How AI decides what your content means and why it gets you wrong

Google as soon as attributed two of Barry Schwartz’s Search Engine Land articles to me — a misclassification on the annotation layer that briefly rewrote authorship in Google’s techniques.

For just a few days, while you looked for sure Search Engine Land articles Schwartz had written, Google listed me because the creator. The articles appeared in my entity’s publication listing and had been linked to my Data Panel.

What occurred illustrates one thing the web optimization business has nearly totally missed: that annotation — not the content material itself — is the important thing to what customers see and thus your success.

How Google annotated the web page and bought the creator unsuitable

Googlebot crawled these pages, discovered my title prominently displayed beneath the article (my creator bio appeared as the primary acknowledged entity title beneath the content material), and the algorithm on the annotation gate added the “Publish-It” that categorised me because the creator with excessive confidence.

That is a very powerful level to keep in mind: the bot can misclassify and annotate, and that defines every little thing the algorithms do downstream (in recruitment, grounding, show, and gained). On this case, the problem was authorship, which isn’t going to kill my enterprise or Schwartz’s.

But when that had been a product, a worth, an attribute, or the rest that issues to the intent of a consumer search question the place your model ought to be one of many apparent candidates, when any side of content material is inaccurately annotated, you’ve misplaced the “rating recreation” earlier than you even began competing.

Annotation is the one most necessary gate in taking your model from uncover to gained, no matter question, intent, or engine you’re optimizing for.

What annotation is and why it isn’t indexing

Indexing (Gate 4) breaks your content material into semantic chunks, converts it, and shops it in a proprietary format. Annotation (Gate 5) then labels these chunks with a confidence-driven “Publish-It” classification system.

It’s a realistic labeler and attaches classifications to every chunk, describing:

• What that chunk comprises factually.
• In what circumstances it could be helpful.
• The trustworthiness of the knowledge.

Importantly, it’s principally unopinionated when labeling details, context, and trustworthiness. Microsoft’s Fabrice Canel confirmed the precept that the bot tags with out judging, and that filtering occurs at question time.

What does that imply? The bot annotates neutrally at crawl time, classifying your content material with out figuring out what question will ultimately set off retrieval. 

Annotation carries no intent in any respect. It’s the perception that has fully modified my method to “crawl and index.”

That clearly reveals you that indexing isn’t the last word purpose. Getting your web page listed is desk stakes. Full, right, and assured annotation is the place the motion occurs: an listed web page that’s poorly annotated is invisible to every of the algorithmic trinity.

The annotation system analyzes every chunk utilizing a number of language fashions, cross-referenced in opposition to the online index, the information graph, and the fashions’ personal parametric information. But it surely analyzes every chunk within the context of the page wrapper. 

The page-level matter, entity associations, and intent present the body for classifying every chunk. If the page-level understanding is confused (unclear matter, ambiguous entity, blended intent), each chunk annotation inherits that confusion. Much more importantly, it assigns confidence to each piece of data it provides to the “Publish-Its.”

The alternatives occur downstream: every of the algorithmic trinity (LLMs, search engines like google, and information graphs) makes use of the annotation to resolve whether or not to soak up your content material at recruitment (Gate 6). Every has totally different standards, so it’s worthwhile to assess your personal content material for its “annotatability” within the context of all three.

And a small however telling element: Again in 2020, Martin Splitt prompt that Google compares your meta description to its personal LLM-generated abstract of the web page. After they match, the system’s confidence in its page-level understanding will increase, and that confidence cascades into higher annotation scores for each chunk — considered one of 1000’s of tiny indicators that accumulate.

Annotation is the important thing midpoint of the 10-gate pipeline, the place the scoreboard activates. Every thing earlier than it’s infrastructure: “Can the system entry and retailer your content material?” Every thing after it’s competitors:

When you think about what occurs on the annotation gate and its depth, hyperlinks and key phrases grow to be the unsuitable lens totally. They describe the way you tried to affect a rating system, whereas annotation is the mechanism behind how the algorithmic trinity chooses the content material that builds its understanding of what you might be.

