Does your lead score pick out the people who buy?
Lead scoring measures the score you already have against what really happened to those people, then shows what a score built from your own records would have caught instead. Nothing here is modelled by AI: every figure is counted, and every claim has to clear a volume floor and a confidence test. If the question is where people stop rather than who to call first, read Where do people drop off, and how long does each step take?
Before you start
Open Analysis, choose your tool, and click Lead scoring. It reads records already synced into your workspace, so a sync is what moves the figures.
Check the business profile strip at the top of the tab before you read anything. It carries your business model, deal size band and what “qualified” means here, and the words on this page follow it. It reads all synced records rather than the Created range, which is why its counts differ from the ones below.
Choose what “worked” means
The row headed Judge scoring against changes the question, because marketing, sales and the board each ask a different one of the same score. Three choices, each with the number who reached it printed on the pill.
| Choice | The question it answers | The population |
|---|---|---|
| Became a customer | Does a high score mean someone becomes a customer? | Everyone read. Success is the lifecycle stage saying Customer, or being on a won deal — the sentence under the tab says which of the two supplied how many. The pill uses your business profile’s word for a buyer, so it reads Became a client on a services profile and Became a paying customer on a product-led one |
| Is on a sales deal | Does a high score mean sales opens a deal with them? | Everyone read. Counted in people, not deals: one deal can carry several people |
| Is on a won deal | Among people whose deal has closed, does a high score mean it was won? | Only people whose deal has closed. Open deals are left out because nobody knows yet how they end |
A choice is available only when at least 30 people reached it and at least 30 did not. Below that the pill is greyed and its tooltip prints both counts. Where the one you asked for cannot be judged, the first one that can is used instead.
Take care
The two deal-based choices count people, not deals. Several people can sit on one won deal, so Is on a won deal is never the number of deals you won. Say which of the three you are quoting whenever you repeat a figure from this tab.
Which fields are read as scores
There is no fixed list. Every field whose name suggests a ranking is read — score, scoring, grade, rating, rank, propensity, likelihood, probability, priority, fit, ICP, health, intent, engagement level and points among them — from three places, so a field is not missed because it is new or empty.
- Your tool’s full field list, saved by the sync. This is what finds a scoring field nobody has ever filled.
- The keys actually present on the most recent 3,000 records, which catches a field added since the last catalogue was saved.
- Whatever the data audit already recognised as a score field.
Up to 60 fields are taken, with the ones literally called “score” first, so a cap can only ever trim the long tail. The name opens the door and nothing more: each field is then judged on its values. The note under the card prints how many were read in all, and names every field that was set aside with the reason.
- Set aside as “a satisfaction, survey or social rating, not a measure of who will buy”: anything whose name says CSAT, NPS, CES, survey, satisfaction, Klout, review or feedback.
- Set aside as “its values are not a ranking”: a field holding text, dates or identifiers.
- Marked as calculated by your tool rather than by your team: HubSpot’s predictive, messaging, email, analytics, social, sequences and sales scores. Those are worth judging and cannot be rebuilt by you, so a defect in one is not a defect in your model.
How a score is judged
A numeric score is split into four equal-sized groups from lowest to highest, by how many records sit in each rather than by the range, so a score where most records sit at zero does not produce three empty groups. A letter grade is grouped by its letters instead. The share of each group that reached the chosen outcome is then compared.
The verdict on each card is one of eight words, and each carries a plain sentence with the rates in it.
| Verdict | What it means |
|---|---|
| Working | The groups climb in order and the top group’s pessimistic end still clears the bottom group’s optimistic end, at 95% confidence |
| Partly working | Top and bottom separate, but the groups do not climb in order — the middle of the ranking is not reliable |
| Ranking backwards | The bottom group converts reliably better than the top |
| Not telling you anything | Even the most generous reading of the intervals cannot reach twice the bottom group’s rate. This is a claim, so it needs evidence too |
| Too early to tell | Fewer than 30 records with an outcome in a group, or the intervals still allow a working score. No verdict is invented |
| Mostly empty | Filled on under a quarter of records, so it cannot guide anyone yet |
| Not really a score | Two distinct values, or one value on 90% of records. It is a flag, not a ranking |
| Never filled in | The field exists and holds no value on any record. Scoring was set up here and never switched on, or the tool has never calculated it for this account |
Every card also says what kind of score it is and who sets it — a fit score, an engagement score, a grade, a composite, a data-quality score or a manual priority — with the range its values run over.
What your buyers have in common
Below the scores, MarqIQ tallies the attributes on your own records against the outcome and gives each pattern a weight. The columns are More likely to buy and Less likely to buy, each line carrying the value, the points, how it compares with your average, and the counts behind it.
- A pattern earns points only when it appears on at least 30 records and its conversion is reliably above or below your average at 95% confidence. Anything thinner is left out rather than guessed at, and the count of what was left out is printed under the table.
- Points are deliberately coarse — a scale of 20, in steps — because a rubric of fives and tens gets implemented in a CRM and a rubric of 7.3s does not.
- A score of your own that reads Working or Partly working becomes an input, grouped into quarters of its own range, so long as at least 120 of the records in the building half carry a number in it. The proposal builds on what works rather than replacing it.
