A growing database should offer more than more names.
Can knowing how people are connected help a salesperson reach the right person?
A job title helps identify someone's role. Relationship information can add context about the people around them. ExecGraph is testing whether that additional information produces better contact recommendations than profile comparison alone.
The long-term aim is practical: a platform that helps sellers understand whom to approach, who may provide relevant context, and which possible connections deserve attention. The recorded network has expanded across each study snapshot. The results below show four recorded snapshots, three fixed-reference performance checkpoints, and the separately classified September 20 diagnostic.
August 21 to October 4, 2026
The graph is compounding faster than the database.
Across the fixed-reference endpoints, the number of people grew 38.8%, while recorded relationships grew 853.5%. The relationship count is now 9.53 times the August level.
Connected coverage rose from 19.63% to 60.81%. About 61 in every 100 people had at least one recorded connection by October 4, compared with about 20 in August.
This is the compounding opportunity: the contact base is growing, but relationship context is accumulating much faster around it. That gives a model that uses connections more relationship structure to evaluate. It does not, by itself, prove that the resulting sales recommendations are better.
Four snapshots. Three fixed-reference checkpoints.
The complete table keeps September 20 visible while preserving the scientific boundary. October 4 is the third observation under the fixed reference; September 20 remains the changed-setup diagnostic.
| Measurement | * | |||
|---|---|---|---|---|
| People represented | 47,326 | 57,083 | 60,367 | 65,682 |
| Recorded relationships | 21,348 | 54,267 | 80,238 | 203,549 |
| People with at least one link | 9,289 | 17,366 | 24,477 | 39,944 |
| Connected coverage | 19.63% | 30.42% | 40.55% | 60.81% |
| Average degree, including isolates | 0.902 | 1.901 | 2.658 | 6.198 |
| GraphSAGE ROC AUC | 0.8302 | 0.8802 | 0.9153 | 0.9192 |
| GraphSAGE average precision | 0.8731 | 0.9071 | 0.9208 | 0.9271 |
| Semantic ROC AUC | 0.8824 | 0.9058 | 0.8971 | 0.8911 |
| Semantic average precision | 0.8740 | 0.9061 | 0.8977 | 0.8917 |
| Test positive pairs | 2,077 | 5,291 | 7,886 | 20,383 |
| Sampled negative pairs | 2,077 | 5,291 | 7,886 | 20,383 |
| Test setup | Fixed reference | Fixed reference | Supplementary diagnostic* | Fixed reference |
Source: reviewed ExecGraph Energy experiments. August 21, September 4 and October 4 are the fixed-reference series. *September 20 is a separately classified diagnostic shown in place for context. The semantic comparator is profile similarity, not full Warm Path V2 accuracy.
ROC AUC: fixed reference through October 4
Average precision: fixed reference through October 4
Under the fixed reference, GraphSAGE ROC AUC moved from 0.8302 to 0.8802 to 0.9192. Average precision moved from 0.8731 to 0.9071 to 0.9271. The model went from 5.22 AUC points behind profile similarity in August to 2.81 points ahead on October 4.
That crossover is a meaningful longitudinal observation because the fixed recipe did not change. It is consistent with the hypothesis that relationship information can help the model distinguish recorded connections better than profile similarity alone. It does not isolate relationship density as the cause, and it does not establish better introductions or sales outcomes.
Comparison notes and September 20 status
August 21, September 4 and October 4 form the fixed-reference performance series. The graph snapshots and held-out pairs change with the underlying database, while the recovered reference recipe, comparator and metric definitions remain fixed.
September 20 original-protocol result: not measured. Its diagnostic remains visible because it supplied additional ranking endpoints, but it is not inserted into the fixed-reference line.
No confidence intervals or independent-seed significance claim accompanies these point estimates. Score points equal a 0-to-1 metric difference multiplied by 100.