The body has to shift. You’re educating algorithms. They behave like kids, studying from what you constantly, clearly, and coherently put in entrance of them. With constant, corroborated info, they construct an correct understanding.

Given inconsistent or ambiguous indicators, they be taught incorrectly after which confidently repeat these errors over time. Constructing confidence within the machine’s understanding of you is a very powerful variable on this work, whether or not you name it web optimization or AAO.

In 2026, each AI assistive engine and agent is that very same baby, working at a larger scale and with increased stakes than Google ever had. Educating the algorithms isn’t a metaphor. It’s the operational mannequin for every little thing that follows.

For a extra tutorial perspective, see: “Annotation Cascading: Hierarchical Model Routing, Topical Authority, and Inter-Page Context Propagation in Large-Scale Web Content Classification.”

5 ranges of annotation: 24+ dimensions classifying your content material at Gate 5

When mapping the annotation dimensions, I recognized 24, organized throughout 5 practical classes. After presenting this to Canel, his response was: “Oh, there may be undoubtedly extra.”

After all there are. This taxonomy is constructed via commentary first, then naming what constantly seems. The [know/guess] distinctions comply with the identical logic: check hypotheses, get rid of what doesn’t maintain up, and maintain what stays.

The 5 practical classes kind the muse of the mannequin. They’re easy by design — when you perceive the classes, the size comply with naturally. There are probably further dimensions past these mapped right here.

What follows is the taxonomy: the classes are directionally sound (as confirmed by Canel), whereas the particular dimension assignments replicate noticed conduct and stay incomplete.

Stage 1: Gatekeepers (get rid of)

• Temporal scope, geographic scope, language, and entity decision. Binary: move or fail. 
• In case your content material fails a gatekeeper (unsuitable language, unsuitable geography, or ambiguous entity), it’s eradicated from that question’s candidate pool immediately. The opposite dimensions don’t come into play.

Stage 2: Core identification (outline)

• Entities, attributes, relationships, sentiment. 
• That is the place the system decides what your content material means:
  • Who’s being mentioned.
  • What details are acknowledged.
  • How entities relate.
  • What the tone is. 
• With out clear core identification annotations, a bit carries no semantic weight in any downstream gate.

Stage 3: Choice filters (route) 

• Intent class, experience degree, declare construction, and actionability. 
• These decide which competitors pool your content material enters.
  • Is that this informational or transactional? 
  • Newbie or professional? 
• Mistaken pool placement means competing in opposition to content material that could be a higher match for the question, and also you’ve misplaced earlier than recruitment or rating begins.

Stage 4: Confidence multipliers (rank)

• Verifiability, provenance, corroboration depend, specificity, proof kind, controversy degree, and consensus alignment. These scale your rating throughout the pool. 
• That is the place validated, corroborated, and particular content material outranks correct however unvalidated content material. 
• The multipliers clarify why a well-sourced third-party article about you usually outperforms your personal claims: provenance and corroboration scores are increased.
• Confidence has a multiplier impact on every little thing else and is essentially the most highly effective of all indicators. Full cease.

• Sufficiency, dependency, standalone rating, entity salience, and entity position. These decide how your content material seems within the ultimate output. 
• Is that this chunk an entire reply, or does it want context? Is your entity the topic, the authority cited, or a passing point out? 
• Extraction high quality determines whether or not AI quotes you, summarizes you, or ignores you.

Throughout all 5 ranges, a confidence rating is hooked up to each particular person annotation. Not simply what the system thinks your content material means, however how sure it’s.

Readability drives confidence. Ambiguity kills it.

Canel additionally confirmed further dimensions I had not initially mapped: viewers suitability, ingestion constancy, and freshness delta. These sit throughout the prevailing classes slightly than forming a sixth degree.

In 2022, Splitt named three annotation behaviors in a Duda webinar that map straight onto the five-level mannequin. The centerpiece annotation is Stage 2 in direct operation: 

• “We now have a factor known as the centerpiece annotation,” Splitt confirmed, a classification that identifies which content material on the web page is the first topic and routes every little thing else — supplementary, peripheral, and boilerplate — relative to it. 
• “There’s just a few different annotations” of this kind, he famous. 

