Entity resolution in B2B SaaS becomes problematic when the product, company, suite, editions, and integrations are described inconsistently. A model or search engine may end up combining two products with similar names, assigning a feature to the wrong company, or using an old page for a rebranded name.
A good experiment does not start from the assumption that the markup or sameAs "resolves the entity". It starts from concrete cases of ambiguity, defines an intervention, and measures whether observed errors fall on a stable set of questions.
Define the entities that go into the test
Build a simple map: organization, core product, modules, plans, integrations and any historical names. For each element, write down the canonical name, the main URL, and the relationship to the other entities.
If the product has the same name as the company, write this explicitly. If the suite contains multiple applications, don't use the same term for all of them without context.
Identify real errors before intervention
It collects examples where the entity is confused, a function is attributed to another product, an acquisition is treated as an independent brand, or an integration is described as its own product. Keep the prompt, response, sources, and date.
It does not include the mere absence of a mention in the same category. Entity resolution looks at identity and relationships, not just frequency of occurrence.
Formulate the hypothesis
Example: If the organization and product pages use consistent names, clear relationships, and stable references, the proportion of responses that confuse the product with the competing entity will decrease in the tested set.
The hypothesis must specify the error and the population. "We will improve GEO" cannot be tested.
Choose control
Keep a comparable page family or by-product without main intervention. The control helps distinguish specific change from general variation in responses.
In a large portfolio, the control may be another product line with an already established identity. Document why it is comparable and where it is not.
Intervention 1: Clarify visible names and relationships
On treated pages, eliminate name inconsistencies, explain the relationship between the company and the product, and use the same names in navigation, headings, and documentation. Do not overload the text with the brand name; clarity is more important than repetition.
If a product has been rebranded, keep the historical context where users can look for it, but make the current name clear.
Intervention 2: Align Organization and SoftwareApplication where applicable
Google documents Organization structured data, and Schema.org defines SoftwareApplication and properties such as identifier, owner, and sameAs. Use only the properties that reflect real and visible information.
Markup should not introduce relationships that the page does not explain. If the product's owner or identity is ambiguous to humans, JSON-LD is not a substitute for page clarification.
Intervention 3: Eliminate self-contradictory sources
A knowledge base can use an old name, a pricing page can use the new name, and release notes can have abbreviations. These contradictions are a problem of internal provenance.
Create a list of controlled sources and reconcile material terms. Don't rewrite the historical record for no reason; adds context and links to the current designation.
Define the observation window
Do not rate immediately after publication. Keep multiple runs of the same prompts in a stable period. Record releases, campaigns, and media appearances that may change available sources.
If the product simultaneously releases a major feature, the entity resolution test can be contaminated by interest and new content.
Measure error, not just mention
Useful metrics include: number of correct entity responses, product confusions, wrong company-product relationships, old names without context, and inconclusive cases. Report the denominator.
An increase in mentions without reducing confusions does not confirm the hypothesis of entity resolution.
Stop rules
Stop the test if you change the prompt set, if the product is rebranded again, if the control pages receive the same intervention, or if you can no longer verify the sources. Keep the partial result and describe why the experiment is no longer comparable.
Don't retroactively adjust the error definition just to get a favorable result.
What success means
Local success is a reproducible decrease of confusions in the tested set without the appearance of new contradictions. It does not mean that the entity is universally "solved" on the Internet.
Repeat the experiment on a second family of questions before wider rollout.
Check the effect on the buying committee
After measuring entity confusions, also check the questions that different roles on a buying committee would ask. Security may use the product name other than procurement or engineering. If the intervention reduces confusion only in brand prompts but not in those about integration and ownership, success is limited. Report by category and preserve examples where the organization-product relationship remains ambiguous. This layer helps the team decide if the rollout is worth expanding, or if the issue is localized to a single documentation area.
What do you keep after the experiment
Even if the result is inconclusive, it keeps the entity map, confusion list and change log. These artifacts can become the baseline for the next round and help sales engineering or support teams use more consistent names. But don't turn the documentation of the experiment into a marketing obligation to over-repeat the brand. The goal is clarity of relationships between organization, product and modules.
Claim ledger
- FACT/EVIDENCE: Google documents Organization structured data and Schema.org defines SoftwareApplication and identity relationships like sameAs.
- PRACTITIONER GUIDANCE: control/treatment design and confounding metrics are the experimental methodology proposed here.
- INFERENCE: more consistent names and relationships may reduce entity ambiguity, but the effect should be observed.
- NOT PROVEN: that Organization or SoftwareApplication markup directly produces AI citations or better ranking.
