Digital Marketing » Articles » Articles By » The Strategy Sandbox » Three Recruitment Logics: Why Annotation Has to Speak Differently to Knowledge Graphs, LLMs, and Search Engines

Three Recruitment Logics: Why Annotation Has to Speak Differently to Knowledge Graphs, LLMs, and Search Engines

Status: Original concept, first publication. Strategy Sandbox, jasonbarnard.com. Date: May 2026.

The annotation gate has been the focus of the AI Engine Pipeline conversation for over a year, ever since I laid out how AI decides what your content means and why it gets you wrong for Search Engine Land back in April. Annotation is where AI assigns meaning to a chunk of content: what entities are present, what attributes those entities carry, what relationships connect them, what topical context surrounds them. Get annotation right and your content is legible to the algorithms. Get it wrong and the rest of the pipeline collapses against you.

But annotation is only the upstream half of the story. What happens to annotated content next, at the Recruited gate I covered as part of the five competitive gates hidden inside rank and display, is where most brands lose the recommendation, and the reason most brands lose it is that they treat recruitment as one decision when it is in fact three parallel decisions made by three different components of the Algorithmic Trinity for three different reasons.

Knowledge Graphs, Large Language Models, and Search Engines all draw recruitment candidates from the same annotated pool. They do not apply the same recruitment criteria to that pool. The brand that produces one piece of annotated content and hopes all three components find use for it is operating on a category mistake. The brand that understands the three recruitment logics shapes annotation specifically for each, and the brand that understands which engines weight which components most heavily shapes the annotation accordingly.

Two stacked weightings sit underneath every recommendation outcome, and both need to be named.

All three components draw from one shared Web Index, with cleaned subsets feeding the Knowledge Graph and the LLM

Before the three recruitment logics make sense, the architecture underneath them has to be named clearly. Search Engines do not have separate indexes. Google’s index, Bing’s index, and the indexes of every other Search Engine are built by crawlers (the bots) that visit the web, fetch content, and feed it into a single shared Web Index for that engine. The bot is the agent that fills the index. The Search Engine reads from the index it has been given. Every Search Engine ranking decision, every Assistive Engine grounding query, every real-time validation pass against current claims, all draw from this shared, continuously-updated Web Index that the continuous algorithmic annotation layer has been annotating as the bot crawls.

I have been writing about this architecture publicly since 2019, sourced directly from Microsoft Bing engineering. The relationship started at SMX London in 2019 with Nagu Rangan, then leading core ranking work at Bing, where the architectural detail of how candidate sets compete inside the SERP first went on the record. In April 2020 I extended that into a five-engineer working interview series at Bing, published on Search Engine Journal. How Bingbot Works: Discovering, Crawling, Extracting & Indexing was sourced from Fabrice Canel, Principal Program Manager, naming exactly how the bot fills the index. How Bing’s Whole Page Algorithm Works was sourced from Nathan Chalmers, Program Manager Search Relevance Team, naming the algorithm that organises content from the shared index before showing it to the user, and confirming that Bing has an internal algorithm called Darwin that handles the multi-vertical assembly. The same architecture sits underneath Google. In a 2022 episode of Search Off the Record, John Mueller referred to the system as “the super search engine” and Gary Illyes used the term “universal mixer,” both naming the whole page algorithm that sits between the index and the surfaces brands try to win in. In February 2022 I extended the Bing-sourced thinking to Google in How Google Universal Search Ranking Works - Darwinism In Search on SEJ, where the survival-of-the-fittest framing made explicit that the multi-vertical surfaces all draw from the same underlying index. The most recent operational expression of the same architecture is in chunks, passages and micro-answer engine optimization wins in Google AI Mode for SEL in June 2025, where AI Mode synthesises answers from multiple background queries pulling passages from across the shared index. The substrate is shared. The mechanism is named. The architecture has been on the public record for more than six years.