Beyond profile similarity: the shortlist is diverging further
The supplementary recipe was repeated on October 4 only after the September 20 run reproduced within the documented numerical replay contract. Both diagnostics compare GraphSAGE with the pinned Warm Path V2 implementation over the same eligible people within each target.
| Measurement | September 20 | October 4 |
|---|---|---|
| Targets compared | 632 | 632 |
| Eligible target population | 4,910 | 30,694 |
| Same first-ranked person | 8.07% | 2.06% |
| Average overlap among first three | 15.88% | 4.75% |
| Average overlap among first five | 20.73% | 6.49% |
| Identical top-five membership | 6.80% | 0.63% |
| Targets with changed top-five membership | 589 / 632 | 628 / 632 |
| Independently judged usefulness | Not measured | Not measured |
By October 4, 628 of 632 targets had at least one different person in the top five, and the same first-ranked person appeared only 2.06% of the time. The models are increasingly producing different shortlists.
These are disagreement measurements, not recommendation-quality measurements. They show that the network model and pinned V2 rank the existing eligible pool differently. They do not establish which shortlist is better.
Coverage, denominators and limitations
The October comparison covered 943,285 target-candidate occurrences with no missing target or candidate embeddings. GraphSAGE moved 2,955 candidate occurrences into top-five lists from the existing eligible pools; these are not necessarily unique people and do not represent outside-pool discovery.
Some compared pairs were observed training edges, so this is a description of ranking differences, not an independent accuracy holdout. The comparison used a pinned repository implementation, not live customer API responses. It did not evaluate a particular seller's relationships, introduction willingness, independently judged relevance, meetings, revenue, or sales-cycle improvement.
Why a more connected database matters
For example, a profile search might suggest another person with a similar title. Relationship information might instead suggest a colleague with relevant shared work history. That is an illustration of the question being tested, not a verified customer success.
GraphSAGE is a published method introduced by William L. Hamilton, Rex Ying, and Jure Leskovec. It learns from information about a person and the people connected to them.[1][2] ExecGraph applies the method to Energy relationship data; it did not invent GraphSAGE.
The platform opportunity is to turn accumulating relationship evidence into better decisions. More recorded connections provide more context to evaluate. Establishing which connections are useful is the product and research challenge, rather than assuming that a larger graph automatically gives better advice.
Three questions the testing must answer
A higher test score or a different contact list does not, by itself, mean a better sales recommendation. The fixed-reference relationship tests address the first question. The September 20 and October 4 Warm Path comparisons address the second. Independent usefulness judgments remain to be collected.
What AUC and average precision measure
ROC AUC measures how often a labeled recorded relationship scores above a sampled non-link, with tie handling. It is not the percentage of useful contacts or successful introductions.[3]
Average precision summarizes how strongly labeled relationships appear in the ranking as more of them are recovered. It is not precision among the first five contacts. Its meaning depends on the test's positive/negative mixture.[4]
Historical evaluation details
The August test used 2,077 positives and 2,077 sampled negatives; September 4 used 5,291 of each; October 4 used 20,383 of each under the fixed reference. Validation and test positive edges were excluded from message passing. Known relationships were excluded from negative sampling, and the degree feature used training edges only. The exclusion policy used the recorded snapshot, so this was not a strictly temporal forecasting test. An absent relationship record is not proof that two people do not know one another.
Validation average precision selected the checkpoint. Same-seed repeats were reported, but an independent multi-seed study was not performed. Identical cross-date runtime environments were not established from the recovered records. The fixed-reference dates have different people and test pairs, so the longitudinal comparison does not isolate the effect of adding relationships.
The semantic comparator uses profile similarity. GraphSAGE also uses structured context and supervised learning. Neither that comparison nor its AUC should be relabeled as a test of the complete Warm Path V2 product.
What the measurements leave open
Did more relationship data cause the higher scores?
The recorded network became more connected. How much that contributed to model scores has not been isolated from other changes. People, links and test pairs changed across dates. September 20 also changed the test setup, while October 4 restored a fixed-reference observation and separately repeated the September 20 diagnostic.
Would the result hold with harder examples?
Unrecorded relationships can be hidden true links, and plausible same-context alternatives can be harder than random pairs. Published evaluation research highlights the importance of strong baselines, consistent splits and realistic negatives.[5]
Does a repeatable score establish confidence?
