Privacy
The short, plain version of what we store, who is responsible for it, and how long it lasts, checked against the running code rather than written once and left to drift.
What this service does
Linktrail follows people across three layers: email-open pixels, link clicks, and website activity captured by a snippet installed on a customer's site. Every site the snippet runs on has one of two modes, chosen by the customer:
- Basic mode (the default). We record people who arrive through a link we tracked for the customer — a link addressed to one named contact, or a link the customer posted publicly. A stranger who clicks a public link is still recorded, with no name and no email attached — see “People we cannot name” below.
- Extended mode. Also records a visitor who arrives on their own, with no tracked link at all. See “People we cannot name” below for exactly what is stored.
It is not general website analytics: in basic mode, a visitor who arrives with no tracked link, and has never clicked one before, is not recorded at all.
Who is responsible
The account — our customer — decides who to track and why, and is responsible for having a lawful basis and any consent that requires. We are the processor. We do not decide who gets tracked, and we do not use the data for our own purposes.
The honest position is that the customer's own site needs consent, not that none is needed: recognising a returning visitor requires storing a persistent identifier on their device, which sits squarely within ePrivacy Article 5(3), the EU rule governing storage on a visitor's device. There is no audience-measurement exemption for following a named prospect after a cold email.
Extended mode raises the same bar for every visitor, not only the ones who click a link. Once it records data, the customer needs consent from anyone who lands on that site — not just the contacts they emailed.
Child accounts. An account can be given the power to create child accounts for its own customers or profiles (a “partner” — GTM Outreach is the case this exists for). For everything recorded inside a child account, the partner that created it is the controller, exactly as an account is for its own data, and we remain the processor. A child account has no sign-in and no real email address of ours behind it, so the partner is the only party we can reach about it.
Accounts with profiles. The signed-in owner of an account that has profiles (child accounts) can read, but not change, the opens, clicks and page activity recorded for those profiles, never including IP addresses, and such reads are not logged separately.
What is stored
- A contact: an opaque reference the account gives us. We never learn a name or an email address — the account is the only party that can map the reference back to a real person.
- Email opens, one row per tracked message, with a coarse verdict (
human/proxy_open/prefetch/scanner/unknown) rather than a bare yes or no. - Link clicks and website activity — pages visited, active time — recorded for an identified contact when the link or the visit belongs to one, and with no name attached when it does not (a link shared publicly, clicked by a stranger). See “People we cannot name” below for what that second case means and how long it lasts.
- A network attribution label for some opens — e.g. that a fetch looks like it came from Google's, Apple's, or a corporate scanner's infrastructure — matched in-request against a small bundled list of known provider IP ranges. This is a separate mechanism from the country lookup below and does not by itself produce a country.
- A persistent identifier, stored in first-party storage on the customer's own site, so a returning visitor is recognised as the same person coming back — whether or not we know their name. This is the piece that needs consent.
- IP addresses, for link clicks and email opens. We store the address behind a click or an open, in exactly one column (
events.ip; one database index holds the same value and is cleaned in step with it, see the next item). For a click, a country alone cannot answer how many different people clicked a link that was shared publicly, or whether the same person clicked it five hundred times. For an open, the address is what tells a fetch by a data centre, a mail provider's image proxy or a security scanner apart from one by a person, and one reader apart from another. We keep an address for 60 days by default (adjustable per customer, capped at 90 days), then an automatic sweep removes it — see “Retention” below. This does not apply to website page views: we have no need for it there. - A connection kind and a reader label on each email open and link click. The moment an open or click arrives, we sort the address it came from into one of five kinds — a data centre (including VPNs and relay or CDN exits), a home or mobile connection, a mail provider's image proxy (Google, Apple, Microsoft), a security scanner, or unknown — using an offline list that ships with the service and never sends the address anywhere. The kind is stored beside the event and kept as long as the event is. It is an inference from the address, not a lookup of who was behind it: “home or mobile” means only that the address is on none of our lists. Separately, each open or click of an email gets a reader label (r1, r2, ...): a plain number, per email, assigned in order of first appearance to each address-and-device-type pair, and never computed from the address (a number derived from an address could be reversed by trying every possible address, and would then be the address). It means only “the same reader as, or a different one from, another open of this same email”; r1 of one email says nothing about r1 of another. Opens through Gmail's image proxy, and any other mail proxy on our list, get no label — every reader behind a provider shares that provider's addresses, so we say nothing rather than something false — and neither do opens with no usable address. Apple Mail Privacy Protection opens usually show up as a data centre and do get labels; because Apple's relay addresses rotate, one Apple reader can appear as several labels. The label is kept as long as the email's own record exists — longer than the address — together with its first and last times and whether any of its opens looked like a person; erasing the contact removes it. Matching ends when the address does: recognising a returning reader needs the address, so once the sweep has removed it (60 days by default), a reader who comes back gets a new label, so a count of readers per email can be too high. It can also be too low (two people behind one address and device type count once, and an email past its reader limit stops labelling), so it is a count of labels, not of people. While an address is stored, it also sits in a database index used for that matching; the index only holds rows that still carry an address, so it is emptied in step with the sweep and is not a second copy that outlives it.
