What the words mean

Stitch uses a few terms precisely, because being loose about them is how you get a confident wrong answer. This page says what each one means in plain English. The screens now lead with these plain names; the precise term is kept alongside where it appears in the API or the filter grammar.

The two questions

Everything in Stitch answers one of two questions, and they are never added together.

Who first found this person? — one answer per person, ever. Set the first time we see them and never changed. If a paid Google ad brought someone in 2024, that ad keeps the credit for every loan they ever do, on either brand. Called the acquisition ledger; labelled per person.

Who produced this loan? — one answer per loan. Recalculated as new information arrives, and adjustable by the model you choose. Called the production ledger; labelled per loan.

Adding them together produces a number that looks like money and means nothing, which is why the app makes you pick one.

Confidence in the source

Every visit carries a rating for how sure we are where it came from. The screens show the plain name; the letter is the spec’s and appears in the API.

Name	Letter	What we actually have
Certain	A	The click itself — an ad click ID, a signed link, or a logged-in session
Strong	B	The landing page carried our campaign tags
Likely	C	We saw which site sent them, but not which campaign
Self-reported	D	The borrower told us where they heard about us
Unknown	E	Typed the address in, or came from a site we do not recognise

“Certain only” in the filter bar means: show me the number I could defend in a meeting.

How a loan got credited

When we credit a loan to a channel, one of four things happened. The screens name them; the API calls them fallback levels 1–4.

Name	What happened
Matched directly	A real visit inside the look-back window. The ideal
Matched to older history	Nothing in the window, so we used the first channel this person ever came from
Inherited by a repeat borrower	A returning borrower with no new visits, credited to whoever originally found them
No source found	Nothing to credit anywhere. Counted and shown rather than hidden

How a person was first found

The same four-way split, for the other ledger.

Name	What happened
We paid to find them	Their first visit was a paid, non-brand ad
We earned them	Search, social or a partner site, without paying for the click
They already knew the name	Their first visit already searched for us by name
Origin unknown	No first visit we can point to. Phone-in and bought-in leads

Words you will see

Visit / session — one sitting on the site. A new one starts after 30 minutes of inactivity, or when someone arrives from a different channel.

Touch / touchpoint — a visit that could earn credit. Hopping between our own sites does not create one, or a visitor moving from the marketing site to the application would overwrite whatever actually brought them.

Look-back window — how far back before a lead we look for visits to credit. 90 days by default. A visit older than the window does not count towards that loan.

Attribution model — how credit is split when there are several visits. Last non-direct (the default) gives it all to the last visit that was not someone typing the address in. First touch gives it to the first. Linear splits it evenly.

Newly won vs inherited — a repeat borrower who came back through a fresh ad was newly won. One who just showed up again is inherited: credited to whoever found them the first time.

Brand search — someone searching for “Beeline” by name. It defends demand rather than creating it, so it is kept out of acquisition cost and reported separately. Otherwise it flatters itself in every calculation and slowly eats the budget that generates the demand it harvests.

Untracked spend — money paid to a campaign that never showed up in a single site visit. Usually a tracking template someone edited, a campaign renamed without its tags, or tracking removed from a landing page.

Cost per funded loan — what a channel cost divided by the loans it actually funded. Distinct from cost per lead, and channels reorder dramatically between the two: cheap leads that never fund are not cheap.

Unlicensed spend — advertising aimed at states we are not licensed to lend in. Money that cannot become a loan no matter how well the ad performs.

Declared source — what a borrower says brought them, when asked. Counted and cross-tabbed against what we tracked; never credited, because a large share of people say “Google” when they mean they typed the brand name in after seeing something else. See [[declared-source]].

Brand halo — demand created by one channel and harvested by another. The declared-vs-tracked matrix is how you see it: people who say Social media while we tracked them as Paid Search were moved by the social ad.

Introduced / closed — the channel whose visit started a journey, and the channel that got the credit for finishing it. Every other number in the product names one channel per loan, so a channel that starts journeys others finish reads as zero; the assist matrix is the only place the difference is visible. Nurture touches — a reminder email, a retargeting ad — cannot introduce anybody, because they are how somebody comes back rather than how they were found.

Still alive / censored — a deal that has not died yet, and what to do about the fact that we do not know how it ends. The survival curve counts an open loan toward every day up to today and no day after — it is evidence the deal survived that long, not evidence it failed. Dividing dead by total counts it as a failure at day 30, 60 and 90 alike, which makes a channel that started spending last month read as the deadliest in the book. Funding is not a death either: a funded loan leaves alive, so the curve steps down only when deals actually die.

Alive at N days — of the loans watched at least that long, the share that had not died. Read the at-risk count with it: past the point where only a handful are still in play, one death moves the line a long way, so the curve stops there rather than carrying on quietly.

Speed to lead — minutes from a lead arriving to somebody making contact, against how often those leads fund. A prompt rather than a proof: the leads called back fastest are also the ones who pick up, so only a trial that randomises who gets called fast separates the speed from the leads.

Pipeline — loans in flight: started, not funded, not dead. Its expected value is each loan’s predicted value times how often loans at that step have actually gone on to fund. It assumes the next few months look like the last few, which is the one thing it cannot check.

Concentration — how much of funded revenue comes from the single biggest channel, or the biggest three. Not a number to optimise; one to know before somebody else tells you.

Lag — days from someone’s first visit to their loan funding. The number that says whether the look-back window is long enough: a journey longer than the window has its first visit credited to a later one instead.

In flight — a loan still working through a step, neither a win nor a loss yet. Held out of the next step’s conversion rate, because counting it as a failure would make every recent month look worse than an old one whatever actually happened.

Merge candidate — a pair of people who might be one person, proposed automatically and always shown to a human before anything is joined. Only ever proposed from weak evidence — same home network, same kind of device, one application picked up where another stopped — because the alternative reading is a household, and a spouse looks exactly like a second device.

Cohort — everyone we first found in the same month. A cohort is followed for the rest of its life, so “March’s people” means the same group whether you look at them in April or two years later.

Payback month — how many months after we found someone before the money they produced covered what we spent finding them. Month 0 is the month they arrived. It stays blank until it actually happens: a month that has not paid back is not a large number, it is no number yet.

Consent — two answers Stitch stores and enforces: may we record this visit, and may we share it with an ad platform. Stitch reads them; the banner that asks the question belongs to a consent management platform.

GPC (Global Privacy Control) — a signal a browser sends saying “do not sell or share my data”. Legally binding in California. It stops conversions going to ad platforms; it does not stop us measuring the visit, because that is not what it refuses.

Withheld — a conversion that was built and then not sent, because the person opted out of sharing. Shown on the export health card beside sent and failed, because a conversion that disappears silently looks exactly like a broken integration. See [[consent]].

Nurture spend — retargeting and CRM campaigns, aimed at people we already have. It appears beside a month’s cost rather than inside it, because the revenue it produces belongs to whichever earlier month first found those people. Charging March for it would make every month look permanently worse than it was.

Where the terms come from

The spec (Spec/beeline-attribution-spec.md) is the source of truth for behaviour and uses the precise terms throughout. This page is the translation, not a replacement — if the two ever disagree, the spec is right and this page needs fixing.