The Knowledge Graph and the LLM do not maintain their own separate copies of the web. They work from cleaned subsets taken from the Web Index, with additional engineer-driven annotation passes layered on top. When engineers refresh the Knowledge Graph, they pull from the Web Index, clean it, apply additional structural annotation specific to graph topology, and update the graph on a periodic cycle. When engineers train or fine-tune an LLM, they pull from the Web Index, clean it, apply additional annotation specific to the model’s training objectives, and produce a refreshed parametric snapshot. The Web Index is the shared substrate. The Knowledge Graph and the LLM are downstream artefacts of engineer-curated extractions from that substrate, refreshed on cycles the brand never sees in advance. The Search Engine, by contrast, reads the Web Index live, in real time, on every query. I treat the relationship between continuous algorithmic annotation and periodic engineer-curated annotation in full in the companion Sandbox piece on the two annotation layers.

This architecture matters because it changes what recruitment actually means at each component. The Knowledge Graph and the LLM recruit when their respective engineering teams curate their training corpora from the Web Index. The Search Engine does not recruit periodically; it reads continuously, every time a query lands or an Assistive Engine triggers a grounding pass. With that distinction in place, the three recruitment logics make structural sense.

Recruitment is differentiated, but the differentiation is gradient not partition

Knowledge Graphs do not only recruit facts. LLMs do not only recruit gap fills. Search Engines do not only recruit novelty. Every component does some of every kind of recruitment. The differentiation captures dominant weighting, not strict functional partition. Knowledge Graphs sometimes ingest novelty when an emerging entity needs first-pass coverage. Search Engines sometimes prioritise corroborated facts when grounding an Assistive Engine’s claim against multiple sources. LLMs sometimes recruit purely confirmatory content to strengthen parametric knowledge that already exists.

That said, the dominant weighting is real, it is operationally useful, and it tells the brand where each component does its heaviest recruitment work. With the gradient honesty in place, the three logics are these.

Knowledge Graphs recruit for facts, attributes, relationships, and corroboration

A Knowledge Graph recruits because its core job is structural completeness. Its recruitment logic is built around four parallel questions. Is there an entity here I do not yet have? Is there an attribute on an entity I already have that I do not yet hold? Is there a relationship between two entities I have that I do not yet record? Does this content corroborate facts I already hold, raising my confidence in those facts?

Each of the four answers triggers a different recruitment outcome. New entities go through a creation pass. New attributes extend existing entity records. New relationships build the graph topology. Corroboration raises the confidence score on existing facts, which determines how willingly the Knowledge Graph surfaces those facts when other components query it.

Brands optimising for Knowledge Graph recruitment need annotation that names entities precisely, declares attributes explicitly, expresses relationships in machine-legible structure (schema.org, semantic HTML, structured citations), and corroborates the same facts across multiple independent sources. Vague annotation produces uncertain Knowledge Graph reads. Specific annotation produces confident Knowledge Graph entries. The differentiator is structural specificity, and the recruitment outcome depends on how well your annotation answers the four structural questions, with the engineer-curated extraction from the Web Index doing the actual recruitment on a periodic refresh cycle.

LLMs recruit for gap fills, confirmation, and bridges

LLMs recruit when engineering teams select training corpora to refresh or fine-tune the model, and the selection is driven by gaps the model itself has shown through usage. The LLM’s recruitment logic operates on three parallel questions. Does this content fill a gap in my parametric knowledge that the engineering team has identified through usage? Does this content confirm what I already hold parametrically, strengthening the pattern? Does this content bridge between something I know and something I could plausibly need to know, completing an inferential arc?

The gap-filling recruitment is the easiest to understand: engineers identify domains where the LLM underperforms, and content that addresses those gaps gets weighted in subsequent training corpus selection from the Web Index. The confirmatory recruitment matters because LLMs do not just need new information, they need information that strengthens what they already encode, because confidence in the existing pattern affects whether the model generates from it confidently or hedges.

The bridging recruitment is the most overlooked and the most strategically valuable. LLMs reason inferentially. If the model knows fact A and fact B but does not have content connecting A to B explicitly, the model’s reasoning across the gap is fragile. Content that bridges A to B gets recruited because the model can use the bridge to reason confidently across the gap thereafter. This is also the operational answer to the Framing Gap I named in why topical authority isn’t enough for AI search and unpacked in the framing gap: why AI can’t position your brand: the framing gap is not just a brand-side problem of supplying frames the AI cannot generate, it is also an LLM-side recruitment opportunity, because the engineering teams selecting training corpora are actively looking for bridging content, and the brand that supplies the bridge gets recruited to fill the inferential gap when the next training cycle runs.