Annotation runs earlier than recruitment, which implies a bit categorised as non-centerpiece carries that verdict into each gate that follows. Boilerplate detection is Stage 3: content material that seems constantly throughout pages — headers, footers, navigation, and repeated blocks — enters a special competitors pool primarily based on its structural position alone. 

• “We determine what seems to be like boilerplate after which that will get weighted otherwise,” Splitt stated 

Off-topic routing closes the image. A web page categorised round a major matter annotates each chunk relative to that centerpiece, and content material peripheral to the first matter begins its personal competitors pool at an obstacle earlier than Recruitment begins. 

Splitt’s instance: a web page with 10,000 phrases on pet food and a thousand on bikes is “in all probability not good content material for bikes.” The system isn’t ignoring the bike content material. It’s annotating it as peripheral, and that annotation is the routing choice.

The multiplicative destruction impact: When one near-zero kills every little thing

In Sydney in 2019, I used to be at a convention with Gary Illyes and Brent Payne. Illyes defined that Google’s high quality evaluation throughout annotation dimensions was multiplicative, not additive. 

Illyes requested us to not movie, so I grabbed a beer mat and famous a easy calculation: for those who rating 0.9 throughout every of 10 dimensions, 0.9 to the facility of 10 is 0.35. You survive at 35% of your unique sign. In the event you rating 0.8 throughout 10 dimensions, you survive at 11%. If one dimension scores near zero, the multiplication produces a outcome near zero, no matter how nicely you rating on each different dimension.

Payne’s phrasing of the sensible implication was higher than mine: “Higher to be a straight C scholar than three As and an F.”

The beer mat went into my bag. The precept turned central to every little thing I’ve constructed since.

The multiplicative destruction impact has a direct consequence for annotation technique: the C-student precept is your information. 

• A model with constantly satisfactory indicators throughout all 24+ dimensions outperforms a model with sensible indicators on most dimensions and a near-zero on one. The near-zero cascades. 
• A gatekeeper failure (Stage 1) eliminates the content material totally. 
• A core identification failure (Stage 2) misclassifies it so badly that top confidence multipliers at Stage 4 are utilized to the unsuitable entity. 
• An extraction high quality failure (Stage 5) produces a bit that the system can retrieve however can’t deploy usefully. The failure doesn’t should be dramatic to be deadly.

On the annotation stage, misclassification, low confidence, or near-zero on one dimension will kill your content material and take it out of the race.

Nathan Chalmers, who works at Bing on high quality, advised me one thing that places this in a special gentle totally. Bing’s inside high quality algorithm, the one making these multiplicative assessments throughout annotation dimensions, is literally called Darwin. 

Pure choice is the express mannequin: content material with near-zero on any health dimension is chosen in opposition to. The annotations are the health check. The multiplicative destruction impact is the choice mechanism.

How annotation routes content material to specialist language fashions

The system doesn’t use one large language mannequin to categorise all content material. It routes content material to specialised small language fashions (SLMs): domain-specific fashions which might be cheaper, quicker, and paradoxically extra correct than common LLMs for area of interest content material. 

A medical SLM classifies medical content material higher than GPT-4 would, as a result of it has been educated particularly on medical literature and is aware of the entities, the relationships, the usual claims, and the crimson flags in that area.

What follows is my mannequin of how the routing works, reconstructed from observable conduct and confirmed rules. The existence of specialist fashions is confirmed. The precise cascade mechanism is my reconstruction.

The routing follows what I name the annotation cascade. The selection of SLM cascades like this:

• Website degree (What sort of website is that this?)
• Refined by class degree (What part?)
• Refined by web page degree (what particular matter?)
• Utilized at chunk degree (What does this paragraph declare?)

Every degree narrows the SLM choice, and every degree both confirms or overrides the routing from above. This maps on to the wrapper hierarchy from the fourth piece: the positioning wrapper, class wrapper, and web page wrapper every present context that influences which specialist mannequin the system selects.

The system deploys three sorts of SLM concurrently for every matter. That is my mannequin, derived from the conduct I’ve noticed: annotation errors cluster into patterns that recommend three distinct classification axes. 

