The Corporate-Family Rebuild Teardown: CRM Accounts to a Sourced Parent Tree
Take one corporate group's CRM records, pull its LEI, Exhibit 21, and UK register tree, label every record, and write back sourced, dated parent links.
Key takeaways
- In the LEI file, reporting exceptions outnumber parent relationship records about 9.3 to 1, so a null parent is the expected result and the exception reason code is the field to read.
- Apple's LEI family has 20 ultimate children but only 8 direct children, so pulling direct children from the top entity alone misses 60% of the family.
- Exhibit 21 may legally leave out subsidiaries that together would not be significant, so an entity missing from the exhibit can still belong to the group.
- UK PSC dates are notification dates, which can trail the real control change by up to 14 days under the current rule and up to 28 days under the old 14+14 rule.
- Build the legal tree before merging any account records, because merging a local legal entity into its brand record destroys the registration identifier the tree depends on.
- Salesforce's ultimate-parent formula only goes 10 levels deep, so deep multinational groups break the stamp even though the hierarchy itself has no hard depth limit.
This guide covers one job: rebuilding a single corporate group's parent-child tree in a CRM from public evidence. It is for the RevOps, recruiting-ops, or data owner who has dozens of account records filed under brand names, local legal names, and acquired-company names, and who will be asked where each link came from. I carry one real group's public LEI family through the whole rebuild, show the forks and wrong turns at each step, and give you a count log to fill in for your own family.
What is a sourced corporate-family tree?
A sourced corporate-family tree is a set of parent-child links between CRM account records where every link names the public record it came from and the date that record was true. A link that cannot cite a source does not go in the tree. It stays in a labelled holding state until it can.
Terms used throughout:
- Direct parent: the entity that immediately consolidates or controls this one.
- Ultimate parent: the entity at the top of the chain, which nothing above it consolidates.
- LEI: the Legal Entity Identifier, the global identifier for a legal entity, published by GLEIF.
- Level 2 data: the "who owns whom" layer of the LEI system. Each entity reports its direct accounting consolidating parent and its ultimate accounting consolidating parent.
- Reporting exception: the record an entity files in place of a parent relationship, with a reason code explaining why no parent is reported.
- Exhibit 21: the subsidiary list a US SEC registrant attaches to its annual report.
- PSC register: the UK register of people with significant control, held at Companies House. It can name a corporate controller.
- Brand-only record: a CRM account named after a product or trading brand that is not a legal entity.
By the end of the rebuild, every CRM record carries exactly one of four labels: matched to a legal entity, brand-only, unmatched, or insufficient evidence. Parent links get written as fact only for matched records.
This guide builds the legal tree and nothing else. Whether sales covers a subsidiary as its own account or folds it into the parent's territory is a separate decision, covered by the Subsidiary Ownership Verdict. Keep the two apart. Once a legal tree gets bent to fit territory politics, it stops being evidence.
Which public sources hold parent links, and how does each one lie?
Most parent links come from three free public sources: LEI Level 2 records, the Exhibit 21 subsidiary list, and national registers that name corporate controllers, such as the UK PSC register. Each one proves something narrow, and each one fails in its own way, either by staying silent or by running late.
