Blog · Revenue Operations · October 20, 2026 · By Ahmad Zabir

CRM Data Cleanup: Fix the Deals Before You Trust Your Pipeline

Most CRM cleanup guides scrub contacts, but pipeline reports are built from deals. Here are seven checks, the five deal fields that break forecasts and a one week plan to fix them in any CRM.

A coaching marketplace we work with had 1,523 opportunities in its CRM, and 99% of them never moved past stage one. No pipeline total built on that list could be believed, and it says as much about the stage rules as about the deals. A contact cleanup would not have changed any of it.

We read the five guides that kept coming up for this search, from HubSpot's blog, Truelist, New Breed, Zapier and Intelemark. All five clean contacts. None covers stale deals, close dates or stage definitions, and pipeline reports are built from deals.

This plan works on deals, in any CRM or even a spreadsheet. Where HubSpot behaves in a documented way we say so, and five quick checks show how your own CRM behaves. Pick your version first:

• You sell most deals yourself, or have under about 50 open deals. Run the seven counts and the Day 4 triage in one afternoon (3 to 4 hours, our estimate), then scrub for 30 minutes a week. Since nobody checks you, have a co founder or board member pick 5 deals a month and ask for the evidence.
• You have 3 to 15 reps. Follow the one week plan.
• You have more reps or over 800 open deals. Same plan, with more time on Days 3 and 4.

Check in 15 minutes whether your pipeline report is lying today

Counts 1 to 6 come from one export of open deals. Count 7 also needs your invoice list. The healthy lines are our judgment calls; we found no published benchmarks, so adjust them after your first month.

1. Open deals with a close date before today. The close date is the day the deal is expected to sign. Healthy: zero. Each one is a deal nobody looked at after its date passed.
2. Open deals whose close date is the last day of the month they were created in. That is the mark HubSpot's default leaves when a rep skips the date. By chance about 3% of dates land on a month's last day (12 of 365); a much bigger share is a prompt to ask owners about 10 of them, not a verdict, since real deals bunch at month end too.
3. Open deals with no activity inside your stale window. The stale window is how long a deal can go without buyer contact before it counts as stale. Healthy: under 10% of open deals. Day 2 shows how to set it.
4. The most repeated amount. Healthy: no single figure on more than 20% of open deals. Above that it is probably a placeholder (a value typed only to fill the field).
5. Deals older than your stale window still sitting in stage one. Healthy: under half. Above that, stage one is a parking spot, not a stage.
6. Two open deals on one company with similar amounts, whoever owns them. Healthy: zero pairs. Each pair counts the same money twice.
7. Closed won last quarter against invoices issued. Healthy: every closed won deal matches an invoice within a month. If most revenue is repeat orders from existing customers, give those their own pipeline or match invoices by customer.

If three or more counts land outside the line, treat every number in the report as the most it could be, not what it is, until the cleanup is done.

Five fields decide whether your forecast is true

A forecast is a sum. Each deal's amount is multiplied by the probability of its stage (that product is the weighted amount), deals are filtered by close date into a month or quarter, and totals are split by owner. Most numbers in that sum are fields someone typed.

Two HubSpot defaults matter before you touch anything. Per its default deal properties documentation, updated 29 September 2026, a deal created in an open stage without a close date gets the last day of the current month. And when a deal is marked won or lost, HubSpot writes today's date into its close date. Checks 1 and 2 tell you whether your CRM does the same.

Owner breaks quietly. Deals still owned by people who left belong to nobody, so nobody updates them.

Last activity date is the least biased of the five, because the CRM fills it when a call, email or meeting is logged. But reps must log their calls, and if an email sequencing tool (software that sends automated follow ups, such as Apollo or Smartlead) logs its sends against deals, a dead deal looks alive. On Day 1, also count staleness from the last buyer reply.

Stage probability is the input nobody checks. HubSpot's pipeline settings put one on each stage. Replace the defaults with your own win rate by stage; the questions at the end cover small numbers.