- At most 40 values are tested per attribute, and at most 24 lines are shown.
Note
Where no attribute stands out, the card says which of the two reasons applies: fewer than 30 people reached the outcome, or the fields that describe your buyers are too sparsely filled to learn from. Neither is dressed up as a result.
The honesty rule behind “If you scored this way instead”
A scoring built from some records and then tested on those same records always looks good. So the points above are built on half of your records — every other one — and the proposed score is then run over the other half, which it has never seen. The figure on screen is what it did on those.
The card needs at least two scored patterns and at least 60 records in the test half. It states the size of that half in its own note, and the sentence it prints compares the proposed top group with the top group of your best existing score, naming that score. Where the two are close, it says so: your current scoring is already capturing most of what the data can show.
Is a score rewarding clicks no person made?
Security software at many companies opens every link in an email before anyone sees it. A score that gives points per click collects them for those people. The card headed Is a score rewarding clicks no person made? compares contacts whose only email clicks came from security software against contacts who never clicked at all.
- It needs HubSpot or Marketo, because those are the two tools that flag scanner activity per contact.
- It needs the Email analysis to have been opened at least once for that tool. The groups come from a reading the email reader saves, so on a workspace that has never opened Email this card is absent rather than empty.
- It needs at least 30 contacts in the scanner-only group and 30 in the never-clicked group. Below that it says so, with both counts.
- A score is called out only when the cautious end of the range still puts more than half the scanner-only contacts above the typical never-clicked contact.
Take care
The card states its own weakness in the finding itself: companies running this kind of security software tend to be larger, and size alone can raise a score. Treat it as a reason to read that score’s rules for points per email click, not as a verdict on the score.
What the figures are measured on
- Up to 80,000 people are read. Above that the read is an even spread across your whole database, never the newest records — recent people have had the least time to buy, so the newest records would make any score look worse than it is. The sentence under the headline prints the sample and the population.
- The average conversion printed beside every band is your own account’s, over the records read and the outcome chosen. There is no industry benchmark anywhere in MarqIQ.
- Each band prints its own hits and denominator. Two bands on one card routinely rest on different numbers.
- The method at the foot of the tab states how “bought” was decided for this workspace, with the counts from each route.
Where it stays blank, and why
| Situation | What you see |
|---|---|
| Nobody is marked as a customer and no won deal links to a contact | A headline saying there is no outcome to learn from yet, and the steps to mark buyers in your tool |
| Fewer than 30 people reached the outcome | The count so far, and a note that scoring can be judged fairly from 30. Nothing is claimed |
| Scoring fields exist and all are empty | Each one listed as Never filled in, with a headline saying scoring has been set up and never switched on |
| No score field at all, on a tool that supports scoring | A Not set up yet card: what scoring is called in that tool, where it lives, which edition it needs, the steps to build it, and what you should see when it works |
| No score field, on a tool MarqIQ has not verified for scoring | A sentence saying MarqIQ has not verified how scoring works in that tool and will not guess at the setup, and that nothing there is a judgement about your data |
What it will not do
- It changes nothing in your tool and writes no score.
- No number here comes from a language model. Every figure is counted.
- It never claims a pattern on fewer than 30 records.
- It does not advise creating a field before checking whether you have an empty one to test in. Where you do, it names it.
If it did not work
All three Judge scoring against pills are greyed out
Why: Each needs 30 people who reached that outcome and 30 who did not. The tooltip on each pill prints both counts.
Fix: Make sure buyers are marked as such in the tool — a lifecycle stage of Customer, or a won deal with the person attached. Then run a sync from Data sources.
A field you use as your score is not listed
Why: Either its name carries none of the words that open the door, or its values are not a ranking and it was set aside.
Fix: Read the note under Is your scoring working?: it names every field set aside and why. A field whose label gives no hint of ranking is easiest fixed by renaming the label in your tool.
Your score reads Mostly empty and you know it runs
Why: It is filled on under a quarter of the records read, which usually means the rule fills only a subset, or it stopped running for most records.
Fix: Check when the rule last fired on a recent person. Is each field being used for what it is for? names a score field that stopped updating, with the month it stopped and what used to write it.
There is no Is a score rewarding clicks no person made? card
Why: That check needs HubSpot or Marketo, and it needs the Email analysis to have been opened at least once for that tool, because the groups come from a reading the email reader saves.
Fix: Open the Email tab for the same tool and let it finish, then come back. Where the tool is neither HubSpot nor Marketo, the card cannot appear: no other tool flags scanner activity per contact.
The proposed scoring did not separate buyers
Why: The patterns are real but not yet strong enough to score on, judged on the half of your records the proposal never saw.
Fix: Read the points table as findings rather than as a model. Nothing here should be implemented on a card that says the test half did not separate.
The counts differ from the business profile strip at the top
Why: The strip reads all synced records; everything below it follows the Created range you set.
Fix: Click Clear on the Created filter to make the two agree, or read the population sentence under the headline, which prints what the figures below it were counted on.
Next steps
Related articles
Was this article helpful?