- A country, derived from that same click address, using a small offline lookup that never sends the address anywhere else. Unlike the address itself, this is stored only for link clicks — never for email opens (an address tied to an email open is often a mail provider's or a corporate scanner's own infrastructure rather than the recipient's, which would make a country derived from it actively misleading, not merely unused) and never for page views. Not every address resolves to a known country; when it does not, we simply store nothing for that field.
- A child account, for partners only. The partner's own reference for it, an optional display label the partner chooses (stored as given — put a name there only if you mean to hold it here), its status, and a synthetic placeholder email address that is not a real mailbox, is never mailed to and is never returned by any API. We also keep a partner audit trail of what a partner key and an admin did, recorded by numeric ids and the short visible key prefix only — never a name, label, email address or raw key — kept for as long as the partner account exists, then deleted with it. It holds no personal data.
“Exactly one place” includes our own infrastructure logs, not only the database. The web server in front of this service keeps an access log so a bad deploy can be diagnosed — by default that log records the same visitor address a second time, outside events.ip, outside the retention sweep above, and outside erasure. It is configured to drop the visitor's address before a line is ever written, so the log still exists for diagnosing a deploy, but is not a second place an address lives.
People we cannot name
Not every visit belongs to someone we can identify. This section says plainly what we keep about a person with no name attached, and for how long.
A stranger who clicks a shared link
When someone clicks a tracked link that belongs to no one in particular — one shared in a document or posted publicly — and lands on a site running our snippet, we create a record for them with no name, no email address, and no contact attached. They are identified only by a random token stored in that site's own browser storage. It lets us recognise that the same person came back, without ever knowing who they are.
A visitor who never clicked a link
A customer can turn on extended mode for one of their sites, to additionally record visitors who arrive entirely on their own — no tracked link, no prior visit. The first time someone in extended mode has no identifier already stored, their own browser mints itself the same kind of random identifier described above — entirely on their device, before we ever hear about it. We only start a record for it once that browser reports real activity on the page — a page visited, time spent, how far they scrolled — never merely because a page loaded. From that point on, the pages they visit and how long they spend on them are recorded against that identifier, with no name and no email attached, the same way a stranger who clicks a shared link is recorded. In basic mode none of this happens: a visitor with no tracked link is not recorded, however long they stay on the site.
Turning extended mode off stops this immediately, not only for visitors it has not met yet. The moment a customer switches a site back to basic, anyone who was only being recorded because extended mode was on stops being recorded on their very next visit — we do not wait for anything to expire first. This does not delete what was already recorded about them; retention (below) is what removes that. It also does not affect a stranger who is recorded because they clicked a shared link — that keeps happening in basic mode, on the same site, because it never depended on extended mode being on.
When a stranger becomes a known person
If someone we already hold an anonymous record for later clicks a tracked link that does name a contact, on that same browser, everything they did anonymously becomes attributed to that contact — their earlier page views join their history, under their name.
Be clear about what that means: this data was never truly anonymous. It was collected so that it could later be matched to a name, which is different from data that can never be linked back to anyone. From the moment it is written, it is data that is capable of naming someone — not only from the moment it actually gets matched.