Brands optimising for LLM recruitment need annotation that fills domain gaps, confirms existing knowledge in the model’s training distribution, and bridges between concepts the model already holds. The differentiator is inferential helpfulness, and the recruitment outcome depends on how well your content closes inferential arcs the model is reasoning across when the engineers next select the corpus.

Search Engines recruit for grounding, ranking, and real-time freshness at the passage level

Search Engines recruit differently from the Knowledge Graph and the LLM, because Search Engine recruitment is real-time rather than periodic, and increasingly it operates at the passage level rather than the page level. The bot has already filled the Web Index by the time a query lands. The Search Engine’s job at recruitment is not to fill the index - that is the bot’s job, running continuously in the background, as I described back in 2020 in How Bingbot Works: Discovering, Crawling, Extracting & Indexing on SEJ, sourced from Fabrice Canel of Microsoft Bing. The Search Engine’s job is to read the index live, decide which annotated content best serves the query in front of it, and surface that content as a ranked result, a grounding citation, or a real-time validation against current claims.

That makes the Search Engine’s recruitment logic operate on three real-time questions. Which annotated content in the index best answers the user’s query right now? Which annotated content best grounds the claim the Assistive Engine has just generated and needs to validate before showing it to the user? Which annotated content carries the freshest, most timestamped, most specific signal for the niche the query is about?

Google AI Mode adds a fourth dimension that sharpens the real-time recruitment logic considerably. As I covered in chunks, passages and micro-answer engine optimization wins in Google AI Mode, AI Mode synthesises answers from multiple background queries, with each background query pulling the best passage from the index for one facet of the user’s intent. The Search Engine is no longer recruiting whole pages for one query; it is recruiting passages for cascades of related queries and stitching the result. Inclusion at the passage level is what wins the assistive output. A brand that does not rank for the primary query can still appear in the AI Mode response if a single passage of its content wins recruitment for one of the cascading background queries. The recruitment unit has shrunk from the page to the passage. The recruitment logic remains real-time, but the granularity is smaller and the opportunity to be recruited is larger.

The ranking recruitment is what most SEOs already understand: relevant content, authoritative sources, the index pulling the best match for the query. The grounding-on-demand recruitment is the newest and most operationally important. When ChatGPT or Google AI Mode generates a claim and is uncertain, it triggers a real-time search against the live index to ground the claim against current corroborated sources. The content that gets grounded is the content that becomes the citation. The brand that produces grounding-ready content (clear claims, recent timestamps, structured corroboration, accessible to crawlers) wins the citation that becomes the assistive output. The freshness recruitment matters because the Search Engine reads the index live, which means the most recently crawled and annotated content has a genuine recruitment advantage when the query is time-sensitive.

Brands optimising for Search Engine recruitment need annotation that ranks well against query intent, presents grounding-ready claims with structural clarity at the passage level, and surfaces the kind of niche specificity and recency that wins real-time recruitment passes. The differentiator is real-time accessibility and currency, and the recruitment outcome depends on whether your content is reachable, readable, and timestamped well enough that the Search Engine reaches for it the moment a query needs it. As I put it in the SEL piece, inclusion at the passage level is everything.

Different engines weight the three components differently, which adds a second optimisation axis

Google AI Mode leans heavily on Knowledge Graph plus Search Engine, with LLM doing language-generation work over the top of grounded factual outputs. ChatGPT leans more on LLM with Search Engine grounding triggered when the model is uncertain. Perplexity’s architecture is search-grounded LLM, which means Search Engine recruitment is upstream of every claim the system makes. Claude relies heavily on its parametric LLM knowledge with retrieval-augmented generation when grounding is needed. Microsoft Copilot leans on Bing’s index plus the model. Gemini integrates Google’s Knowledge Graph more aggressively than most. Grok leans on real-time access to X (formerly Twitter) plus its own LLM.