| Source | What it proves | How it lies |
|---|---|---|
| LEI relationship record | Reported accounting parent, LEI to LEI | Rare; most entities file an exception |
| LEI reporting exception | Why no parent is reported | Read as "no parent" when it means "not reported" |
| Exhibit 21 | Registrant lists it as a subsidiary | Non-significant subsidiaries may be omitted |
| UK PSC (corporate) | A UK company controls this one | Notification dates; UK controllers only |
| UK company profile | Branch's parent company number | Covers branches only |
LEI Level 2: exceptions are the normal output
Every LEI entity except branches must report both parent relationships, and each one is filed either as a relationship record or as a reporting exception. Exceptions cover three cases: the entity has no parent, the child opts out for exceptional reasons, or the parents have no LEI. The reason codes include:
- NO_LEI
- NATURAL_PERSONS
- NON_CONSOLIDATING
- NO_KNOWN_PERSON
- LEGAL_OBSTACLES
Exceptions have outnumbered relationship records in every public snapshot of the files:
| Snapshot | Relationship records | Reporting exceptions | Exceptions per record |
|---|---|---|---|
| Feb 2018 | ~113,000 | ~1,368,000 | 12.1 |
| Jul 2026 | 481,845 | 6,258,523 | 13.0 |
| Nov 2026 | 678,859 | 6,317,492 | 9.3 |
The July and November rows come from different file products (the golden copy and the concatenated file), so the change between them is not a trend. The July snapshot also counts 3,384,122 legal entities. That gives each entity two parent slots, and 6,258,523 exceptions fill roughly 92.5% of them. The figure is approximate because branches sit in the denominator. The LEI Regulatory Oversight Committee once reported an LEI-to-LEI parent for 5 to 6 percent of entities. No current official percentage exists, so if you need one, derive it from the latest file counts.
Here is why: any entity whose parent has no LEI must file an exception. Expect most of your tree to come back null on the first pass. The reason code is the field that tells you where to look next.
To pull the data you can use the GLEIF API (documented at api.gleif.org/docs), the golden copy downloads, the gleifr R client, or the gleif-mcp-server. The API includes endpoints for direct-parent-reporting-exception, ultimate-parent-reporting-exception, and ultimate-children, among others.
Exhibit 21: incomplete by design
Filers may leave out subsidiaries that, considered in the aggregate as a single subsidiary, would not be a significant subsidiary. "Significant subsidiary" is defined in Rule 1-02(w) of Regulation S-X. Some exhibits state the omission basis outright. Greif's, for example, states its omission threshold as of its fiscal year end. An entity missing from Exhibit 21 tells you nothing about whether it belongs to the group.
UK PSC: one hop at a time
A PSC entry whose kind is corporate-entity-person-with-significant-control names a company as the controller. That company's number sits in identification.registration_number, so you can follow the chain upward one hop at a time. The dates need care. ceased_on is the date Companies House was notified of the cessation, not the date control changed. PSC information is now held only on the central register, and changes and cessations must be reported within 14 days. The earlier "14+14 rule" allowed up to 28 days. Check Companies House guidance for the rule currently in force. This route only works when the controller is UK-registered. For branches, the company profile carries branch_company_details.parent_company_number.
Worked example: what does Apple's LEI family actually contain?
Apple's LEI family has 20 ultimate children but only 8 direct children. The reason is that intermediate holding companies such as Apple Operations International sit between Apple Inc and many of the entities below it. I use Apple here because those counts are public and you can check them. Run the same pulls on your own group and log your numbers next to these.
Apple's LEI family, outermost first
- Ultimate parentApple Inc, the top of the LEI tree
- Intermediate holding companyApple Operations International, a direct parent that differs from the ultimate parent
- LEI family20 ultimate children, of which 8 are direct children of the top entity
- CRM recordsMatched, brand-only, unmatched, or insufficient evidence
Fork 1: which LEI is the anchor?
A legal-name search returns the top company and also subsidiaries with similar names. The anchor is the entity with nothing above it. Prove that by reading its own ultimate-parent slot. If the slot holds an exception, read the reason code. NO_LEI and NON_CONSOLIDATING mean a parent was not reported as an LEI relationship. They do not mean the entity has no parent.
Wrong turn: anchoring on Apple Operations International because it shows up as the parent on many children's records. It is a direct parent, and its ultimate parent is Apple Inc. Anchoring there would put Apple Inc outside its own tree. If an entity appears as a parent across many records, check whether it has an ultimate parent above it before treating it as the top.
Fork 2: direct-children or ultimate-children?
Wrong turn: calling direct-children on the anchor and stopping there. For Apple that returns 8 entities. Ultimate-children returns 20. Stopping at direct-children misses 12 of the 20, or 60% of the LEI family. The correct pull uses ultimate-children to set membership, then reads each member's direct parent to rebuild the structure inside the tree.
