Systems and records
Duplicate opportunities in pipeline metrics: count each buying event once
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Duplicate opportunities bend pipeline metrics in predictable directions: a copy closed as lost drags win rate down, an open copy inflates pipeline value and forecast coverage, and a merge that keeps the newer record shortens measured cycle time. Restate each metric per buying event from records frozen before the merge, publish recorded and restated figures side by side, and keep every record of one purchase in a single training or test split.
System export
Duplicate opportunities
Key takeaways
- Count the buying event, not the CRM row: every pipeline metric that duplicates distort is repaired by restating it per purchase.
- A duplicate closed as lost adds a loss that never happened; in an illustrative case, 100 such rows turn a 25 percent win rate into 22.7 percent.
- Pipeline value and forecast coverage overstate one deal for every day two records for it are open; store that overlap interval and count the amount once.
- Cycle time shrinks when a merge keeps the newer record, because Odoo and EspoCRM leave the survivor with its own creation date; start the clock at the group's earliest date.
- None of the three open-source merge tools read here stores why records were merged, and all take the losing record out of normal use, so restating depends on a snapshot taken first.
- Assign whole buying events to one split; duplicated test data inflated reported metrics by up to 100 percent in a study of code datasets.
How do duplicate opportunities distort pipeline metrics?
They push each metric in a predictable direction, and the direction depends on how the duplicate was cleaned up. A copy closed as lost adds a loss that never happened, copies left open inflate pipeline value and forecast coverage for as long as both stay open, and a merge can shorten the recorded sales cycle when the record kept is the newer one.
Illustrative arithmetic, not data from any company: a team closes 1,000 buying events in a year and wins 250, a true win rate of 25 percent. If 100 of those events also had a second record that someone closed as lost with a duplicate reason, the CRM shows 250 wins in 1,100 closed opportunities, or 22.7 percent. Had both copies been closed as won instead, bookings for those 100 events would be counted twice.
Cycle time can move the other way. Suppose one rep opened a record in March, a colleague opened a second in May for the same purchase, the deal closed in September, and the May record is the one kept. Odoo's merge takes the first non-empty creation date in its own sort order, which is the kept record's own, and EspoCRM never offers the creation date as a merge choice. A deal that took six months now reads as four.
Whether two look-alike records describe one purchase, and which of them should survive, is a separate decision that the site's guide to sales duplicate opportunities covers. This article takes that decision as given and asks what it does to the numbers. The matrix below shows the direction of the error for each metric under each cleanup path:
| Metric | Copy closed as lost | Both copies left open | Copies merged into one |
|---|---|---|---|
| Win rate | Falls: a loss that never happened joins the denominator | No effect until a copy closes | Right only if the kept outcome is the event's real one |
| Pipeline value | Overstated from the second record's creation until the copy closes | Overstated by the full amount for every day both are open | Overstated before the merge; afterwards one record's amount |
| Forecast coverage | Overstated while the copy sat in a forecast category | One commitment counted twice in category totals | Doubled before the merge; afterwards it depends on which amount was kept |
| Cycle time | Adds a cycle that never happened to lost-deal averages | Deal age read from the newer copy looks younger than it is | Shrinks when the newer record is kept |
| Stage conversion | Adds a drop-out at whatever stage the copy reached | One deal passes the same stages twice | The kept record's path stands in; the other is moved or left on a hidden record |
| Bookings and deal size | Lost-deal size averages gain a row that never existed | No effect until close; two wins would double bookings | One record's amount, never a sum, which may differ from the signed order |
Five rules for restating pipeline metrics per buying event
Restate every metric from the corrected unit, the buying event, and publish the recorded and restated versions side by side so the size of the correction stays visible. Each rule reads its inputs from snapshots taken before the merge, never from the record that survived it.
At a glance
Five rules for restating pipeline metrics per buying event
Two habits make restated figures checkable. Report duplicate closures as their own count instead of burying them, so a reader sees how many rows left each denominator, and write the counting rule out in words next to every restated number. The rules are the minimum:
- Win rate: one outcome per buying event; rows closed only because they were duplicates leave the denominator and are reported as their own count.
- Pipeline value and forecast coverage: for every day two records for one event were open, count the event's amount once, in the forecast category of the record that carried the deal, and store that overlap interval with the correction.