Stages named after what the seller did inflate the pipeline

Most pipelines we inherit have stages like Discovery done, Demo done, Proposal sent. Each records something the seller did. None records whether the buyer moved, so the stage advances, the probability climbs and the forecast grows on seller activity alone.

Dave Kellogg, who writes about SaaS sales on his blog Kellblog, makes the fix simple: every stage should be a buyer milestone, something a manager can check by asking the buyer one question. Our rule adds proof. A deal leaves a stage only with a dated note, an attached email or a booked meeting, never the rep's say so, and that includes a dropdown the rep ticks. Written at the time, a dated note takes a minute or two (our estimate).

If deals start from an inbound request (a demo request, a trial or a quote form), four stages are enough. Request received, with first contact logged and timed. Offer sent, with the dated quote or proposal attached. Offer reviewed, with a buyer reply. Order or signature, with a purchase order number or signed date. Add a field for minutes from request to first contact and report it by rep; our speed to lead post covers what a reasonable reply time looks like.

A one week CRM data cleanup plan a small team can run

The owner is whoever will run the weekly scrub afterwards: a founder, a head of sales or an operations lead. For 3 to 15 reps and a few hundred open deals, we estimate 19 to 22 owner hours plus 30 minutes per rep; we found no published benchmark. At larger volumes, budget about a minute per flagged deal on Day 4 (our estimate), so 800 flagged deals is roughly 13 rep hours.

If you use HubSpot, read the box below before Day 1. If you use another CRM, skip it; your answers to the five checks tell you which of these traps apply.

Day 1: freeze the data and take a baseline

Export every deal, company and contact with its record ID (the unique number your CRM gives each record; keep that column in the export), and store the file where nobody edits it. Pick 100 deals: 70 at random and the 30 oldest, which have had longest to go stale. Save the 100 record IDs and score the two groups separately: the random 70 estimates the whole CRM, the oldest 30 shows the worst.

HubSpot users then switch off close date automation for the pipeline (its docs, updated 9 June 2026, say every plan has the switch) and pause workflows that fire on field changes. Last, tell whoever reads the forecast (your board, if you are the founder) that the number is about to drop, and why.

Day 2: write the definitions

Two hours with whoever sells. Agree the new stages and their exit evidence, the close date rules below and the $0 amount rule. Build dropdowns for lost reason, lead source, industry and country, because free text cannot be counted. Add a number field called times pushed.

Then write down one stale window and use it everywhere. It is 60 days, or your median sales cycle if that is longer, capped at 120 days (our line). Your median sales cycle is the middle value of days from create date to close date across deals you won in the last four quarters: export them, subtract the dates, take the middle one. A firm whose won deals take a median of 150 days uses 120. If a buyer gives a dated reason to wait (a board meeting, a contract renewal, a budget year), record that date on the deal; the deal is not stale before it.

Day 3: merge duplicates in the right order

Companies first, then contacts, then deals, using the match keys in the next section. Test 20 pairs by hand before any bulk merge, and log the counts. Then read 20 of the pairs your rules rejected. A rule that is too strict looks exactly like one that works until you read what it threw out.

Day 4: triage stale deals

Build one view: open deals with a past close date, no activity inside the stale window, or times pushed of two or more. Each rep decides every deal on it within 48 hours of it being flagged: redate it with a date the buyer gave, move it back a stage, move it to nurture (a status outside the forecast for buyers who are not ready), or close it lost with a reason.

Before any deal closes, the rep writes the date and sender of the last buyer reply on it. If your CRM logs email, the owner can confirm by searching the log for the buyer's domain; nobody opens a colleague's inbox. Deals with a buyer reply inside the stale window stay open. The rest close lost as "Stale: cleanup", after a second person reads the list.

Day 5: standardize, repair and fix lead source

Move free text into the new dropdowns. Set missing amounts to $0. Reassign deals owned by people who left. If a tool such as Clay (a data enrichment tool), Apollo or a sequencer writes into your CRM, it should match on record ID rather than create new records; the plan needs none of them. Nothing here emails anyone. Adding outside contact data, or emailing law firm or healthcare contacts, raises consent and privacy questions this post does not cover and is not legal advice on; take those to counsel first.

