How to Measure Entity Resolution Rate for a Brand
Measure entity resolution rate with a reproducible brand-scoring worksheet. Keep identity accuracy, non-mentions and missing answers in separate denominators.
To measure entity resolution rate for a brand in AI search, divide resolved in-scope brand mentions by all in-scope brand mentions under a matching rule fixed before scoring. Preserve the answers and reference identity facts, then report results by engine, product, query panel and date. This is observed identity accuracy among mentions—not how often the brand appears, a hidden provider confidence score or a citation guarantee.
The canonical Entity Resolution Rate definition holds the formula; the AuthorityTech glossary gives the short explanation. This guide owns the measurement job: a reproducible scoring worksheet, explicit exclusions and comparison rules. For the intervention decision after scoring, use which entity-recognition solutions to try first. Machine Relations Research and AuthorityTech are both Bramwell companies. The formula lives on Machine Relations Research.
The metric matters because AI-mediated brand discovery compresses research into generated answers. If a system cannot distinguish the brand being asked about, the answer may omit it, miscategorize it, or cite a source that frames it incorrectly. Entity resolution rate gives teams a practical way to test that output without pretending the engines publish one shared threshold or one deterministic source hierarchy.
Harvard Business Review's March 2026 analysis of brand readiness for agentic AI documented the practical risk: brand data in AI systems can be incomplete or incorrect before teams notice the issue. That supports direct brand-output auditing; it does not prove that any one source class or entity signal determines a recommendation.
Entity resolution rate was introduced by Jaxon Parrott, founder of AuthorityTech, as part of the Machine Relations measurement framework. Machine Relations names the operating work of making brands legible, retrievable, and citable to machine readers. Entity resolution rate is the measurement doorway: it asks whether that work is visible in real AI answers.
Set the identity record before scoring
Write a dated identity record with sources for every required field: canonical name and domain, category, founder, product or relationship as appropriate to the question. Freeze the query set, accepted aliases, engine or product, location, account state, inclusion rule and matching rule before collection. Brand-named and unbranded questions are different cohorts; keep their results separate. Do not require an answer to state a fact unrelated to the question.
The term adapts entity resolution from database and knowledge-graph work, where systems decide whether different records refer to the same real-world entity. In AI search, the practical question is narrower: when a person asks a category, comparison, use-case, or brand-adjacent question, does the answer system connect the available evidence to the right brand and characterize that brand accurately?
Key takeaways
- Entity resolution rate is observable: score the answer, cited sources, and brand characterization; do not infer a provider's hidden confidence threshold.
- The denominator is in-scope brand mentions: this worksheet counts at most one target-brand assessment per answer. Disclose that unit, engine, query cluster, date and scoring rubric.
- Third-party coverage is useful source evidence, not an automatic outcome: citation inventories show where sources appeared, not why a provider selected them.
- Earned, owned, structured, and knowledge-base signals play different roles: each should be tested against retrieval, citation, recommendation language, referral, and commercial outcome separately.
- The Machine Relations framework provides the operating model: improve source evidence, entity clarity, and machine-readable owned surfaces, then remeasure the same panel.
How ChatGPT, Perplexity, Gemini, and Claude resolve brand entities
AI answer systems can draw on multiple evidence surfaces when representing a brand: indexed pages, citations in retrieved documents, knowledge-base records, structured data, current web search, and model priors. The exact mix is provider-, feature-, query-, and date-specific. Because the systems do not expose one complete entity-resolution formula, a brand team should measure the output and inspect the sources that appear with it.
The academic framing is well documented. Research from Dong Liu and Sreyashi Nag at arXiv on query brand entity linking in e-commerce search describes the need to detect a brand mention and disambiguate which entity is meant when multiple matches exist. That research supports the entity-linking lens; it does not say that a public LLM uses one published threshold for every brand answer.
For practical brand strategy, this translates into one measurement question: does the documented identity online match the actual identity, and do AI answers reproduce it correctly? This is the core of machine resolution as a brand discovery framework.
Why fixed entity-resolution thresholds are unsafe
There is no public, universal 80/60 rule for whether AI systems cite or omit a brand. A high observed rate across a stable panel is valuable evidence that the brand is being represented consistently in that measurement environment. A low observed rate is a diagnostic signal, not proof that one hidden confidence floor rejected the brand.