- Cycle time: start the clock at the earliest creation date in the group, read from the pre-merge snapshot, and stop it at the event's single close.
- Stage conversion: follow the stage path of the record that carried the deal, and keep the other path as context tagged with its source record.
- Bookings and deal size: take the amount from the signed order where one exists, and never add two records' amounts for one event.
Why a restated metric needs the records frozen before the merge
Because the merge itself removes most of the inputs those rules need. We read the merge code of Odoo, EspoCRM and SuiteCRM at their October 2026 commits: none stores a reason for the merge, all three take the losing opportunity out of normal use, and none adds the two amounts together.
Odoo sorts the selected records by what its code calls confidence level, with lost leads last and opportunities ahead of leads, and keeps the first. Expected revenue and creation date take the first non-empty value in that order, so the kept record's own date wins. The other records are deleted once their messages and activities have moved across, each moved message retitled From and the old record's name, under a note listing the loser's revenue, probability, stage and dates. The merge wizard asks only for the records, a salesperson and a sales team.
EspoCRM saves whichever value the user picks for each differing field but never offers creation or modification times, so the kept record keeps its own. Only posts and sent or received emails follow the merge; change notes for the opportunity fields it audits, including amount, stage and close date, stay with the losing record, which is soft-deleted and, under the default configuration, purged with its notes once its last change is more than two months old.
SuiteCRM takes one value per field from the merge form, keeps a creation date only if the form posts one, relinks related records except the account link on opportunities, and marks each losing record deleted. It audits fields including the US dollar amount, sales stage, close date and probability, yet the merge code never touches audit rows, and the module that runs the merge is defined against a table named does_not_exist, with no fields, so the merge leaves no row of its own.
The table maps each finding to the metric it damages; the methods note gives the commits and the tests applied:
| What the merge leaves | Odoo | EspoCRM | SuiteCRM | What a restated metric loses |
|---|---|---|---|---|
| A reason for the merge | None: the wizard takes records, salesperson and team | None: the request carries record IDs and chosen values | None: the merge module has no table and no fields | Any way to audit whether two purchases were collapsed into one |
| The losing opportunity | Deleted from the database | Soft-deleted; default cleanup purges it after two months unchanged | Marked deleted; the row stays in the table | Its open interval and outcome drop out of normal reports |
| Amount on the survivor | First non-empty value in confidence order | The value the user picks | The value the merge form posts | The other record's amount, outside Odoo's merge note, unless frozen first |
| Creation date on the survivor | First non-empty value, in practice the survivor's own | The survivor's own; not offered in the form | The survivor's own unless the form posts another | The true start of the cycle when the newer record is kept |
| The loser's history | Messages and activities moved, retitled From plus its name | Posts and emails moved; field-change notes stay behind | Related records relinked; audit rows left on the deleted record | A clean second stage path for conversion analysis |
What to freeze for each restated metric, and what to leave out
Freeze every record in the group, with its native ID, at the moment someone decides to correct it, and keep only what a restated metric or an evaluation needs: status and close fields, dated amounts, creation dates and stage history for each record, plus the decision itself, which none of the CRMs read here records.
Store match evidence as yes-or-no flags rather than as the personal data behind them. Odoo's duplicate hints show what such checks rely on: an exact match on the email domain, an exact match on a normalized phone number, or the same commercial entity. A dataset needs to know which matched, not the address or the number, so keep flags such as same account, same email domain and same phone, plus product overlap and the gap in days between creation dates. Contact names, email addresses and phone numbers stay out.
The map below pairs each input with the metric that depends on it:
| Input to freeze | Metric that depends on it | Where it sits before the merge | Treatment in the dataset |
|---|---|---|---|
| Status, close date and close reason of each record | Win rate | Stage and close fields of each opportunity | Snapshot every record with its native ID |
| Dated amount and forecast category of each record | Pipeline value, forecast coverage, bookings | Opportunity fields and their change history | Keep each change with its date so the overlap interval can be rebuilt |
| Creation date of each record | Cycle time | The opportunity row | Snapshot it; never read it from the record kept |
| Stage history of each record | Stage conversion | Audit rows, tracked messages or stream notes | Export per record, tagged with the source ID |
| Match evidence | Model evaluation and hard negatives | Account, email domain, phone, products, dates | Yes-or-no flags and day gaps; contact values dropped |
| Decision, decider role and reason code | Separating duplicate closures from real losses | Nowhere, by default | Written before the merge runs, with the settling document |
How each correction state enters the metrics
Record the correction state apart from the metric arithmetic, because the same pair of records counts differently depending on what was done to it. A merge can be right while its outcome accounting is wrong, and a pair left apart can still distort pipeline if nobody records that it was reviewed.