Fork 3: where Exhibit 21 and the LEI list disagree
I did not retrieve Apple's Exhibit 21 count for this guide, and nobody has published a typical exhibit-versus-LEI comparison, so I am not printing a number. What the diff produces is three buckets, and each bucket tells you something different:
- In both: the strongest evidence of membership. Two independent sources agree.
- LEI only: expected, because the exhibit may legally omit non-significant subsidiaries.
- Exhibit only: the entity either lacks an LEI or filed an exception. Go to the national register for its parent link.
Fork 4: CRM records against the tree
CRM-side counts describe your data, not public facts. No published benchmark says how many CRM records a group of a given size breaks into, or how often those records are misparented. The published reference points are broad. One vendor case study found 38% of 514K CRM records were duplicate or fragmented entities, and one data-cleanup guide recommends an alert threshold of 3% duplicates on accounts. Record your own counts in the count log in the procedure section. If you take your duplicate rate from the export snapshot, it is the baseline you will measure the rebuild against.
A null parent in the LEI file is a missing answer, not an answer.
Who owns a corporate-family rebuild?
Two functions usually share this job. RevOps owns the CRM fields, routing, and territories. A master-data or data-stewardship function owns legal-entity data. Name an owner on each side before Step 1, because the person who signs off the legal tree and the person who signs off the write-back are rarely the same person.
| Segment | Count | Most common title (sample of 25) |
|---|---|---|
| US RevOps | 2,710 | Director of Revenue Operations (18/25) |
| UK RevOps | 346 | Revenue Operations Manager (14/25) |
| US Master Data / Data Steward | 1,812 | Master Data Analyst (9/25) |
In Refolk's index, US RevOps titles outnumber UK RevOps titles 7.8 to 1, and outnumber US master-data titles 1.5 to 1. In the sample, master-data roles cluster in regulated enterprises such as AbbVie and JPMorganChase. My inference, which the index alone does not prove: in many companies the legal-entity data and the CRM fields sit in different functions, and that is why rebuilds stall at the handoff. The title breakdowns come from a sample of 25 per segment, so treat them as indicative.
You may not have the second function in-house. If you need to hire a steward, or borrow a peer who has already rebuilt a hierarchy after an acquisition, Refolk takes that background as a plain-English request. You don't have to piece together title variants and keyword strings across LinkedIn, GitHub, and the open web. Refolk returns the people, and you can check their history against the job before you reach out.
The rebuild procedure, step by step
The procedure runs in eight steps: freeze the records, build the legal tree from public sources, label every CRM record against it, then write back and log. Merging comes after the tree is built, never before.
Corporate-family rebuild
- Freeze and export the account setExport a CSV snapshot of every candidate account with record IDs, names, domains, countries, and current ParentId. Make no edits to these records until write-back.
- Resolve the anchor LEISearch LEI records by legal name and country and confirm the top entity by reading its own ultimate-parent slot. About 15 minutes.
- Pull the LEI family from the topCall ultimate-children for membership and direct-children for structure, and record every child's direct parent and exception reason code. Under an hour.
- Diff Exhibit 21 against the LEI listCompare the parent's subsidiary exhibit entity by entity with the LEI list and count in both, LEI only, and exhibit only. One to two hours.
- Chain UK controllers through the PSC registerFor UK entities without an LEI parent, follow corporate PSC entries upward by registration_number, logging notified_on at each hop. About 30 minutes per chain.
- Match and label every CRM recordMatch each record using the strongest key available, or label it brand-only, unmatched, or insufficient evidence. A reviewer signs off every match below the top key tier.
- Write back links with source and as-of dateWrite direct parent, ultimate parent, source, and as-of date onto every matched record, then recalculate the ultimate-parent stamp for the whole family.
- Log counts and schedule the refreshComplete the count log for every step, store it with the snapshot, and put a dated refresh on the calendar with an owner.