Quality-threshold research can still be useful when it stays inside its design. The GEO-16 framework analysis by Kumar et al. studied citations across Brave, Google AI Overviews, and Perplexity using a defined prompt and page-quality benchmark. Treat its quality index and reported citation-rate differences as findings within that study, not as a provider-wide rule that determines brand inclusion.
| Observed resolution band | What the measurement can indicate | What it does not prove |
|---|---|---|
| High for a query panel | The brand is repeatedly identified and represented accurately in the tested engines and prompts | That every engine uses the same threshold or that future answers will keep citing the brand |
| Mixed across engines or prompts | Resolution may depend on query wording, source availability, location, freshness, or provider behavior | That one source type alone explains the variance |
| Low for a query panel | The brand may have entity clarity, source coverage, retrieval, or positioning gaps worth diagnosing | That the brand is invisible everywhere or that a single intervention will recover inclusion |
Evidence roles: what the cited studies actually show
The strongest entity-resolution work starts by assigning each source the right evidence role. Citation studies, source inventories, and earned-versus-owned comparisons are useful because they show where AI answers and citations were observed. They do not disclose the full selection mechanism behind any provider, and they do not guarantee that a source placement will produce a brand recommendation.
Muck Rack's Generative Pulse analysis classifies links inside an observed answer sample and reports earned, owned, and paid-source composition within that sample. That is a source-composition observation inside Muck Rack's taxonomy and dataset. It does not establish that earned coverage causes citation, that named outlets will be cited for a specific brand query, or that editorial reputation is a hidden selection mechanism.
Ahrefs' analysis of ChatGPT's most-cited pages describes the domain-rating distribution and related correlations among pages that already appeared in its citation inventory. The measured unit is an inventory of cited pages. It does not establish that domain rating causes citation, that low-rating pages cannot be cited, or that domain authority alone determines citation value.
The Fullintel and University of Connecticut AI media citation study classified sources in sampled AI responses and reported substantial journalistic, earned-media, and unpaid-source shares in that sample. These are source-composition observations within sampled responses. They do not establish a universal engine-selection mechanism, prove that earned coverage causes a brand citation or recommendation, or identify journalism as the primary mechanism behind AI recommendation.
Stacker and Scrunch's earned-vs-owned AI citation-rate pilot, reported on Machine Relations, reports a bounded observed rate comparison between eight stories republished across third-party outlets and their brand-only versions (about 34% of answers against about 8%). The measured unit is an observed rate comparison inside that dataset. It does not establish that earned media caused the citations, guarantee a fixed multiple for any brand, or prove the primary mechanism behind every AI-search citation.
Those studies support a practical hypothesis: third-party evidence often appears in AI citation environments and is worth testing as part of a Machine Relations program. The evidence stops there unless a brand runs a controlled measurement that ties a dated source intervention to changed retrieval, citation, recommendation language, referral, or commercial outcome.
Moz's 2026 analysis of Google AI Mode citations found substantial non-overlap between AI Mode citation URLs and organic search results in its sample. That finding supports measuring AI citations separately from rankings; it does not prove that SEO is irrelevant or that one citation population governs every AI answer surface.
Entity signal workstreams to test
Entity resolution work should be organized as testable workstreams, not as a claim that AI engines cross-reference every signal or weight every source the same way. The useful question is whether improving a workstream changes the measured output for a fixed query panel.
| Workstream | What it improves | How to measure it |
|---|---|---|
| Entity clarity on owned properties | Consistent names, categories, descriptions, founders, locations, and canonical URLs | Crawl the owned pages and compare answer descriptions against the intended entity profile |
| Third-party evidence | Independent descriptions of the brand from category-relevant sources | Record which sources are retrievable, cited, and used to support claims in AI answers |
| Knowledge-base anchors | Structured identifiers where they are appropriate and accurate | Test whether the brand is distinguished from similarly named entities in brand and category prompts |
| Citation architecture | Answer-first passages, supported claims, tables, FAQ answers, and clean machine-readable representations | Verify crawl access, Markdown/API surfaces, retrieval, and citation separately |
| Query relevance | Alignment between the language buyers use and the evidence available about the brand | Compare category, use-case, comparison, and problem-aware prompts over time |
For deeper implementation context, see how to improve entity resolution rate in AI search and how AI search engines evaluate source credibility. Use those guides as operating hypotheses, then test against real outputs.