The trajectory toward this engine-mix-weighting argument goes back to 2017, when I formalised Answer Engine Optimization (AEO) through a Trustpilot webinar and white paper that named the concept and put the framing on the working record. Search Engine Watch’s February 2018 coverage of the rise of AEO referenced the Trustpilot work and put the term on the public record by name, five years before ChatGPT existed. AEO is the upstream frame from which the Algorithmic Trinity (coined 2024) and the engine-mix observation that follows here are direct extensions, working the same architectural problem with more components named and sharper recruitment logic. I put the three-part trinity treatment on the public record in Search, answer, and assistive engine optimization: a 3-part approach for SEL in April 2025, laying out how to optimise across LLMs, knowledge graphs, and modern search engines. I extended it in chunks, passages and micro-answer engine optimization wins in Google AI Mode two months later, where I noted that the difference between Google AI Mode, ChatGPT, Perplexity, Microsoft Copilot, and the other assistive engines is simply the balance of the Algorithmic Trinity they use. What is original to this Sandbox piece is the next move: pairing the engine-mix observation with the differentiated recruitment logics per component, which gives the brand a structural map for deciding which annotation work matters most for which engine outcome.

Same brand. Same annotated content. Different engines weight the three components differently, which means the brand’s recruitment outcome varies by engine even when the annotation is identical. The strategic implication is that brands cannot optimise for one component without thinking about which engines that component most heavily serves, and brands cannot optimise for one engine without thinking about which components that engine most heavily uses.

Two stacked weightings, both gradient, both important to optimise across.

The brand running all three logics across the engines that matter wins recruitment that one-or-two-component brands lose

The brand that wins recruitment across the trinity runs all three logics in parallel. Knowledge Graph optimisation alone gives you facts the LLM can ground but not the bridging content the LLM needs to reason. LLM optimisation alone gives you bridges and confirmatory content but not the structural completeness Knowledge Graphs require for confident retrieval. Search Engine optimisation alone gives you currency, niche coverage, and grounding-readiness but not the parametric depth that affects how AI engines reason about your category before any specific query is asked.

The push layer entry modes I covered in the push layer returns: why ‘publish and wait’ is half a strategy are how the systematic operational practice runs continuously across all three logics. IndexNow accelerates the bot’s discovery and re-crawl cycles, which feeds the shared Web Index that all three components draw from. WebMCP handles the inference layer placement that affects LLM training corpus selection at the next periodic refresh. Structured data and the Entity Home covered in the entity home: the page that shapes how search, AI, and users see your brand handle the Knowledge Graph structural completeness layer at the next graph update. Each entry mode serves a different recruitment logic, and the brand running all three is the brand whose annotation reaches all three components at the recruitment cadence each component requires.

The operational practice that runs all three logics, weighted by the engines the brand most needs to win in, is what produces consistent recruitment across the trinity. The operational practice that runs one or two and hopes the third sorts itself out is what produces the inconsistent AI recommendations Rand Fishkin’s tests have been documenting since 2024 and that I covered in Rand Fishkin proved AI recommendations are inconsistent - here’s why and how to fix it.

Related reading from the AI authority series at Search Engine Land

Status: Original concept, first publication. Strategy Sandbox, jasonbarnard.com. Date: May 2026. The framework articulating differentiated recruitment logics across the three components of the Algorithmic Trinity is original to Jason Barnard. The Algorithmic Trinity itself was coined by Jason Barnard in 2024. The two-weightings argument (component weighting within the trinity, engine weighting across the components) is articulated here for the first time. Cite as: Barnard, J. (2026). Three Recruitment Logics: Why Annotation Has to Speak Differently to Knowledge Graphs, LLMs, and Search Engines. Strategy Sandbox, jasonbarnard.com.

SANDBOX PIECE B

Two Annotation Layers: Why Engineering-Driven Updates Like the Killer Whale Update Reward Brands With Systematic Operational Practice

Status: Original concept, first publication. Strategy Sandbox, jasonbarnard.com. Date: May 2026.

In July 2023, the Knowledge Graph went through an update that produced a roughly threefold increase in the volume of person entities I was tracking across Kalicube Pro™. I named it the Killer Whale Update, because the scale of the shift was orca-sized in a category that had been showing whale-shaped behaviour for years. The update was not announced. It was not documented in Google’s public release notes. It was visible only to people watching the Knowledge Graph closely enough to notice the volumetric shift, and it had specific operational consequences for the brands that had systematic person-entity work in place at the moment of the update versus the brands that did not.

