickstream sessions to MAIDs or hashed emails?
A MAID and a hashed email (HEM) are person-level identifiers. Whether sessions can be linked to them depends on the law, the notice given to the visitor and the consent behind the data. Many teams roll sessions up into cohorts first. Others match consented records inside a clean room, where neither side sees the other's raw rows. See the glossary entries for MAID and HEM for plain definitions.
Stakes differ by statute, so involve counsel before any linkage. As one reference point, GDPR Article 83 lets regulators set fines that scale with a company's worldwide annual turnover for infringements of the basic processing principles, including the conditions for consent. State laws in the United States differ again. This guide is not legal advice.
What is the next step for a clickstream buyer?
Request a matched clickstream sample and run the five checks above before you compare vendors on price. GSDSI describes its offer on the clickstream and web intent data page, and the team can scope a sample for your use case through the contact page. Ask for the scrubbing rules in writing at the same time.
Frequently Asked Questions
Can teams join clickstream sessions to MAIDs or hashed emails?
It depends on the law and on the consent attached to the data. The GDPR, for one, requires a lawful basis, and consent is only one of the possible bases. State laws differ. Most teams aggregate session paths into cohorts first, or match consented records in a clean room. Ask counsel to approve the final design.
How much delivery lag does clickstream processing add?
Scrubbing personal data, removing bots and tagging categories all take time, so a cleaned feed arrives later than a raw one. The lag differs by vendor and by delivery method. Ask any supplier to state its lag in hours, per table, in writing, and test it by comparing event timestamps with file arrival times in your sample.
Why not just count visits to a domain?
Raw visit counts mix people, scripts and one-off clicks, so a bot loop can look like ten buyers. Path, referrer, dwell time and scroll depth give a truer read of intent. They also let you compare sessions fairly across panels whose device mixes differ, which a plain visit count cannot do.
What should a buyer ask for first?
Ask for a sample file and the vendor's written scrubbing rules. Scan the sample for emails and tokens, check that safe search terms survive, and audit the category tags on a set of pages you know. A vendor that cannot show these results has probably not tested them, and that is worth knowing before you sign.
Does removing query strings solve the PII problem?
Not on its own, and it can throw away useful intent. Deleting every query string also removes site-search terms. Scrub the identifiers instead: emails, tokens and session keys. Then check the referrer, because the previous page's address can carry the same values. Verify the result by scanning a sample, not by trusting the rule.
✓ Opt-Out Request Honored via Global Privacy Control
We use cookies and similar technologies to improve your experience. No marketing cookies are pre-selected. Learn more in our Privacy Policy. Your Privacy Choices.