Pairs reviewed and kept apart deserve as much care as merges. They are the hard negatives for any model that proposes corrections, and a set built only from merges would teach it that every look-alike pair is one purchase.
| Correction state | Win rate | Pipeline and forecast | Cycle time |
|---|---|---|---|
| Merged into one record | One outcome for the event | Amount once; overlap interval stored | From the earliest creation date in the group |
| Closed as a duplicate, not merged | Out of the denominator; reported as its own count | Amount once from the day the second record opened | The copy's own cycle dropped from lost-deal averages |
| Reviewed and kept apart | Two outcomes, each counted | Both amounts count | Each record's own dates |
| Earlier merge reversed | Outcomes restored to each event; the bad merge kept as an error example | Amounts restored for the merged period | Each event's own dates restored |
| Unresolved | Held out and reported as a count | Shown as a range with and without the pair | Held out of averages |
Keeping one buying event out of both the training and test sets
Split by buying event, never by CRM row. If one record of a duplicate pair lands in training and the other in the test set, the model is graded on a purchase it has already seen.
The effect is large in other fields. Allamanis's study of code datasets (arXiv, 2018) found reported performance metrics sometimes inflated by up to 100 percent when models were tested on duplicated corpora rather than de-duplicated ones. Opportunity records carry the same risk in a quieter form, because two records for one purchase share the account, the products, the amounts and much of the timeline.
Give every record a buying-event ID during correction and assign whole events to one split. Keep pairs reviewed and kept apart on the same side as well, so the test set does not reward a model for recognizing an account it met in training. The site's guide to task family IDs applies the same principle to workflow tasks.
Grade against restated targets too. A win-probability or forecast model scored on recorded outcomes is rewarded for predicting cleanup artifacts, such as a duplicate's phantom loss, so score it on the restated outcome and report the recorded score beside it. Judge a model that proposes corrections against the evidence available when each correction was made, not against the deal's final result, which says little about whether two records were one purchase.
Measuring how far duplicates moved the numbers: a design
Measure the duplicate rate and the size of each metric correction with a fixed sampling rule and blind review; SourceX has not published results from this design. Draw candidate pairs from a stated window, for example opportunities created in one fiscal year, using the match flags above, and add an equal number of random pairs so the rate is not estimated only from suspicious ones.
Two reviewers assign each sampled pair a correction state from the pre-merge snapshots without seeing the final outcome, and their agreement is reported with a chance-corrected statistic such as Cohen's kappa, beside the count of pairs in each state.
Report four numbers, each with its denominator: the share of closed opportunities that were duplicate copies, win rate before and after restatement, the median change in cycle time for merged groups, and the largest single-day pipeline overstatement in the window as a share of restated pipeline. State the window, the match rule, the CRM and its version, and the export date, and keep unresolved pairs visible rather than dropping them.
Asking a supplier for restated deal histories
Ask for the correction ledger and both versions of each metric together with the opportunities, and make the ledger's completeness part of acceptance. A ledger exists only where a company ran its corrections deliberately and kept notes, so the request itself tells a supplier which histories qualify.
For buyers, SourceX finds records like these at businesses after a request arrives instead of holding an inventory, and the request form collects the scope, volume, format, timing and licensing terms needed. The company supplying records approves the scope and terms and signs before any delivery, and whether its correction history is complete enough is settled by reviewing that history, not assumed.
A request should name the CRM and version, the years covered and the correction states wanted, and it should list the tests a delivery must pass:
- Every opportunity ID that disappeared or was closed as a duplicate appears in exactly one correction row.
- Each correction row holds pre-merge snapshots of all records in the group, a reason code and the decider's role.
- Win rate, pipeline value and cycle time arrive twice, as recorded and as restated, with the counting rule written out.
- No buying event appears in more than one split.
- Contact names, email addresses and phone numbers are absent; match evidence arrives as flags.