What the Killer Whale Update revealed, when I sat with it for a few months and traced what had actually happened, is that there are two annotation layers feeding the Algorithmic Trinity, not one. The annotation gate I described for SEL is the continuous algorithmic layer running across the Web Index in real time. The second layer sits on top of it: a periodic engineer-curated annotation pass that selects, cleans, and labels training corpora for the LLM and the Knowledge Graph based on engineering priorities the brand never sees in advance.

Both layers matter. The continuous layer is what brands can shape through systematic operational practice. The periodic layer is what produces demand spikes the brand cannot anticipate but can be ready for. Understanding the two layers and how they relate is the difference between optimising for what the algorithms do every day and optimising for what the engineers occasionally decide they need.

The continuous layer is what brands shape through systematic operational practice

Continuous algorithmic annotation runs across the Web Index in real time. As content is crawled, it gets annotated by the algorithm: entity recognition, attribute extraction, relationship inference, topical classification, sentiment analysis, language detection, and so on. The annotation runs continuously, on every newly crawled URL and on URLs that have been re-crawled. It is the same annotation gate I covered in the SEL piece on how AI decides what your content means, and it is the gate that determines whether your content is legible to the rest of the 10-gate AI Engine Pipeline.

This layer is what brands shape through operational practice. Schema markup, semantic HTML, clear entity declarations, structured citations, consistent identifiers, accurate attribute statements, all the things SEO and entity optimisation cover. Every piece of annotation work the brand puts into this layer compounds across every gate downstream of it. The annotation is the upstream control point, and the brand that runs systematic annotation operationally is shaping a continuous stream of inputs to everything that happens further down the pipeline.

The continuous layer never stops. Every day, new content gets annotated. Every day, the index updates. Every day, the brand running systematic operational practice is feeding the annotation engine fresh material that gets weighted into the trinity over time.

The periodic layer is what engineers control and brands cannot anticipate

Engineers select, clean, and annotate the training corpora that feed the LLM and Knowledge Graph upstream of the systems brands interact with. The selection is not random. It is driven by engineering priorities: where the LLM is currently underperforming, which Knowledge Graph categories need denser coverage, which emerging fields need foundational data, which entity types are coming through the existing pipeline weakly enough to need a top-up.

The selection also draws from the underlying Web Index, which means the brand’s continuous annotation work feeds the periodic layer indirectly even though the brand never sees the engineer’s selection criteria. When the engineering team decides that person-entity coverage needs strengthening, they construct a corpus weighted toward person entities, and the corpus they construct draws from the Web Index pool the brand’s continuous annotation has been feeding. Brands with strong person-entity annotation in place at the moment of corpus selection get over-represented in the training data. Brands without it get under-represented. Both effects compound across the model’s lifecycle, because the training data shapes the model’s parametric knowledge for as long as the model runs.

The Killer Whale Update of July 2023 is the worked example. Engineers identified a coverage gap in person entities, constructed a corpus weighted accordingly, and the resulting Knowledge Graph update produced the threefold increase in person-entity volume I observed across Kalicube Pro. Brands with strong person-entity annotation in their continuous layer (clear About pages, structured author markup, consistent identifiers across the web, corroborated attribute statements) got recruited at scale during the update. Brands without it watched the window close.

Continuous work feeds periodic windows, but periodic windows cannot be timed

The continuous layer is shapeable by every brand that decides to run systematic annotation work. The periodic layer is unannounced, unpredictable in timing, and selective in what it amplifies. Brands that have systematic annotation work running continuously are positioned for whatever the next periodic update happens to need. Brands that try to retrofit annotation work in response to an announced update are too late, because the update is already complete by the time it shows up in the wild.

This is the asymmetry: continuous work feeds periodic windows, but periodic windows cannot be timed. The brand that waits for the announcement misses the update. The brand that runs systematic work catches every update, because the work was already in place when the corpus selection happened. There is no efficient just-in-time strategy for the periodic layer. There is only continuous preparation that catches windows when they open.