Two things limit how far this reaches, so a shared computer cannot misattribute one person's browsing to whoever next uses it:
- A time limit. Matching only reaches back 30 days. On its own this is a weak limit — “researched us three weeks ago, then clicked today” and “someone else used this computer three weeks ago” look identical from a clock alone — which is why the next point matters more.
- One identifier, one person, permanently. Once a browser's identifier is attached to a named contact, it stays theirs. If a different, second person is later recognised on that same browser — a shared computer, a family device — they get a brand-new identifier and a brand-new record. The first person's history is never handed to them, and theirs is never added to it.
We never use an IP address or network to decide that two website visits belong to the same person, or to attach a visitor to a contact — only the identifier already sitting in that visitor's own browser. An office, a phone network or public wifi can put many unrelated people behind one address; matching on it would attribute a stranger's browsing to the wrong person, with no way for anyone — including us — to tell that it happened.
The one place the address is used to match, and what limits it. Within a single email, the reader label (see “What is stored”) IS decided from the address plus the type of device, and only among the opens and clicks of that one email. It never joins an email open to a website visit or to a visitor's identifier, and it never reaches across emails. It can be wrong in both directions: too low when two people behind one address and device type look like one reader (or when an email has reached its reader limit), and too high when one person opens on two networks, when Apple's rotating relay addresses give one reader several labels, or when a reader returns after the address has been deleted. That is why it is called a label, and a count of readers is a count of labels, never of people; the API says when a count spans longer than the address is kept and when it is capped.
What we do not store
- A contact's real name or email address — we only ever hold the opaque reference the account gave us.
- Any location more precise than the country named above — no city, no region, no coordinates.
- A country for anything other than a link click — not for an email open, not for a page view.
Erasure
Erasure is genuine, not a flag. Deleting a contact removes every row that carries their reference — messages, links, visitors, nonces, events, reader labels, website activity — plus the daily rollups tied to their messages and links. Every table that carries a contact's data is required to be covered, and a test enforces it. See the API docs for the actual endpoint.
An audit trail records that an erasure happened and how many rows it removed, keyed by a one-way hash of the contact's reference — never the reference itself.
Child accounts. Erasure is unchanged inside a child account — per contact, using that child's own key. A child account itself cannot be deleted through the service today, and an account that still has child accounts is refused if it asks to delete itself: erasing a child's contacts removes its people and their data, and leaves the account's own record (its reference, label and placeholder address).
The honest limit: a record with no name cannot be erased by name. The anonymous records described above — a stranger's visit, not yet matched to anyone — have no contact to erase. “Delete this contact” can never reach them, because there is no contact. Retention is what removes them instead (see “Retention” below). The moment one of these records is matched to a contact, it becomes reachable by that contact's own erasure request, like anything else.
Backups. Nightly on-box database snapshots already run: a full copy of the database is taken automatically every night, and normally only the two most recent are kept. So an erasure is no longer complete and immediate everywhere the moment it runs — corrected here 2026-09-23. It removes the contact from the live database straight away, but a copy of what was erased can remain in an on-box snapshot afterwards.
The honest limit: we cannot promise a fixed number of days for that on-box copy. The clean-up that removes an older snapshot only runs once a new one has succeeded — so if nightly snapshots start failing (the box running low on disk is the realistic way that happens, and is also exactly when a snapshot attempt fails), the last snapshot taken before an erasure is kept until snapshots start succeeding again, with no automatic outer limit on how long that could take. Separately, a small number of one-off copies made by hand for a specific, unrepeated reason (for example, immediately before a one-time maintenance operation) sit outside this automatic system entirely and are removed only by hand, on no schedule at all.
Off-site backups are built, not yet running. The code that ships an encrypted copy of each night's snapshot to separate storage (Backblaze B2) has been written and reviewed, but nothing has been turned on in production yet — nobody has entered the storage credentials on the box. Once it is: an erasure will still reach the live database and the on-box snapshot on the same schedule as today, and the off-site copy is locked for 30 days in a way the box's own access cannot override — even someone who broke into this box could not use what is stored there to delete or shorten that lock early. Published as a firm ceiling of 35 days — the 30-day lock plus a margin for the daily check that removes a copy once its lock expires — because a privacy notice needs an upper bound it cannot be wrong about.
