Short answer: There is no universal "entity salience score" published by Google for publishers, and no official percentage that says how well your company has understood an AI engine. But you can measure observable properties: name and attribute consistency, brand query coverage, factual accuracy, sources cited, and representation stability over time. For B2B SaaS, they are more useful than a composite index that hides errors.

First define what you call salience

In this audit, `entity salience' is not a claimed ranking factor. It is an operational term for how clearly and consistently an entity appears in the set of surfaces we are measuring.

If a commercial tool provides its own "salience score", treat it as that tool's metric until the methodology is explained and validated. Don't make it an official search engine metric.

The minimum baseline

Construct a fixed set of questions in four groups:

  1. identity: "what is Brand X?";
  2. category: "what products does Brand X offer?";
  3. relationships: "what does Brand X integrate with?";
  4. evaluation: "for what type of company is Brand X suitable?".

Add non-brand buyer journey questions, but keep them separate. Otherwise you won't know if you're measuring brand identity or category competitiveness.

Metrics that can be defined

Fact accuracy rate

Choose a set of verifiable claims: product name, category, important integrations, deployment model, regions or other public attributes. For each observation mark correct',incomplete', wrong' orcannot be verified'.

The denominator must be the number of eligible claims, not the total number of sentences generated.

Owned source citation rate

When the AI product displays sources, it measures the proportion of observations in which a first-party URL appears. Don't confuse this with brand mention rate.

Third-party corroboration rate

Measure how many responses use external sources that correctly confirm identity or attributes. An old review that misdescribes the product does not become useful just because it is third-party.

Naming consistency

Define the accepted form of the brand and product name. It counts material changes, not non-impact capitalization differences.

Attribute conflict rates

Compare claims about the same property in the website, docs, marketplace and cited sources. Conflict is a governance issue even before it is a visibility issue.

Observation window

Set the duration before changes. For a B2B SaaS site with moderate traffic, don't promise that one week is enough. The frequency of recrawl, the variation of responses and the volume of queries differ.

It's more important to keep the method the same: the same basic prompts, the same language, the same engines, and the same coding rules.

How do you test an intervention

Let's say you've found three problems: inconsistent organization markup, an outdated About page, and a marketplace listing with the old product name. Intervention can correct all three, but must be registered as a package.

After the change, track fact accuracy and naming consistency. If they improve, you can say that the observed representation has improved in your sample. You cannot automatically say that the scheme was the sole cause.

False attribution risks

  • an AI engine changes its search model or system;
  • new articles about the brand appear;
  • the company launches a product and receives coverage;
  • a marketplace updates data independently;
  • the query mix changes;
  • the prompts did not remain comparable;
  • human raters use different rubrics.

These do not invalidate monitoring. It just forces you to separate observation from explanation.

A better dashboard than a single score

It uses five separate series: accuracy, owned citations, external corroboration, naming consistency, and conflicts. Show the trend and examples. A single score of 82/100 may hide the fact that the brand is mentioned frequently, but a critical integration is misrepresented.

For leadership, add the operational impact: what finding requires correction in docs, in structured data, in third-party profiles or in the product.

What role does structured data play?

Google says Organization markup can help disambiguate the organization and recommends properties that describe identity and online presence. This is a verifiable basis for the audit. There is no justification for inventing a saliency score, nor for promising a certain representation in AI systems.

How do you construct the denominator without manipulating the result

A percentage can look great if the denominator is chosen according to the result. That's why, before measuring, fix the list of eligible queries and the exclusion rule. If an engine does not offer web search for a run or the answer has no visible sources, it decides in advance whether the observation enters the brand mention rate' but not theowned citation rate'.

The same for fact accuracy: don't just note claims where the product is described favorably. Create an inventory of material attributes before the test: category, integrations, constraints, plans, regions, deployment models. If some attributes are not public, they are not included in the score.

Keep raw observations. The dashboard must be able to be recalculated if you change a rule. A series without raw data and without the definition of the denominator is difficult to audit and easy to cosmeticize.

Claim ledger

  • FACT/EVIDENCE: Google documents Organization markup for understanding and disambiguating organizations.
  • FACT/EVIDENCE: sameAs, url, logo and other properties are documented as relevant information when applicable.
  • PRACTITIONER GUIDANCE: measuring on separate series reduces the risk that a single index hides material errors.
  • INFERENCE: a more consistent identity can reduce ambiguity in retrieval, without guarantee of ranking or citation.

Conclusion

It measures entity salience as a matter of observability, not as a magic number. If you can show that the facts are more correct, the name is more stable, the conflicts decrease, and the sources are more consistent, you have a real improvement. The rest must be demonstrated, not assumed.

Sources reviewed

Related insights