No. Same-seed repeatability is not independent-seed robustness. Shared people and contexts also make graph observations dependent. The reported point estimates do not include a significance claim.
Does it help a salesperson?
That question is still open. No independently judged shortlist quality, successful introduction, meeting, revenue or sales-cycle improvement was measured. All recorded experiments kept GraphSAGE at zero production traffic and left customer rankings unchanged.
What is available publicly?
Only reviewed aggregate observations. Person-level records, graph connections, feature specifications, software settings and checkpoints remain private. The September 20 and October 4 executions retain fuller private evidence packages. Public aggregates do not expose person-level records, implementation settings, or model artifacts. The public downloads are not replication datasets.
Building toward better-informed sales decisions
The direction is a platform that becomes more useful as it accumulates relevant evidence, not simply a database with more names. The growing network provides the foundation. The next tests must establish whether the recommendations deliver on that opportunity.
Keep the fixed-reference series unchanged. Future checkpoints should repeat the same reference recipe so the trend remains interpretable. Candidate methods can be tested alongside it, not substituted for it.
Now judge whether the changed contact lists help. Reviewers should assess recommendations without seeing which model produced them. They need to check relevance to the opportunity and evidence supporting the proposed connection.
The goal is validated added value: a useful, evidence-supported option that the existing shortlist missed. New tests of discovery outside the existing pool would be a separate experiment.
Sources and research lineage
ExecGraph Energy conducted the applied experiments. The references below explain the published method, metrics and evaluation context; their authors did not evaluate ExecGraph's results. Citation does not imply endorsement or affiliation.
- [1] Inductive Representation Learning on Large GraphsWilliam L. Hamilton, Rex Ying, and Jure Leskovec. NeurIPS 2017; arXiv:1706.02216.Original GraphSAGE method. Not a source for ExecGraph's experiment results.
- [2] GraphSAGE projectStanford Network Analysis Project. Original project documentation.Node features, neighborhood aggregation, and inductive representations.
- [3] Classification: ROC and AUCGoogle for Developers. Machine Learning Crash Course.ROC AUC definition and interpretation.
- [4] average_precision_scorescikit-learn. Official metric documentation.Recall-weighted, non-interpolated AP definition; not a claim about the experiment's software implementation.
- [5] Evaluating Graph Neural Networks for Link Prediction: Current Pitfalls and New BenchmarkingJuanhui Li, Harry Shomer, Haitao Mao, Shenglai Zeng, Yao Ma, Neil Shah, Jiliang Tang, and Dawei Yin. NeurIPS 2023, Datasets and Benchmarks.Motivation for shared evaluation settings, stronger baselines, and more realistic negative examples. Not an independent validation of ExecGraph.
Citation for this research note
ExecGraph Energy (2026). Beyond more contacts: a compounding Energy relationship network. EXG-RN-001, version 1.3. Published September 5, 2026; updated October 4, 2026. Offline research note, not peer reviewed.
@techreport{execgraph2026energygraph,
author = {{ExecGraph Energy}},
title = {Beyond more contacts: a compounding Energy relationship network: Network growth and model results from August 21 through October 4},
institution = {ExecGraph Energy},
year = {2026},
number = {EXG-RN-001},
version = {1.3},
url = {https://execgraphenergy.com/research/energy-graphsage-replay-2026-09},
note = {Originally published September 5, 2026; updated October 4, 2026. Fixed-reference observations on August 21, September 4 and October 4; separately classified September 20 and October 4 diagnostics. Not peer reviewed.}
}
Revision history
Version 1.0 · September 5, 2026. First publication of the August 21 baseline and September 4 replay.
Version 1.1 · September 20, 2026. Added the separate September 20 diagnostic and a clearer business explanation. Original reference values and metric definitions remained unchanged.
Version 1.2 · September 20, 2026. Integrated September 20 into the visible comparison while retaining its changed-test classification.
Version 1.3 · October 4, 2026. Added October 4 as the third fixed-reference observation, updated network-growth trends, and added the fresh October supplementary reranking diagnostic. September 20 remains separately classified.
Historical values are preserved rather than overwritten. Report a correction.