Dan Petrovic
Primary Bias
Primary bias is what an AI model already believes about your brand before it searches for anything. It is the model's ungrounded confidence in an entity, formed during training and baked into its weights, and it fires the moment a question is asked, before a single source is retrieved. In AI search this turns out to be the single largest factor in whether your content gets selected.
The clearest illustration came from a model reasoning about us. Reading our name, “Dejan”, it pattern-matched to the Balkans and began generating test queries for Serbian and Slovenian cities, when we are in fact Australian. The judgement landed before any other context was processed: as we described in our account at the time, the model had already formed an opinion before the conversation began.
Primary versus secondary bias
It helps to split the influences on selection into two layers. Primary bias is the model's inherent relevance perception of an entity, its pre-retrieval instinct. Secondary bias is everything about how your content is formatted, structured, and presented once it has been retrieved. The distinction matters because they move on very different timescales: secondary bias is addressable now, at the margin, while primary bias is slow to shift because it depends on training data.
Why it is hard to move
Primary bias lives in the model's weights, so changing it means changing what future models learn. Pre-training data is now heavily curated and very unlikely to respond to ordinary SEO. The realistic lever is fine-tuning, which authority-building can influence over roughly three to six months, with major model releases arriving about once a year. Our research on Selection Rate found that a brand strong in the training corpus can earn a high selection rate even with mediocre content, while a weak or confused brand struggles even when its page is retrieved.
How it connects to Selection Rate
Primary bias is measured through its effect on the Selection Rate: how often the model picks your source out of the grounding candidates. We frame primary bias as the dominant input to that rate, and use a probability-path method (our “Tree Walker”) to surface a brand's weakest associations, the high-uncertainty spots where reinforcement helps most.
What you can do about it
- Build a consistent, authoritative presence in the kinds of sources that feed training data.
- Disambiguate your entity explicitly, so a name does not collapse into the wrong association.
- Earn citations in authoritative contexts, across academic, industry, and media.
- Persist over time, so the signal survives across training cycles.
The uncomfortable part
Primary bias is pre-judgement in the literal sense, and it inherits the patterns in its training data, human ones included. Names, locations, and demographic signals can override the actual context in front of the model. The question is not whether these systems carry bias; they do. For anyone trying to be visible in AI answers, it is a force to understand and work with, not one that can be edited away on the day.
Related concepts
Evidence and sources
Grounding Snippets
When an AI system answers a question with current information, it does not recall the web from memory. It runs a search, pulls in pages, reads short extracts from them, and writes its answer from those extracts. Those extracts are grounding snippets, and they are the atomic unit of visibility in AI search. You can rank first in traditional search and still be invisible here, because the model, not the user, is now the reader, and a selective one.
The grounding funnel
Every platform runs the same basic pipeline, tuned differently: search query → pages received → pages with readable content → pages cited. Received is every URL the search step returned; readable is the subset the model actually obtained text for; cited is the few whose sources appear in the answer. The gap between received and cited is where each system shows its character.
What a grounding snippet is
A grounding snippet is built by extractive summarization, not abstractive: the system pulls exact sentences from your page rather than paraphrasing. The unit of extraction is the individual sentence, scored against the query, and the top sentences are stitched together. Where the chosen sentences are not next to each other on the page, they are joined by an ellipsis, producing the familiar segment … segment … segment shape. This is not unique to Google: testing Claude shows it returns the same ellipsis-joined, sentence-stitched format, so the pattern appears to be a shared convention across assistants.
How Google builds them
The pipeline runs prompt → query fanout → retrieval → extractive summarization → context assembly → synthesis and attribution. The observed traits of Google's extraction:
- Query-focused. Sentences close to the query are preferred; off-topic sections are skipped, and the same page yields different extracts for different fanout queries.
- Lead bias. Opening paragraphs are extracted almost wholesale, regardless of content.
- Structural noise. Tables of contents, headers, and stray markers get scored as if they were sentences.
- Confidence scores. Each chunk carries a relevance score from 0.1 to 1.0.
We replicate this behaviour closely by fine-tuning the open cross-encoder model microsoft/deberta-v3-large.
Query fanout
Before retrieval, the model breaks one prompt into several single-intent sub-queries, a separation of concerns where a multi-faceted question is split into individual dimensions of intent. Each sub-query retrieves its own set of sources, typically five to twenty. Because of fanout, a page can be grounded for one angle of a question and absent for another.
Only about a third survives
Most of your page never reaches the model. Across one sample analysis the system cited about 32% of the available characters, with per-source coverage ranging from roughly 21% to 65%. What gets kept is core service information, process steps, pricing and examples; what gets dropped is navigation, boilerplate, time-sensitive promotions, off-topic sections, and verbatim customer quotes.