Lead source goes on every deal, because it shows which work pays. For referrals, use a dropdown of referrer types (existing client, partner firm, event, website), name the referring person or firm, and report closed revenue by referrer each quarter. For ads, copy the form's campaign fields onto the deal, set blanks to Unknown, and compare 90 days of closed won revenue by campaign with its spend. Under about 10 closed deals, a source's number is a hint (our line). Our post on offline conversion tracking covers sending those sales back to Meta and Google.

Day 6: lock the door

Four controls, in plain words. Use your CRM's settings if it said yes to checks 4 and 5; otherwise do them by hand in the weekly scrub.

• Required fields at stage change: close date with its source from Budget confirmed onward, amount above $0 from Proposal reviewed, a next step date at every open stage, and lost reason when a deal is lost.
• Pipeline rules, so nobody skips a stage or moves a deal backwards without a reason.
• Dropdowns for every field a report groups by.
• A stale deal task. When a deal passes the stale window, a task goes to the owner. For the first quarter it stays a task, never an automatic close. Read a month of what it flags and count the false alarms, meaning healthy deals it should not have flagged. If more than about 1 in 10 flags are false alarms, lengthen the window before anything closes on its own (our line).

Required fields invite junk typed to get past them; the weekly report's five most repeated values show it.

Day 7: score the same 100 deals again

Rescore the 100 record IDs you saved on Day 1, so before and after compare the same deals. Score a merged record as the one you kept and a closed deal as it now stands. Rerun the seven counts, present the smaller number, and book the weekly 30 minute scrub before that meeting ends.

How to score the sample

Score each line on its own. An average hides the failure that matters.

• Completeness: sampled deals with every required field filled, divided by deals sampled.
• Uniqueness: one minus duplicates divided by total records.
• Staleness: open deals with no activity inside the stale window, divided by open deals.
• Past due: open deals with a close date before today, divided by open deals. The target is zero.
• Accuracy: fields confirmed against an outside source (invoice, buyer email, LinkedIn), divided by fields checked. 19 right out of 20 reads as 95%, but a sample that small fits anything from about 76% to 99% (a Wilson 95% confidence range, the band the true figure very probably sits in; our arithmetic). At 95 out of 100 it narrows to about 89% to 98%. Report the range with the sample size.
• Perfect record: sampled deals with no error in any of 10 to 15 key fields. Thomas Redman, a data quality consultant, calls this the Friday Afternoon Measurement; Salesforce Ben published a one hour version for CRMs.

Merge duplicates in order: companies, then contacts, then deals

Match keys are the fields that decide two records are the same person, company or deal. Order matters, because duplicate companies hide duplicate deals: two deals on two copies of one company never sit side by side in a report.

Per HubSpot's duplicates documentation, updated 25 September 2026, its duplicate tool covers contacts and companies, not deals. Detection needs a Professional or Enterprise plan (HubSpot's two top paid levels), and custom rules (up to two per object, using up to nine fields) need Data Hub, a paid add on whose pricing page listed Professional at $720 to $800 a month as of October 2026. In GoHighLevel (a CRM and marketing platform), a setting called Allow Multiple Opportunities per Contact decides whether one contact can hold two deals in the same pipeline, per its help article. If yours allows it, duplicates can build up there too.

In any CRM, deals need a spreadsheet, because the comparison runs row against row:

1. Export open deals with record ID, company, pipeline, owner, amount and create date.
2. Sort by company, then pipeline, then create date.
3. Flag any row that matches the row above on company and pipeline, was created within 30 days of it, and has an amount within 5%. Put pairs with two different owners at the top, since two reps on one account is an easy way to count one deal twice.
4. Open each flagged pair, read the notes, and merge by hand.

HubSpot states that merged records cannot be unmerged, so export the IDs first and test 20 pairs. If you are still choosing a CRM, our GoHighLevel vs HubSpot comparison covers cost and setup side by side.

Close dates and stale deals need written rules, not reminders

A close date should be the date the buyer said they will decide or sign. Two things pull it elsewhere: a CRM default and the date the rep needs for quota.