Timing: steps 1 to 5 have published or practical estimates, given above. Step 6 does not. The only published manual benchmark is about 30 duplicate records merged per hour. That is a merge rate, not a match-and-label rate, so treat it as a floor on effort rather than a forecast.
One published cleanup sequence runs the other way: deduplicate, enrich, validate, govern. That order suits contact data. For corporate families it breaks things, because the duplicates you would merge are often different legal entities that share a brand. Build the tree, label every record, and only then merge records that share the same legal identifier.
Family: Snapshot file | row count | exported by | export date: Anchor LEI | ultimate-parent slot: relationship / exception (reason code): LEI ultimate-children of anchor: LEI direct-children of anchor: Exception reason codes seen: NO_LEI __ | NATURAL_PERSONS __ | NON_CONSOLIDATING __ | NO_KNOWN_PERSON __ | LEGAL_OBSTACLES __ Exhibit 21 entities: __ | in both: __ | LEI only: __ | exhibit only: __ UK PSC chains followed: __ | total hops: __ CRM records in scope: __ matched to a legal entity: __ brand-only: __ unmatched: __ insufficient evidence: __ matched but sources conflict: __ Maximum hierarchy depth: __ (flag if over 10) Records written back: __ Records grouped by ultimate parent match the tree: yes / no Next refresh due | owner:
One log per family. Fill every line before write-back and store it beside the export snapshot.
How do you label each CRM record against the tree?
Match each record on the strongest identifier it carries, and treat any match below the top tier as a candidate for review, not an automatic link. A parent link needs a legal identifier and a source. A name or domain alone justifies a review, never a link.
My key ladder, strongest first:
- LEI on the record or found for the entity.
- National registration number, for example a Companies House number.
- Exact legal name plus country, confirmed against a register entry.
- Web domain. It points to a brand, not to a legal entity.
- Name stem. Useful only for finding candidates.
Anything at rungs 3 to 5 goes to a reviewer, who either finds a rung 1 or 2 identifier or labels the record insufficient evidence.
Label decision for one CRM record
What each label proves, and how it lies:
- Matched: the record is a specific legal entity in the tree. It lies when the identifier was copied from a same-name company in another group.
- Brand-only: the record names a trading brand with no legal entity behind it. It lies when the brand is also the name of a small local entity that never got an LEI. Check the national register before you settle on this label.
- Unmatched: no entity in the tree fits. It lies when the entity is real but appears only in an exhibit or register you have not pulled yet.
- Insufficient evidence: a plausible match exists but cannot be proven. This is the honest label for "exhibit only" entities until LEI or register data confirms them.
Record ID: Label: matched / brand-only / unmatched / insufficient evidence Legal identifier: LEI or national registration number Direct parent record ID: Direct parent source: LEI relationship / Exhibit 21 / UK PSC / UK company profile Direct parent as-of date: Direct parent date type: file publication / annual report / register notification Ultimate parent record ID: Ultimate parent source: Ultimate parent as-of date: LEI exception reason code (if any): Joint venture: yes / no | other owner and source: Reviewer and review date: Conflict note:
Map each line to a custom field. Source and as-of date belong on every link, not only the top one.
What does the CRM do to the tree once you write it back?
Neither major CRM stores an ultimate parent natively, and both have limits that can quietly distort a correct tree. Know them before you write back, because the failures look like data errors rather than system limits.
| Capability | Salesforce | HubSpot |
|---|---|---|
| Direct parent | Parent Account lookup | One parent per child |
| Ultimate parent | Custom field, not native | No global parent property |
| Children and depth | No hard depth limit; View Hierarchy shows 2,000 | 10,000 children per parent; multi-level |
| Roll-up | Not native for custom stamps | None |
Salesforce
The Parent Account field holds only the direct parent. Ultimate parent is a separate field, usually custom and usually filled by a formula. The hierarchy has no hard depth limit, but the formula approach only climbs 10 levels, so deep families break the stamp while the tree itself stays intact. An ultimate-parent formula that returns an ID gives a 15-character ID. CASESAFEID can't be used because it hits the formula size limit, so any join on that field has to handle the 15-character form. LeanData reports that the Account object allows 10 unique cross-object relationships across formulas, rules, and lookup filters, so an ultimate-parent formula competes with your other automation for that budget. LeanData also documents what breaks when accounts are reparented. I did not retrieve a Salesforce primary document on reparenting behaviour, so test a reparent in a sandbox before you trust it.