Opting out
Unauthenticated, and takes the visitor's own persistent identifier. What happens next depends on whether that identifier belongs to a known contact:
- A known contact. Every identifier belonging to them (one per device) is deleted, so none of their devices are recognised any more, and they are marked opted out. Pixels and links keep working — a tracking pixel breaking in a real inbox is not an acceptable way to honour an opt-out — they simply stop recording anything for that contact from that point on, even under a brand-new identifier.
- A stranger with no name attached. Their one identifier is deleted, so that browser is no longer recognised. There is no contact to mark opted out; a later visit starts a fresh, unrelated record, like any other new stranger.
Either way, opting out stops future tracking — it does not erase history already recorded. Erasing history needs a contact to erase (see “Erasure” above). See the API docs for how to call it from your own unsubscribe page.
What already happened is not special-cased by opting out — it ages out on the normal clock, not a faster one. An IP address already stored against one of their link clicks survives opting out; only the ordinary sweep removes it, 60 days by default and adjustable per customer, capped at 90 (see “Retention” below). For a known contact, their past page views, clicks and opens stay attributed to them the same way, under the same retention policy as anyone else's history (see “Retention” below).
Retention
| What | Kept for | Removed by |
|---|---|---|
| Raw events (opens, clicks, page views) belonging to a contact | 13 months, by policy | Already works — an automatic sweep removes it. |
| Daily rollups (per message, per link) | With the message or link they count — for as long as it exists. | Removed when that message or link is deleted — today, that means when the contact is erased. The automatic sweep above does not reach them: it ages raw events by time, and these age with the message or link instead. |
| Account-wide totals (no contact in them) | Kept forever | Not personal data — nobody to remove. |
| Partner audit trail (ids and key prefixes only — see "What is stored") | For as long as the partner account exists | Deleted with the partner account. Not personal data — nothing in it names anyone. |
| Anonymous records (no name attached — see "People we cannot name") | 30 days by default, adjustable per account | Already works — an automatic sweep removes them. |
| IP addresses (link clicks and email opens — see "What is stored") | 60 days by default, adjustable per customer, upper limit 90 | Already works — an automatic sweep removes it. Matching a returning reader to their label ends with it. |
| Connection kind of an open or click, and the reader label on it (see "What is stored") | The connection kind: with the event it describes, 13 months by policy. The reader label and its times: with the email's own record, for as long as it exists. | The kind goes with its event. The label goes when the email is deleted — today, that means when the contact is erased. The automatic sweep above does not reach labels: they age with the email, not with the event. |
| The identifier that recognises a returning visitor (see "What is stored") | 180 days by default, adjustable per account | Already works — an automatic sweep checks every 30 seconds and clears up to 500 expired identifiers per pass, shared across every customer using this service, and detaches each one from any contact's history without touching that history. We also still ask the browser to forget its own copy. The one gap: a copied or replayed identifier can still be recognised until that sweep reaches it — the very next pass in the ordinary case, longer behind a large backlog, for example right after this window is lowered for many people at once. |
| On-box database snapshots (a full nightly copy of everyone's data, including anyone erased since) | Normally the 2 most recent nights; no fixed outer limit if nightly snapshots start failing — see "Backups" above | Deleted automatically once a new snapshot succeeds. One-off hand-made copies are outside this and are removed only by hand. |
| Off-site encrypted backups (Backblaze B2) | Up to 35 days once running (a 30-day lock, plus a margin for the deletion check) | Built, not yet running — see "Backups" above. Locked in a way the box's own access cannot override early. |
Retention length for the events above is a per-account setting (accounts.retention_days), because a processor acts on the controller's documented instructions rather than choosing this unilaterally. The daily rollups follow the message or link they summarise instead — not this setting, and not the sweep that enforces it.
The identifier window is accounts.identity_ttl_days, and it governs only the last row above — not the events, not the rollups. An automatic sweep now enforces it: once the window passes, our own record of the identifier is removed. That sweep checks every 30 seconds and clears up to 500 expired identifiers at a time, shared across every customer using this service. The browser is still asked to forget its own copy too, and a copied or replayed identifier can still be recognised until that sweep reaches it — the next pass in the ordinary case, longer behind a large backlog.