• The topic SLM classifies by subject material — what is that this about? — routing content material into the precise topical area. 
• The entity SLM resolves entities and assesses centrality and authority: who’re the important thing gamers, is that this entity the topic, an authority cited, or a passing point out? 
• The idea SLM maps claims to established ideas and evaluates novelty, checking whether or not what the content material asserts aligns with consensus or contradicts it.

When all three return excessive confidence on the identical entity for a similar content material, annotation price is minimal, and the boldness rating may be very excessive. After they disagree (i.e., the topic SLM says “advertising,” however the entity SLM can’t resolve the entity, and the idea SLM flags the claims as novel), confidence drops, and the system falls again to a extra common, much less correct mannequin.

The important thing perception? LLM annotation is the failure mode. The system needs to make use of a specialist. It defaults to a generalist solely when it will probably’t path to a specialist. Generalist annotation produces decrease confidence throughout all dimensions. 

The sensible implication 

Content material that’s category-clear inside its first 100 phrases, makes use of customary business terminology, follows structural conventions for its content material kind, and references well-known entities in its area triggers SLM routing. 

Content material that’s topically ambiguous or terminologically artistic will get the generalist. Decrease confidence propagates via each downstream gate.

Now, this is probably not the precise manner the SLMs are utilized as a triad (and it won’t even be a trio). Nevertheless, two issues strike me:

• Noticed outputs act that manner.
• If it doesn’t operate this manner, it will be.

First-impression persistence: Why the preliminary annotation is the toughest to right

Right here is one thing I’ve noticed over years of monitoring annotation conduct. It aligns with a precept Canel confirmed explicitly for URL status changes (404s and 301 redirects): the system’s preliminary classification tends to stay.

When the bot first crawls a web page, it selects an SLM, runs the annotation, assigns confidence scores, and saves the classification. The subsequent time it crawls the identical web page, it logically begins with the beforehand assigned mannequin and annotations. I name this first-impression persistence. 

The preliminary annotation is the baseline in opposition to which all subsequent indicators are measured. The system doesn’t re-evaluate from scratch. It checks whether or not the brand new crawl is in line with the prevailing classification, and whether it is, the classification is bolstered.

Canel confirmed a associated mechanism: when a URL returns a 404 or is redirected with a 301, the system permits a grace interval (very roughly per week for a web page, and between one and three months for content material, in my commentary) throughout which it assumes the change would possibly revert. After the grace interval, the brand new state turns into persistent. I imagine the identical precept applies to content material classification: a window of fluidity after first publication, then crystallization.

I’ve direct proof for the correction aspect from the evolution of my very own terminologies. After I first described the algorithmic trinity, I used the phrase “information graphs, giant language fashions, and net index.” Google, ChatGPT, and Perplexity all picked up on the brand new time period and outlined it appropriately.

A month later, I modified the final one to “search engine” as a result of it occurred to me that the online index is what all three techniques feed off, not simply the search system itself. On the level of correction, I had revealed roughly 10 articles utilizing the unique terminology. 

I went again and invested the time to alter each single one, updating each reference, leaving zero traces. A month later, AI assistive engines had been constantly utilizing “search engine” rather than “net index.”

The lesson is that change is feasible, however it’s worthwhile to be thorough: any residual contradictory sign (one previous article, one unchanged social put up, and one cached model) maintains inertia proportionally. Thoroughness is the unlock, slightly than time.

A rebrand, profession pivot, or repositioning is the sensible instance. You possibly can change the AI mannequin’s understanding and illustration of your company or private model, but it surely requires completely and constantly pivoting your digital footprint to the brand new actuality.

In my expertise, “on a sixpence” inside per week. I’ve achieved this with my podcast a number of instances. Fb achieved the last word rebrand from an algorithmic perspective when it modified its title to Meta.

The sensible implication

Get your annotation proper earlier than you publish. The primary crawl units the baseline. A web page revealed prematurely (with an unclear matter or ambiguous entity indicators) crystallizes right into a low-confidence annotation, and altering it later requires considerably extra effort than getting it proper the primary time.