HubSpot
A child company can have only one parent, and one parent can have up to 10,000 child companies. Contacts, deals, and activities do not roll up to the parent, so any reporting by group needs its own logic. One behaviour works directly against a rebuild: a contact whose email domain matches is associated with the oldest company that has that domain. If an old brand-only record shares a domain with the legal entity you just matched, new contacts can attach to the brand record without anyone noticing.
How this goes wrong: failure modes and false positives
Most failed rebuilds come from eight recurring mistakes. Each one produces a tree that looks finished and is wrong. Every item below gives what the mistake looks like and the check that catches it.
- A null LEI parent read as "independent". It looks like a subsidiary sitting alone at the top of its own one-node tree. The check: read the exception reason code for every entity you are about to treat as a top, and only accept it after another source agrees.
- Exhibit 21 treated as complete. It looks like a real subsidiary labelled unmatched because the exhibit left it out. The check: an entity missing from the exhibit stays insufficient evidence, not unmatched, until LEI or register data settles it. Absence from the exhibit is allowed by the rules.
- PSC dates read as event dates. It looks like an acquisition that seems to happen weeks after it closed, so routing and territory changes land late. The check: store notified_on as a notification date and compare it with the acquirer's announcement date. The gap can be up to 14 days now and was up to 28 under the 14+14 rule.
- Attaching a record to a look-alike parent. It looks like a shared name stem or domain pulling a record into a different group with a similar name. The check: no parent link without a registration number or an LEI hop. Remember HubSpot's oldest-company domain rule, which can attach contacts to the wrong company on its own.
- Merging before the tree is built. It looks like a local legal entity merged into its brand record, its registration ID gone. The check: after any merge, confirm every surviving record that represents a legal entity still carries its legal identifier.
- The formula silently truncates. It looks like a blank, or "beyond 10 levels", in the ultimate-parent stamp on the deepest records. The check: compute the maximum depth of every family and flag any family deeper than 10.
- Reparenting without a restamp. It looks like descendants still carrying the old ultimate parent after a branch is moved. The check: count records grouped by ultimate parent and compare the count to the tree. Any mismatch means a stale stamp.
- Joint ventures forced under one owner. It looks like a 50/50 venture attached to whichever owner was linked first. The check: label it a joint venture, store both owners with their sources, and record which one, if any, the native parent field holds and why.
Each of these has a matching false positive. A tree with no nulls, an exhibit that matches the LEI list exactly, or a family with no unmatched records should make you suspicious, not relieved. Real groups produce exceptions, omissions, and leftovers, and an output with none usually means one of the checks above was skipped.
Keeping the tree current after the first rebuild
A sourced tree starts going stale the day it is written, because every source behind it changes on its own schedule. Re-pull from the sources on a fixed cadence rather than editing links by hand. Most drift comes from small manual fixes.
Practical rules for the refresh:
- Re-pull the LEI family in full for each anchor rather than diffing cached data. GLEIF adds relationship data quality checks over time, so the same family can come back with different relationships or exceptions.
- Re-read exception reason codes. An entity that filed NO_LEI may later report a relationship once its parent registers.
- Re-check the subsidiary exhibit each time the parent files a new annual report, and note the omission basis it states.
- Re-check the PSC rule in force before you interpret notification gaps, because the reporting window has changed before.
- Rerun the grouped count after every reparent, merge, or acquisition. The count from step 7 is your regression test.