How to measure entity resolution rate across AI engines
Measuring entity resolution rate requires a fixed query set, repeatable scoring, and separate tracking for each answer surface. Do not mix raw brand mentions, accurate entity resolution, recommendation language, and revenue impact into one score.
A scoring worksheet that keeps the denominators separate
For this worksheet, count at most one target-brand assessment per successfully collected answer. Record repeated mentions and contradictory facts, but do not multiply the denominator for a name repeated in one answer. A protocol that counts individual occurrences has a different unit and must be reported separately. Define how accepted aliases and common-word names enter scope before scoring.
| Preserved answer state | Entity-resolution denominator | Recorded outcome |
|---|---|---|
| Request failed, engine unavailable or no usable answer returned | Excluded: no answer to assess | Missing collection cell; keep attempted and successful collection counts |
| Usable answer does not name the target brand | Excluded: no in-scope mention | Non-mention in the separate mention-rate denominator |
| Eligible target-brand mention has a wrong required identity fact | Included | Unresolved; preserve each mismatch and its source |
| Eligible mention is ambiguous, contradictory or missing a required fact | Included | Unresolved under this strict matching rule; retain the uncertain fields |
| Eligible mention matches every required identity fact | Included | Resolved; score citation, sentiment and recommendation separately |
Copy these columns into the assessment: exact prompt; engine and product; timestamp; collection status; preserved answer; cited URLs; target-brand presence; accepted alias; required field; reference value and source; observed value; match, mismatch, absent or uncertain; final disposition; reviewer. Apply exclusions only under the rule declared before scoring, and publish the excluded counts and reasons. Uncertain eligible mentions stay unresolved; do not remove them just to raise the rate.
Calculate collection coverage over attempted cells, mention rate over usable in-scope answers, and entity resolution rate over eligible target-brand assessments. If there are no eligible mentions, entity resolution rate is unmeasured, not zero. If every named instance matches but most answers omit the brand, report strong conditional accuracy and weak mention coverage—not broad recognition. These are measurement conventions, not claims about any engine’s private architecture.
| Step | Action | Detail |
|---|---|---|
| 1. Define the query panel | Select category, comparison, use-case, and problem-aware prompts | Document wording, geography, engine, account state, and date before running the test |
| 2. Capture the answer | Run each prompt across the engines that matter to the audience | Save answer text, cited URLs, timestamps, and whether the brand appears |
| 3. Score entity accuracy | Evaluate presence, category accuracy, attribute accuracy, and positioning alignment | Separate correct resolution from vague mentions, outdated descriptions, and competitor confusion |
| 4. Attribute hypotheses carefully | Map failures to possible evidence gaps | Inspect crawlability, source coverage, entity consistency, freshness, and prompt relevance before naming a cause |
| 5. Remeasure after one intervention | Change one workstream at a time when possible | Compare the same panel after the new evidence is live and available to crawlers |
For brands that want to measure brand mentions in AI search systematically, identity matching adds a separate accuracy assessment. The AuthorityTech visibility audit is an AI-estimated starting view, not evidence that this preserved-answer worksheet has been executed or that a measured entity resolution rate exists.
Entity resolution rate vs. share of citation
| Metric | What it measures | What it cannot show alone |
|---|---|---|
| Entity resolution rate | The share of in-scope brand mentions that match the declared identity record | Why the system selected or omitted the brand |
| Share of citation | How often the brand or its sources are cited across monitored answers | Whether the citation is favorable, accurate, recommended, or commercially valuable |
| AI share of voice | Relative presence against competitors in a prompt set | Whether the brand is represented correctly as an entity |
Share of voice is not obsolete, but it measures a different dimension. A brand can be mentioned often and still be miscategorized. A brand can be accurately resolved in one high-intent cluster and absent elsewhere. Entity resolution rate keeps the scoring tied to the quality of identification, not merely the count of mentions.
