You spent three years building a brand. Logo, tone of voice, design system, the whole posture. Then ChatGPT recommends three competitors and skips you. Your CMO walks into the next meeting holding a printout of the conversation and asks how this is possible when the brand has 28,000 newsletter subscribers and a 4.7 G2 rating. The room goes quiet because nobody knows.
This is a structural problem, not a marketing problem. The LLM is not reading your G2 rating in real time. It is assembling a sense of your brand’s authority from a layered pile of signals that mostly do not map to what your marketing team measures. Understanding the layers is the only path back into the answer. There are four of them, they compound, and they fail in predictable ways.
What “brand authority” actually means to an LLM
A modern LLM does not store a brand authority score. It reconstructs one on every query. The reconstruction draws from two clocks running at different speeds. The first clock is the pretraining corpus: the text the model saw during its base training, frozen at the model’s knowledge cutoff. The second clock is real-time retrieval: the URLs the answering system pulls into the context window when responding to your query right now.
A brand’s authority on any given answer is the joint signal those two clocks produce. If pretraining is strong but retrieval is weak, the brand reads as historically credible but currently invisible. If retrieval is strong but pretraining is weak, the brand reads as a new entrant whose claims need extra verification. If both are weak, the brand is functionally absent from the answer regardless of how strong its owned channels are.
Four signal layers feed both clocks. Marketers tend to optimize for the wrong two.
Layer 1: Entity Coherence
The first layer is whether the LLM can resolve your brand as a single, distinct entity. This sounds trivial. It is not. Most brands have at least one entity ambiguity problem, and the brands that fix it gain measurable authority lift overnight without writing a single new word of content.
Common ambiguity patterns. A product name that is also a common English word (“Notion,” “Stripe,” “Bolt”). A brand that shares a name with an older, larger company in an adjacent category (“Bench” for bookkeeping versus the woodworking brand “Bench”). A parent-company relationship that the public copy obscures (“Anthropic” versus “Claude,” “OpenAI” versus “ChatGPT” versus “GPT-4”). A founder whose personal brand competes with the company brand for entity resolution (“Stripe” versus “Patrick Collison”).
When the LLM cannot cleanly resolve which entity you mean, it does one of three things. It picks the wrong one, it hedges by listing several entities with the same name, or it skips the answer entirely and recommends an entity with cleaner resolution instead. All three are bad.
The fix is structural. Wikipedia disambiguation page (or your own Wikipedia entry). Wikidata entry with clean parent-company links. Schema.org Organization and Brand markup on your homepage with the official name, the parent company, the founding date, the founders by name, and the official domain. Consistent boilerplate across press releases and bylines. The goal is for every place an LLM might encounter your brand to point at the same canonical entity.
The brands that get this wrong tend to be the ones with the strongest visual branding. Their marketers think a memorable logo equals memorable entity. To an LLM, the logo is invisible. The structured entity graph is everything.
Layer 2: Citation Density Across Independent Surfaces
The second layer is the most important for both clocks and the hardest to game. It measures how often your brand is named by sources the LLM treats as independent of you. Citation density across independent surfaces is the closest LLM-era equivalent to the old PageRank insight: links from sites that don’t have a reason to link to you are worth more than links from sites that do.
The surfaces an LLM treats as independent are a smaller, more specific list than most marketers realize. Major-outlet journalism with named bylines (Reuters, AP, Bloomberg, The Information, NYT, WSJ, FT, The Atlantic, Wired, etc.). Academic papers with author affiliations to universities. Government publications. Industry-trade journalism (Modern Retail, Restaurant Business, Inside Higher Ed, etc.). Independent review aggregators (G2 verified-buyer reviews, Capterra, Trustpilot). Wikipedia. Substack and Ghost newsletters with named authors and consistent publication history.
The surfaces that look like citations but carry near-zero weight in modern retrieval systems include press-release-wire pickups that are mirrors of the same wire copy, vendor-syndicated case studies, Medium republishes of your own blog posts, low-quality directory listings, and content-farm “best of” articles. The LLM has learned to discount these as derivative.
The single highest-impact move under Layer 2 is earning one citation in long-form journalism every quarter. Not a press release pickup. A feature where a reporter names your brand inside an analytical context they are constructing themselves. This is harder than running a paid campaign and worth more than a year of paid attention.
The second-highest-impact move is earning citations in academic or industry-research literature. A study published in Harvard Business Review that names your category and your brand outperforms a hundred Forbes Council pieces because the LLM weights academic sources higher than self-published thought leadership.
Layer 3: Temporal Consistency
The third layer is the one most brands ignore until it bites them. Temporal consistency measures whether the signals about your brand have been stable over a span of time the LLM can verify. New brands with strong recent signals but no historical record read as suspicious. Older brands with stale recent signals read as in decline. The LLM is doing the same pattern-matching a human due-diligence analyst would do, just faster.