When the refresh turns up a new acquisition, run the procedure again from step 1 for the acquired company's records, then attach its anchor under the acquirer with the source and date of the link. If the job needs more hands than you have, the two-function split above is the hiring brief, and Refolk will find people with exactly that background when you describe it in a sentence.
Before you call the family done
- The export snapshot exists and no record in it was edited before write-back.
- The anchor LEI's own ultimate-parent slot was read and its reason code is logged.
- The LEI family was pulled with ultimate-children, not direct-children alone.
- The Exhibit 21 diff has counts for in both, LEI only, and exhibit only.
- Every UK PSC hop has a registration_number and a notified_on date stored as a notification date.
- Every CRM record has exactly one label: matched, brand-only, unmatched, or insufficient evidence.
- Every matched record has direct parent, ultimate parent, source, and as-of date.
- Every surviving legal-entity record still carries its LEI or registration number.
- No family is deeper than 10 levels, or the deep ones are flagged for a non-formula stamp.
- A count of records grouped by ultimate parent matches the tree.
- Joint ventures are labelled and both owners are recorded with sources.
- The count log is complete and a dated refresh has a named owner.
Questions practitioners ask
How do I find the ultimate parent company of a subsidiary using free sources?
Start with the subsidiary's LEI and read its ultimate-parent slot in the Level 2 data. If it holds a relationship record, you have the parent's LEI. If it holds a reporting exception, read the reason code and move on to the parent's annual-report subsidiary exhibit or a national register such as the UK PSC register, following corporate controllers one hop at a time. Record the source and the as-of date for every hop.
Does a blank LEI parent mean the company has no parent?
No. Every LEI entity except branches has to report both parent slots, and most fill them with reporting exceptions rather than relationships. NO_LEI and NON_CONSOLIDATING exceptions mean the parent was not reported as an LEI relationship, not that there is no parent. Read the reason code before calling an entity the top of its tree, and check the exhibit or national register before you accept it.
Should I deduplicate accounts before building the parent-child hierarchy?
Build the tree first. One published cleanup sequence puts deduplication first, but merging before the tree exists tends to fold a local legal entity into its brand record and lose its registration identifier. Without that identifier you cannot prove the parent link later. Once every record has a label and a legal identifier where one exists, merge only true duplicates that share the same identifier.
Why does my Salesforce ultimate parent formula return blank for some accounts?
The formula approach only climbs 10 levels, so accounts deeper than that come back blank or flagged as beyond 10 levels. The hierarchy itself has no hard depth limit. Check the maximum depth of each family. Also note that an ID-returning formula gives a 15-character ID, because CASESAFEID pushes the formula over its size limit. Account formulas also share a cap of 10 unique cross-object relationships.
How do I model a 50/50 joint venture in a HubSpot company hierarchy?
HubSpot allows one parent per child company, so a joint venture gets forced under one owner if you link it at all. Do not pick a parent silently. Label the record explicitly as a joint venture, store both owners and their sources in custom fields, and leave the native parent link empty or document which owner you chose and why. Remember that contacts, deals, and activities do not roll up anyway.
What as-of date should I store for a parent link taken from Companies House?
Store the notification date and label it as one. PSC notified_on and ceased_on record when Companies House was told, not when control changed. Under the current rule, changes must be reported within 14 days. The earlier 14+14 rule allowed up to 28 days. If the acquisition date matters for routing, compare it with the acquirer's announcement and store both dates.
Try it on the search you came here for
Stop building boolean strings. Just describe the person.
Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.
01Describe them
One plain sentence. Role, city, stack, stage, whatever matters to you.
02I read the web live
GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.
03You read the shortlist
Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.
- Staff backend engineers in NYC who shipped Rust in production
- Series A fintechs in SF under 50 people, growing headcount this year
- Maintainers of fast-growing Rust web frameworks on GitHub
- No boolean, no filters, no seat to buy. One box.
- Read at search time, so a profile updated yesterday counts today.
- Every step visible as it runs, every name with its reason.
500 free credits on sign-up. No card, no demo call. See real searches.