Use the scored failure to choose the next test
Once the worksheet distinguishes wrong identity, stale attributes, non-mentions and selection outcomes, use the entity-recognition solutions guide to choose a candidate intervention. Keep the same engine, product, query cohort, identity record and matching rule for a later comparison. Disclose model, retrieval or collection changes instead of pooling them into an apparent improvement.
A better observed rate does not establish the cause. Preserve baseline and follow-up answers, missing collection cells and denominator changes; a before-and-after description is not a controlled result.
The cost of poor entity resolution in B2B buying
Low entity resolution can matter commercially because more buyer research is happening before a vendor conversation. Forrester's 2024 State of Business Buying report describes how much B2B research happens before direct contact with a vendor, and Bain's 2025 consumer research reported broad reliance on AI summaries in search behavior. Those sources support monitoring AI-mediated discovery; they do not prove a specific revenue loss from one omitted answer.
For brands already seeing negative brand sentiment in AI search, entity resolution rate helps separate three problems: the system does not identify the brand, identifies it with stale or incorrect attributes, or identifies it correctly but frames it unfavorably. Each problem requires a different source and measurement strategy.
Oxford and Stupid Human's ChoiceEval research on auditing brand preferences in LLMs found biases and variation in brand recommendations within its tested setup. That finding supports auditing recommendations directly. It does not establish that every recommendation follows one entity-signal graph or that resolution alone determines buyer outcomes.
What a good entity-resolution report includes
Stacker's February 2026 analysis captured the industry shift toward machine-mediated discovery and used the Machine Relations framing in an earned-media context. That publication supports the category narrative; it does not establish how any provider selects one brand or source for a given prompt.
A useful report should be specific enough that another team can rerun it. Include the exact prompt list, engine versions or surfaces where available, dates, location, account state, answer text, cited URLs, scoring rubric, entity facts checked, and unresolved ambiguity. When the report names an intervention, include when the evidence went live and when the same prompt panel was rerun.
The output should separate five stages:
- Retrieval: did the answer system appear to access a relevant source?
- Citation: did the answer cite the brand, source, or supporting page?
- Resolution: did it identify the correct entity and category?
- Recommendation language: did it present the brand as a candidate, leader, caution, or neutral mention?
- Outcome: did referral, pipeline, or revenue move in a way that can be measured separately?
Separating those stages keeps the analysis useful. A cited source is not the same as a recommendation. A recommendation is not the same as a qualified buyer. A better entity-resolution rate is a leading measurement, not a business-outcome guarantee.
FAQ
What is entity resolution rate?
Entity resolution rate is resolved in-scope brand mentions divided by all in-scope brand mentions under a declared identity-matching rule. Report the counting unit, query panel, engine, dates and exclusions. A usable answer without a target-brand mention belongs in the separate mention-rate calculation; a failed collection is missing evidence.
Who introduced entity resolution rate as a brand metric?
Jaxon Parrott, founder of AuthorityTech, introduced entity resolution rate as part of the Machine Relations measurement framework. Machine Relations describes the discipline of earning AI citations and recommendations for a brand by making that brand legible, retrievable, and credible inside AI-driven discovery.
How does entity resolution rate relate to share of citation?
Share of citation measures how often a brand or source is cited across monitored answers. Entity resolution rate measures whether the brand is identified and described accurately. A citation can exist without accurate resolution, and accurate resolution can occur without a commercial outcome.
How do AI systems decide whether to cite a brand?
Providers do not publish one complete rule for brand citation. Citation inventories from Ahrefs, Muck Rack, and Fullintel-UConn show source composition and cited-page patterns in their samples; they do not reveal a universal selection mechanism. Measure crawl access, retrieval, citation, recommendation language, and entity accuracy separately.
Can a brand have high Google rankings and low entity resolution rate?
Yes. Traditional ranking, AI citation, and entity resolution are different measurements. A page can rank well while an AI answer misidentifies the brand, and a brand can be accurately represented in an answer without that answer proving a ranking or revenue outcome.
What is the fastest way to improve entity resolution rate?
Start with a dated measurement, then fix the highest-confidence evidence gap. For many brands that means clearer owned entity facts plus accurate third-party descriptions in relevant sources. Third-party evidence should be measured as one workstream, not treated as an automatic citation or recommendation outcome.