The push count is our version of Kellogg's rolling hairball test, his name for a deal whose close date keeps sliding into the next quarter: two quarters of slipping makes it suspected, three or more confirmed. Counting pushes works for any cycle length. Kellogg also argues for creating every opportunity at zero dollars until a value has been discussed with the customer, because a placeholder adds money no buyer agreed to. The 1.5 times cap and the 48 hour window are our rules of thumb.

Closing stale deals lowers your win rate, and the lower number is the true one. Take 10 deals (invented counts, to show the arithmetic): 3 won, 2 lost, 5 open, of which 3 have had no buyer contact for months. The win rate on closed deals is 3 of 5, or 60%. Close the 3 stale deals and it is 3 of 8, or 37.5%. The buyers decided those deals when they went silent. Give the board 37.5%, with the cleanup closures listed apart.

The rules reps work under decide whether the data stays clean

A CRM holds what sellers typed, shaped by whatever they are measured on. The five guides we read put dirty data down to user error and answer it with an owner, a schedule and automation. None of that touches why the data went bad.

The person who creates a record never pays for its errors. Redman, in Harvard Business Review in 2016, calls the result a hidden data factory: one team creates errors and the next adds steps to cope. In a CRM, reps create records and the founder lives with them.

For the rep, an open deal costs nothing to keep alive. Closing one lowers pipeline coverage (open pipeline divided by the target), which is often what the manager is judged on. Kellogg also describes reps who create far future opportunities to keep an account from being reassigned.

The one survey we found that asks about this directly measures belief, not records. In late 2021 Validity, a vendor of CRM data tools, surveyed 1,241 CRM administration workers online in the US, UK and Australia. 75% said employees fabricate data "to tell the story they want decision makers to hear". 82% said they are asked to find data that supports a specific story rather than accurate data, while 76% rated their own data good or very good.

The sociologist Donald Campbell described the pattern in "Assessing the impact of planned social change" (written 1976, published in Evaluation and Program Planning in 1979): the more a number drives decisions, the more pressure there is to corrupt it. Reps respond sensibly to the rules they work under, so change the rules.

What to change, in order

Where each Day 6 control stands in HubSpot and GoHighLevel:

The weekly hygiene report

One page, every Monday, with the stale window printed at the top. The thresholds are ours, not industry benchmarks.

• Open deals with a past close date: 0.
• Open deals with no activity inside the stale window: under 10%.
• Deals at or past Proposal reviewed with no amount: 0.
• New duplicate pairs: under 5.
• Deals owned by people who have left: 0.
• Deals with lead source Unknown, and the five most repeated amounts.
• Error counts by rep.

Once a month, match closed won in the CRM against invoices issued. Every closed won deal without an invoice needs a name and a reason.

The famous bad data numbers measured something else

Gartner's $12.9 million a year comes from its Magic Quadrant for Data Quality Solutions, Gartner's ranking of vendors in that category. Gartner's own page returned an error to us; the copy we read carries Gartner's title, the date 27 July 2020, document number G00389794 and analysts Melody Chien and Ankush Jain, on a whitepaper site under a file name suggesting IBM handed it out. Its survey covered reference customers (customers the vendors put forward), 154 organizations tied to 16 vendors, and how many answered the cost question is not stated. A 20 person firm does not lose $12.9 million.

IBM's $3.1 trillion a year first appears in a 2011 press release quoting an analyst, Hollis Tibbetts, who called it "my estimate" of what dirty data cost the US economy and gave no method. Redman repeated it in HBR in 2016 as IBM's estimate for that year (HBR is paywalled, so we read a PDF copy hosted by Data Ladder, a data vendor). The 1-10-100 rule, a dollar to prevent a bad record against a hundred once it fails, is a rough rule about scale from a 1992 book, not a measured cost.