Annotation-time grounding: The bot cross-references three sources whereas classifying your content material

The system doesn’t annotate in a vacuum. When the bot classifies your content material at Gate 5, it cross-references in opposition to not less than three sources concurrently. That is my mannequin of the mechanism. The observable impact — that annotation confidence correlates with entity presence throughout a number of techniques — is confirmed from our monitoring information.

The bot carries prioritized entry to the online index throughout crawling, checking your content material in opposition to what it already is aware of: 

• Who hyperlinks to you.
• What context these hyperlinks present.
• How your claims relate to claims on different pages. 

Towards the information graph, it checks annotated entities throughout classification — an entity already within the graph with excessive confidence means annotation inherits that confidence, whereas absence begins from a a lot decrease baseline. 

The SLM’s personal parametric information gives the third cross-reference: every SLM compares encountered claims in opposition to its coaching information, granting increased confidence to claims that align, flagging contradictions, and giving decrease confidence to novel claims till corroboration accumulates.

This implies annotation high quality isn’t nearly how nicely your content material is written. It’s about how nicely your entity is already represented throughout all three of the algorithmic trinity. An entity with sturdy information graph presence, authoritative net index hyperlinks, and constant SLM-domain illustration will get increased annotation confidence on new content material routinely. 

The flywheel: higher presence results in higher annotation, which ends up in higher recruitment, which strengthens presence, and which improves future annotation.

As soon as once more, higher to have a mean presence in all three than to have a dominant presence in two and no presence in a single.

And that is why information graph optimization (what I’ve been advocating for over a decade) isn’t separate from content material optimization. They’re the identical pipeline. Your information graph presence straight improves how precisely, verbosely, and confidently the system annotates each new piece of content material you publish.

In the event you’re pondering “Data graph? That’s simply Google,” suppose once more.

In November 2025, Andrea Volpini intercepted ChatGPT’s internal data streams and located an operational entity layer working beneath each dialog: structured entity decision linked to what quantities to a product graph mirroring Google Buying feeds. 

OpenAI is constructing its personal information graph contained in the LLM. My wager is that they may externalize it for a number of causes: a information graph in an LLM doesn’t scale, an LLM will self-confirm, so the worth is proscribed, a standalone information graph will be simply up to date in actual time with out retraining the mannequin, and it’s solely helpful at scale when it stays present.

The algorithmic trinity isn’t a Google phenomenon. It’s the architectural sample each AI assistive engine and agent converges on, as a result of you possibly can’t generate dependable suggestions with out a idea graph, structured entity information, and up-to-date search outcomes to floor them.

Why Google and Bing annotate otherwise from engines that hire their index

Google and Bing personal their crawling infrastructure, indexes, and information graphs. They will afford grace durations, schedule rechecks, and preserve temporal state for URLs and entities over months.

OpenAI, Perplexity, and each engine that rents index entry from Google or Bing function on a essentially totally different mannequin. They’ve two speeds: 

• A gradual Boolean gate (Does this content material exist within the index I’ve entry to?)
• A quick show layer (What does the content material say proper now after I fetch it for grounding?)

The Boolean gate inherits Google’s and Bing’s annotations. Whether or not your content material seems in any respect depends upon whether or not it was recruited from the index these engines draw from, and that recruitment depends upon annotation and choice selections made by the algorithmic trinity. However what these engines present after they cite you is fetched in actual time.

The sensible implication

For Google and Bing, you’re optimizing for annotation high quality with the good thing about grace durations and gradual reclassification. For engines that don’t personal their index, the Boolean presence is inherited from the rented index and is gradual to alter, however the surface-level show modifications each time they re-fetch.

Meaning what you might be seeing within the outcomes will not be a direct measure of your annotation high quality. It’s a snapshot of your web page in the intervening time of fetch, and people two issues could don’t have anything to do with one another.

The right way to optimize for annotation high quality: The six sensible rules

The web optimization business has spent twenty years optimizing for search and assistive outcomes — what occurs after the system has already determined what your content material means. We ought to be optimizing for annotation. 