Three things compound to produce temporal consistency. The age of your domain (visible in WHOIS and indirectly in archive.org’s earliest crawl). The age of your earliest reliable third-party citation (visible in Wikipedia article creation dates, oldest mentions in major journalism, original incorporation records). The cadence of new third-party citations (whether mentions are spiking, flat, or declining over the last 24 months).
A brand whose domain is 18 months old, whose earliest citation is from a YC announcement six months ago, but whose name is currently appearing in 12 industry articles per month reads as “credible new entrant.” A brand whose domain is 12 years old, whose earliest citation is from a 2014 Wired feature, but whose name has appeared in zero major-outlet pieces in the last 18 months reads as “fading legacy.” The LLM treats the second brand worse than the first for forward-looking recommendations, even though the second brand may be objectively more established.
The actionable read on temporal consistency: do not let your brand’s citation pattern flatline for more than two consecutive quarters. The decay curve is steeper than marketers think. A brand that was prominent in 2022 and quiet in 2026 starts to disappear from recommendation answers because the LLM is inferring that current users would not choose it.
A second actionable read: when you rebrand or rename, expect a 6 to 12 month authority dip. The LLM has to relearn the entity resolution under the new name. Rebrands during a period of weak recent citations are dangerous because the old name’s temporal record does not transfer cleanly to the new name. Plan rebrands during periods of strong citation flow.
Layer 4: Structural Disambiguation
The fourth layer is the one your engineering team can move fastest. Structural disambiguation is the set of machine-readable signals on your site, in your meta data, and in third-party knowledge graphs that tell an LLM exactly what your brand is, what category it competes in, and what attributes describe it.
The schema markup that matters: Organization with sameAs links to your Wikipedia, Wikidata, LinkedIn, Crunchbase, and X. Product schema for each product with category, audience, and keywords properties. Person schema for your named founders and executives, with cross-links to their LinkedIn and to any media bylines they have. Article schema on every editorial page with named authors and Person cross-references. FAQPage schema on pages where you answer concrete questions a user might ask.
What most teams get wrong: they install schema once during a website rebuild, then never update it as the company grows. Founders change, products get renamed, the category positioning shifts. The schema is still pointing at the org chart from three years ago, and the LLM is inheriting that stale signal.
A second common error: stuffing keywords into schema fields hoping the LLM rewards it. Modern retrieval systems detect and discount keyword stuffing in structured data the same way Google did after Hummingbird. Use the fields for what they describe. The LLM is not fooled, and your authority signal degrades when the structured data and the prose data conflict.
The non-obvious move under Layer 4: get your brand into the major business knowledge graphs that pretrain models. Crunchbase, PitchBook, ZoomInfo, Apollo, Bloomberg’s data products, Refinitiv’s company database. These get crawled and licensed into training datasets. A clean Crunchbase profile with correct funding history, founding dates, and category tags is worth more than most marketers think because it gets folded into the model’s training corpus when models update.
How the four layers compound
The layers are not additive. They are multiplicative. A brand that is strong on three layers but invisible on one reads as suspicious, not strong. Specifically, weak Layer 1 (entity coherence) crushes the rest. An LLM that cannot reliably identify which entity you are will not credit the citations, schema, or temporal record you have built up under that identity.
The compounding pattern that produces durable LLM authority looks like this. Layer 1 stable: clean entity resolution, unambiguous Wikipedia and Wikidata presence, consistent boilerplate. Layer 4 supports it: schema markup that mirrors and extends the entity graph. Layer 2 builds on top: regular cadence of independent third-party citations that all name the canonical entity. Layer 3 emerges over time: a multi-year record of citations that all resolve to the same entity, with no rebranding chaos.
When all four are stable, the brand crosses an inflection point where it starts to appear in answers it did not target. The LLM has internalized the entity strongly enough to generalize, and the brand becomes a default recommendation in adjacent query classes. This is the state every brand wants and almost no brand reaches without deliberate work on all four layers at once.
What to fix this quarter
If your authority is weak, fix in this order. First, audit Layer 1 entity coherence. Resolve every ambiguity. Update Wikipedia or Wikidata if it does not exist. Standardize boilerplate. Second, audit Layer 4 schema. Make sure the entity in your structured data matches the entity in your Layer 1 work. Third, plan one Layer 2 citation per quarter, prioritizing long-form journalism with named bylines at outlets whose archives are in the pretraining corpus. Fourth, set up monitoring so Layer 3 does not silently decay. Mention tracking via a service like Brand24, Meltwater, or a custom Perplexity-citation monitor will catch a citation decline before it shows up in your dashboards.
Patagonia is the cleanest commercial example of all four layers working at once. Its entity resolution is unambiguous. Its citation density is high across journalism, academic environmental research, and trade press. Its temporal record stretches back 50 years with consistent flow. Its schema markup and Wikipedia presence reinforce the entity. The result is a brand that LLMs recommend as a default in dozens of query classes (sustainable apparel, B-corp examples, environmental activism in business) without any of those being the brand’s stated marketing focus. That cross-query authority is the prize. The four-layer stack is the path.