I have watched this play out across enough Knowledge Graph updates and enough LLM training cycles since 2015 to be confident the pattern holds. Engineering-driven update cycles produce demand spikes for specific entity types or knowledge domains. The Killer Whale Update was one. The June 2025 Knowledge Graph cleanup I covered for SEL was another, in the opposite direction (entities removed rather than added, but the same engineer-driven mechanism). LLM training cycles produce similar effects: when a model is being retrained or fine-tuned, the corpus selection produces compounding effects for brands that have continuous work in place that matches what the engineering priorities select for.

The continuous layer feeds all three recruitment logics, the periodic layer disproportionately affects the Knowledge Graph and the LLM

The first Sandbox piece in this set names the three differentiated recruitment logics across the Algorithmic Trinity: Knowledge Graphs recruit for facts and corroboration, LLMs recruit for gap fills and bridges, Search Engines recruit for novelty and grounding-on-demand. The two annotation layers fold into that argument cleanly.

The continuous layer feeds all three components of the trinity simultaneously, and each component recruits from the layer according to its own logic. A piece of continuously annotated content can be picked up by the Knowledge Graph for its structural completeness, by the LLM for its bridging value, and by the Search Engine for its novelty, all from the same annotation pass.

The periodic layer disproportionately affects the Knowledge Graph and the LLM, because those are the systems whose training corpora go through engineer curation. Search Engine recruitment is mostly continuous because Search Engine indexing is mostly continuous, but the Knowledge Graph and the LLM go through periodic refreshes that produce step-change effects in what they hold and how confidently they reason from it. The engineering selection criteria themselves often correspond to specific recruitment logics. A person-entity update is a Knowledge Graph structural-completeness pass. An LLM domain refresh is a gap-filling and bridging pass. A Search Engine index restructuring (which happens too, just less periodically) is a novelty and grounding-readiness pass.

The brand that runs continuous annotation work shaped for all three recruitment logics, in all three components of the trinity, is the brand whose continuous layer is ready for whatever periodic engineer-curation the next update happens to produce.

Brands that catch the next Killer Whale Update have systematic work running before the update happens

Most brands optimise annotation reactively. A schema validator flags an error and the brand fixes it. An AI overview gets a fact wrong and the brand updates the source. A competitor outranks them on a query and the brand investigates. This is case-by-case work, and it produces case-by-case recruitment, and the periodic windows close on brands running case-by-case work because the windows do not announce themselves.

The brand that wants to catch the next Killer Whale Update needs to have systematic annotation work running before the update happens. That means the push layer entry modes I covered for SEL running continuously through IndexNow and WebMCP. That means structured data running across every entity-bearing page. That means the Entity Home work I covered for SEL running as the foundational source of truth the rest of the digital footprint corroborates. That means corroboration backbones running across the second-party tier. That means inference layer placement running through thought leadership and academic citation.

The work is not glamorous. It is the operational discipline that produces continuous annotation feeding the continuous layer, which feeds the periodic layer when engineers happen to be selecting from it.

The Killer Whale Update is now nearly three years in the past. Whichever update the engineers run next, the brand that catches it is the brand that has systematic work in place when the corpus gets selected. The brand that does not is watching another window close.

Related reading from the AI authority series at Search Engine Land

Status: Original concept, first publication. Strategy Sandbox, jasonbarnard.com. Date: May 2026. The two-annotation-layers framework (continuous algorithmic at Web Index, periodic engineer-curated at training corpus level) is original to Jason Barnard. The Killer Whale Update of July 2023 was named retroactively by Jason Barnard based on Knowledge Graph person-entity volumetric shifts observed across Kalicube Pro tracking. Cite as: Barnard, J. (2026). Two Annotation Layers: Why Engineering-Driven Updates Like the Killer Whale Update Reward Brands With Systematic Operational Practice. Strategy Sandbox, jasonbarnard.com.

Jason Barnard is a serial entrepreneur, award-winning innovator, bestselling author, and acclaimed keynote speaker. He’s the CEO and founder of Kalicube, a premium Digital Branding Consultancy in France. Jason specializes in personal brand intelligence, giving business leaders control over how decision-makers perceive them on Google and AI when million-dollar decisions are at stake.

Similar Posts