If the annotation is unsuitable, every little thing downstream suffers. When the annotation is correct, verbose, and assured, your content material has a major benefit in recruitment, grounding, show, and, in the end, gained.

1. Set off SLM routing

Make your matter class apparent throughout the first 100 phrases. Use customary business terminology. Observe structural conventions. Reference well-known entities. The purpose: specialist mannequin, not generalist.

2. Write for all three SLMs

Clear indicators for topic (what is that this about?), entity (who’s the authority?), and idea (what established concepts does this hook up with?). Ambiguity on any axis reduces confidence.

3. Get it proper earlier than publishing

First-impression persistence means the preliminary annotation is the toughest to alter. Publish solely when matter, entity indicators, and claims are unambiguous.

4. Construct the flywheel

Data graph presence, net index centrality, LLM parameter strengthening, and proper SLM-domain illustration all feed annotation confidence for brand spanking new content material. Put money into entity basis, and each future piece advantages from inherited credibility.

5. Get rid of noise when correcting

Change each reference. Depart zero contradictory indicators. Noise maintains inertia proportionally.

6. Audit for annotation, not simply indexing

A web page will be listed and nonetheless misannotated. If the AI response is unsuitable about you, the issue is nearly actually at Gate 5, not Gate 8.

Annotation is the gate the place most manufacturers silently lose. The web optimization business doesn’t but have a vocabulary for it. That should change, as a result of the hole between manufacturers that get annotation proper and types that don’t is the hole between constant AI visibility and everlasting algorithmic obscurity.

Why annotation issues a lot and why it ought to be your principal focus

You’ve achieved every little thing inside your energy to create the very best content material that maps to intent of your ideally suited buyer profile, you’ve methodically optimized your digital footprint, your information feeds each entry mode concurrently: pull, push discovery, push information, MCP, and ambient, so they’re all drawing from the identical clear, constant supply

So, content material about your model has handed via the DSCRI infrastructure section, survived the rendering and conversion constancy boundaries, and arrived within the index (Gate 4) intact. Phew!

Now it will get categorised. Annotation is the final second within the pipeline the place you’ve the sector to your self. Each choice in DSCRI was absolute: you vs. the machine, with no competitor within the body. 

Annotation remains to be absolute. The system classifies your content material primarily based in your indicators alone, independently of what any competitor has achieved. No one else’s information modifications how your entity is annotated.

However that is the final time you aren’t competing. From recruitment onward, every little thing is relative. The sphere opens, each model that handed annotation enters the identical aggressive pool, and the benefit you carried via absolutely the section turns into your beginning place within the aggressive race you need to win.

Meaning: 

• Get annotation proper, and also you begin forward, with confidence that compounds via each downstream gate in RGDW. 
• Get it unsuitable, and the multiplicative destruction impact does its work — a near-zero on one annotation dimension cascades via recruitment, grounding, show, and gained. No quantity of wonderful content material, structural indicators, or entry-mode benefit recovers it.

Warning: First-impression persistence (keep in mind, the primary time you might be annotated is the baseline) means you don’t get a clear retry. Altering the baseline requires thoroughness, time, and extra effort than getting it proper on the primary crawl.

Annotation isn’t the gate that almost all manufacturers concentrate on. It’s the gate the place most manufacturers silently lose.

That is the eighth piece in my AI authority collection. 

• The primary, “Rand Fishkin proved AI recommendations are inconsistent – here’s why and how to fix it,” launched cascading confidence. 
• The second, “AAO: Why assistive agent optimization is the next evolution of SEO,” named the self-discipline. 
• The third, “The AI engine pipeline: 10 gates that decide whether you win the recommendation,” mapped the total pipeline. 
• The fourth, “The five infrastructure gates behind crawl, render, and index,” walked via the infrastructure section.
• The fifth, “5 competitive gates hidden inside ‘rank and display’,” coated the aggressive section.
• The sixth, “The entity home: The page that shapes how search, AI, and users see your brand,” mapped the uncooked materials.
• The seventh, “The push layer returns: Why ‘publish and wait’ is half a strategy,” prolonged the entry mannequin. 
• Up subsequent: “The engine’s recruitment choice: What topical possession really means.”