Decay rates measure contact details going out of date. The one everyone repeats is 2.1% a month, about 22.5% a year (1 minus 0.979 to the power of 12), from HubSpot's database decay page, which credits undated research by MarketingSherpa, a marketing research publisher, on email lists. The 70.3% a year in a 2024 Forbes Business Council post credits Gartner but links to a general Gartner page with no study on it. A seminar survey of 1,025 businesspeople found 70.8% had changed something on their business card in a year, mostly job title, and 34.2% had changed employer. Decay changes who you can email. It does not change the deal amounts in a forecast.

The most carefully measured figure comes from three data quality researchers, Nagle, Redman and Sammon, who reviewed 75 data quality assessments in the journal Business Horizons in 2020 (via RePEc, a free index of economics research). Averaged across the 75 assessments, 47% of recently created records had at least one critical error. The abstract does not define critical; to test it on your CRM, use the perfect record score from the plan. The study counted a score as acceptable only at 97% correct or better, and just 3% of scores met that bar. The companies chose to be assessed, so it is not a random sample, but new records are what your reps create every day.

• 47% of recently created records had at least one critical error, averaged across the assessments
• 3% of quality scores met the study's bar of 97% correct or better
• 75 data quality assessments reviewed, chosen by the companies rather than at random

Source: Nagle, Redman and Sammon, Business Horizons 63(3), 2020

What it costs, and when you do not need us

Doing it yourself costs the owner's time: 19 to 22 hours at your own hourly rate for 3 to 15 reps, or an afternoon if you sell most deals yourself. You need no new software; your CRM plus a spreadsheet covers every step.

We are Contwre, a GTM agency in Lahore founded in 2020, and we build revenue systems for B2B companies: outbound, paid acquisition, CRM and the sales process around them. Our public price ranges are consulting from $150 an hour, done with you projects from $700 to $1,300, and retainers from $1,000 to $12,000 a month, most between $3,000 and $4,000. Our case studies show the work with the numbers attached.

Questions people ask

How often should you clean CRM data?

Check the five deal fields weekly in a 30 minute scrub, and clean duplicates and out of date contacts monthly. Quarterly is too slow for close dates, because a deal can sit past due for weeks before anyone finds it.

What is the difference between CRM data hygiene and CRM data quality?

Hygiene is the recurring work: removing duplicates, fixing dates, closing stale deals. Quality is the state it produces, measured with completeness, uniqueness, staleness and accuracy scores. Judge the hygiene by whether those scores move.

Does this plan work in GoHighLevel, Pipedrive or a basic CRM?

Yes. The counts, stages and rules use plain fields every CRM keeps on a deal, and a spreadsheet export covers anything your CRM cannot automate. What differs is the defaults, so run the five checks near the top of this post first and use your answers wherever the plan mentions HubSpot.

How do you count how many times a close date was pushed?

Add a number field called times pushed. If your CRM has automation, a workflow adds one each time the close date moves later. If not, the rep updates it during the weekly scrub. Test the count on 10 deals before any rule depends on it.

How do you set stage probabilities with only a few deals?

Use your own history, but treat small numbers as a range. If 12 deals reached the proposal stage last year and you won 4, that reads as 33%, but with 12 deals the true rate could sit anywhere from about 14% to 61% (a Wilson 95% confidence range, our arithmetic). If your CRM does not keep stage history, start recording it now and use your overall win rate until you have a few quarters.

How should deals at $0 show in a forecast?

They add nothing to the weighted total, by design, because no buyer has seen a price. Report their count on its own line, for example 14 early deals with no price yet, so the board sees the activity without counting money nobody has agreed to.

What should you show the board after a cleanup?

Three lines. Deals with buyer evidence closing this quarter against the target. The cleanup closures, shown apart from real losses. And the old and new pipeline totals side by side with the cleanup date, so the drop reads as a correction, not a bad month.

How much does bad CRM data cost?

No credible single number exists. Gartner's $12.9 million is an estimate by data quality vendors' reference customers, and IBM's $3.1 trillion began as one analyst's estimate for the US economy. Match closed won deals against invoices to see your own gap.

Export your open deals today and count the ones whose close date has already passed. If that number is not zero, fix it before anyone presents a forecast built on those deals.

If you want the stages, rules and reports rebuilt with you, talk to us.

Book a call with Contwre · More from the blog