Methods note: how the three merge implementations were read
The code was read on 2026-10-27 at three commits: odoo/odoo f7780f3 (addons/crm), espocrm/espocrm 9cf566c and salesagility/SuiteCRM be52770 (modules/MergeRecords and modules/Opportunities), each cloned shallow into a scratch folder. A script, merge_audit.py, read the files as text and printed the source line behind each finding; it executed nothing from the repositories.
Counting rules, as the script applies them: a merge stores a reason only if its entry point declares or reads a field whose name contains the word reason (the Odoo wizard's fields, the EspoCRM request's keys, the SuiteCRM merge module's fields). The losing record's fate is the call the merge makes on it: unlink in Odoo, the record service's delete in EspoCRM, mark_deleted in SuiteCRM. Amount and creation date are read from Odoo's merged-field list and its first-not-null rule, from EspoCRM's merge form, which skips createdAt and modifiedAt, and from SuiteCRM's handling of posted form values and a posted date_entered. History handling is read from the code that re-points messages or notes, and from a search of SuiteCRM's SaveMerge.php for the word audit.
Not checked: commercial CRMs, plugins and custom merge flows, database triggers, and behavior under non-default settings, such as EspoCRM with cleanup turned off. The win-rate and cycle-time figures above are illustrative arithmetic, not measurements.
Frequently asked questions
Does closing a duplicate opportunity as lost distort win rate?
Yes. Each duplicate closed as lost adds a loss that never happened to the denominator. In this article's illustrative case, 100 duplicate losses turn a 25 percent win rate into 22.7 percent. Report duplicate closures as their own count and leave them out of win-rate calculations.
How should pipeline value count an opportunity entered twice?
Count the buying event's amount once for every day both records were open, and store that overlap interval with the correction. Take the amount from the signed order where one exists rather than adding the two records' amounts. Publish recorded and restated pipeline side by side so the size of the correction is visible.
Which creation date should a merged opportunity use for sales cycle length?
Use the earliest creation date among all records in the group, taken from a snapshot made before the merge. Odoo and EspoCRM keep the surviving record's own creation date, so a merge that keeps the newer record shortens the measured cycle.
Can pipeline metrics be restated after duplicates were already merged?
Partly, and it depends on the CRM. Odoo deletes the losing record, but its merge note on the survivor lists the loser's revenue, probability, stage and dates; SuiteCRM leaves the row in its table marked deleted; EspoCRM keeps it soft-deleted until its default cleanup purges it. None of the three records why the merge happened, so freeze the group first wherever you can.
How do I keep a duplicate pair from leaking into an AI test set?
Assign each record a buying-event ID during correction and split the data by that ID, so every record of one purchase lands on the same side. Keep look-alike pairs that were reviewed and kept apart on one side too, so the test does not reward recognizing a familiar account.
Sources
- Odoo's merge sorts records by confidence level, keeps the first non-empty value for fields including expected revenue and creation date, moves messages and activities to the survivor and deletes the other records; its duplicate hints match email domain, normalized phone or commercial entity. Source
- Odoo's merge wizard declares only the records to merge, a salesperson and a sales team; it has no reason field. Source
- Odoo's merge note lists each merged record's name, expected revenue, probability, stage and dates on the surviving record. Source
- EspoCRM's merge moves only post and email notes to the target, then deletes each source record through the record service. Source
- EspoCRM's merge form offers differing mergeable fields for the user to choose and skips createdAt and modifiedAt. Source
- EspoCRM's cleanup job removes soft-deleted records whose last change is older than its cleanup period, two months by default, along with their notes. Source
- EspoCRM's default configuration turns deleted-record cleanup on. Source
- EspoCRM audits the opportunity fields amount, account, stage, lead source, close date and campaign. Source
- SuiteCRM's merge saves posted field values, keeps a posted creation date, relinks related records except accounts on opportunities, and marks each merged record deleted. Source
- SuiteCRM's MergeRecord module is defined against a table named does_not_exist with no fields, so the merge leaves no record of itself. Source
- SuiteCRM enables duplicate merge for opportunities and audits fields including amount, sales stage, close date and probability. Source
- Duplicated code corpora inflated reported machine learning performance metrics by up to 100 percent compared with de-duplicated ones. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.