No AI search provider has publicly disclosed a formula for this, so any specific answer describing exact weighting would be fabricated. What can be honestly said is that the plausible mechanism, though undisclosed by any provider, draws on well-established NLP concepts: entity recognition and disambiguation, co-occurrence statistics (how often a brand/entity name appears near other terms, entities, or topics across a large corpus), and sentiment or stance classification applied to text mentioning the entity. Whether or how these are combined into anything resembling a discrete “authority score” inside any current AI search system (Google’s AI Overviews, Bing/Copilot, Perplexity, or others) is not something any of these providers has confirmed, and the idea should be treated as a plausible conceptual mechanism, not a documented one.
The conceptual building blocks, and why they’re plausible in principle
Entity recognition (identifying that a specific string of text refers to a specific real-world entity, such as a brand, person, or organization, and disambiguating it from unrelated uses of similar terms) is a mature, long-established NLP capability, not a speculative one. Search engines have used entity recognition and entity-linking (associating mentions with structured knowledge, like a knowledge graph node) for well over a decade; Google’s Knowledge Graph and its associated entity-understanding work in Search predate the current wave of generative AI search products by many years. It’s well established, and Google itself has discussed this in general terms historically, that a search system can recognize when a brand or entity is mentioned in text even without a hyperlink pointing to that brand’s own website, because entity recognition operates on the text itself, not on link structure.
Co-occurrence analysis, looking at what other words, entities, or topics frequently appear near a given entity’s mentions across a large text corpus, is likewise a standard, decades-old technique in computational linguistics and information retrieval, used for tasks like building topic associations, disambiguating word senses, and inferring relationships between entities. It is plausible, in principle, that an AI search system could use co-occurrence patterns as one input for understanding what an entity is associated with or “known for” in aggregate web text, independent of any specific link pointing at that entity’s site. Sentiment or stance detection (classifying whether text discussing an entity is positive, negative, or neutral) is also a mature, widely deployed NLP task, and it’s conceptually plausible that a system attempting to model an entity’s reputation or trustworthiness might incorporate some signal derived from the aggregate sentiment of text mentioning it.
This general idea, that a brand mentioned frequently in relevant, credible contexts (even without inbound links) could contribute positively to how a search system perceives that brand’s relevance or prominence for a topic, has circulated in SEO industry discussion for a long time, sometimes under the informal label of unlinked brand mentions as a possible ranking or entity-strength signal. That discussion has some grounding in older, general web-search-era commentary (including informal statements from Google personnel over the years acknowledging that Google’s systems can recognize brand mentions in text and that such recognition is not strictly dependent on hyperlinks), but it has never been accompanied by a confirmed, disclosed formula, even for classic Google Search, let alone for any current AI-search product specifically.
What’s genuinely unconfirmed, and should not be overstated
No AI search provider (not Google for AI Overviews, not Microsoft for Copilot, not Perplexity, not any other current system) has published documentation describing a specific “entity authority score,” disclosed how heavily unlinked mentions are weighted relative to linked citations, or confirmed that sentiment classification feeds into any ranking or trust calculation at all. It is entirely possible that some AI search systems incorporate something in this general direction as part of broader entity-understanding or source-evaluation processes, since the underlying NLP techniques are mature enough to make it technically feasible, but “technically feasible and conceptually plausible” is different from “confirmed to be happening in a specific, describable way.” There is no public number, percentage, or named metric that legitimately describes how much unlinked mentions, co-occurrence, or sentiment contribute to any AI search system’s evaluation of a brand, and treating any such figure as real would be inventing something no provider has stated.
It’s also worth noting that “entity authority” itself isn’t a term any major AI search provider uses as a disclosed, formal metric name. It functions in industry discussion as a useful shorthand for a cluster of related ideas (topical association, reputation, prominence, trustworthiness signals) rather than as a documented, singular score any of these systems is known to compute.
A hypothetical illustration
As a hypothetical illustration: suppose a hypothetical outdoor gear company called Northgate Outfitters is frequently discussed favorably in gear-review forums and outdoor-recreation publications, but relatively few of those mentions actually hyperlink back to Northgate’s own site, most are just the brand name mentioned in the context of trail gear recommendations. Under a purely link-based model of authority, this pattern of unlinked mentions would contribute nothing measurable.
Hypothetically, if an AI search system’s entity-recognition layer is scanning a large corpus of outdoor-recreation text, it could plausibly still register that “Northgate Outfitters” co-occurs frequently and consistently with terms like “durable,” “recommended for backcountry trips,” and specific product categories, even without ever following a link to Northgate’s domain. If Northgate’s name also carries broadly positive sentiment across that corpus, again, hypothetically, this could be one input among many contributing to how confidently a system treats Northgate as a credible, recognizable entity in the outdoor gear space when constructing an answer to a query like “durable backpacking gear brands.” The important caveat, consistent with the mechanism described above, is that this is a plausible illustration of how such a system could work in principle, not a confirmed description of how any specific AI search product actually weights Northgate’s unlinked mentions today.
What this means practically, held to the appropriate confidence level
Given the genuine uncertainty here, the reasonable practical stance is to treat brand mention volume, context quality, and sentiment as plausibly useful inputs to how AI systems build an aggregate understanding of an entity, worth paying attention to as a matter of prudent brand-building, without treating any specific optimization tactic as guaranteed to move a confirmed metric:
Encourage and track unlinked brand mentions in credible, topically relevant sources (industry publications, forums, review sites, news coverage) as part of general digital PR and reputation work, on the reasoning that even if the exact mechanism by which AI systems use this is undisclosed, being discussed accurately and positively in the sources these systems draw from is very unlikely to be harmful and plausibly helps with however entity understanding actually works.
Pay attention to sentiment in earned coverage and mentions, correcting misinformation or negative narratives where reasonably possible, again on the general principle that if sentiment plays any role in how these systems characterize an entity, the far bigger risk is decisively negative sentiment than any missed optimization opportunity.
Avoid citing a specific claimed percentage, weighting, or named “entity authority algorithm” in client-facing or public claims, since doing so would be presenting an invented specific as if it were disclosed fact; the honest position to communicate is that this is a plausible, undisclosed mechanism grounded in established NLP concepts, not a documented AI-search ranking formula.