A fixed budget, shared by rank
Grounding behaves like a fixed pie, not an expanding one. From an analysis of 7,060 queries, 2,275 pages and 883,262 snippets:
- Each query has a budget of roughly 2,000 words (median 1,929; p95 2,798), with the average chunk about 15.5 words.
- Rank sets your share. The top source takes about 531 words (28%); the fifth about 266 words (13%). The first source gets roughly double the grounding of the fifth.
- Coverage falls as pages get longer. Under 1,000 words, about 61% of the page is covered; at 3,000+ words, about 13%. Selection plateaus around 540 words, or about 3,500 characters.
The lesson is blunt: density beats length. More content dilutes your coverage without increasing what gets selected; you are competing for share of a fixed pie.
Grounding is transient
The snippets do not stick around. AI search is single-turn transient: the raw extracts are injected into the context for one turn, then purged the moment the answer is finished, to save token space. Ask a follow-up and the model is working from its own earlier summary, not the original page. What persists of you is whatever was captured in that first snippet, filtered through the model's reading of it, not the broader page.
Every platform decides differently
The same query, asked the same day, produced very different evidence on each platform:
- Google (Gemini) is economical, with a near one-to-one funnel: it tends to cite what it received and does not expose pages it did not use. Internally it indexes results as query.result pairs, around ten results per query, holding an ordered cache rather than loose text.
- OpenAI (GPT) casts the widest net and cites the least, a roughly twenty-to-one drop we call a visibility trap. It sees the web as a small structured card per result (title, URL, a one-to-three sentence snippet, an ID) and “scrolls” by opening fixed windows of text, never the full page.
- Anthropic (Claude) grounds in two passes, is the most token-hungry, and keeps some unselected pages visible so you can see what it considered.
A caveat on the evidence
The snippets a model exposes are not always reliable, even about themselves. In one case Gemini recited a grounding citation for a paper that does not exist, hallucinating while reporting its own grounding context. And the quantitative work above comes from our own measurements: we did not control for confounders such as authority and freshness, and we keep the raw data private for client confidentiality. Treat the numbers as strong directional signal rather than settled fact.
What it means for your content
- Write dense, self-contained sentences that stand on their own when extracted.
- Front-load the key information, since opening text is favoured.
- Cover the fanout angles a question splits into, not just the headline query.
- Strip structural noise so menus and boilerplate do not compete with your substance.
- Think in modular blocks, since the unit that wins is the sentence, not the page.
You can see this for yourself: our free grounding snippet tool at snippets.dejan.ai runs a live grounded search and shows which URLs Gemini pulls and the exact sentences it extracts, the practical groundwork behind Selection Rate Optimization.
Related concepts
- Grounding
- Grounding Snippet
- Query Fanout
- Selection Rate
- Selection Rate Optimization
- Primary Bias
- AI Visibility
Evidence and sources
- How AI search grounding works (the grounding funnel)
- What extraction method Google uses for grounding snippets
- How much of your content survives the AI search filter
- How big are Google’s grounding chunks
- Search grounding is transient
- How GPT sees the web
- How Google grounds Gemini
- Grounding snippet extraction tool
- SRO and grounding snippets
Selection Rate Optimization
Selection Rate Optimization, or SRO, is the AI-search counterpart to click-through-rate optimization. Where CTR optimization works to make a person more likely to click your result in a list of links, SRO works to make an AI system more likely to choose your content when it assembles an answer. The discipline aims at preferential treatment of a brand, its products, and its services inside AI search.
In AI search the user rarely sees a plain list of links. A language model sits between your content and the reader, in tools like Google's AI Mode and AI Overviews, Gemini, ChatGPT, and Perplexity. When the model answers a question it is handed several candidate sources to ground its reply on, reviews their snippets, and selects which to use. That selection is the moment SRO targets.
Selection Rate, the new click-through rate
The metric behind the practice is the Selection Rate: how often an AI system picks a given source out of the grounding candidates available to it. It is the AI equivalent of click-through rate. CTR measured a human decision among blue links; Selection Rate measures a model's decision among grounding choices. Raising it means your content is chosen, and therefore cited and represented, more of the time.
Why it matters
The model acts as an interpretive layer between your content and your audience. If it never selects your pages, your brand is absent from the answer no matter how well the page reads for people. Optimizing Selection Rate is how a brand regains some control over when and how it is presented across these systems.
How it works
The work runs in three broad stages.
- Project setup. Define the key entities, generate the fan-out queries and prompts a topic produces, identify the grounding candidates that surface for them, set the client as the optimization target and competitors as the field, then benchmark current performance.
- Optimization. Run a snippet optimization cycle, analyse the results, and produce a content optimization brief.
- Implementation. Apply the changes, measure the metrics, and iterate.
Grounding snippet reconstruction
The core of the pipeline rebuilds the snippet an AI system would extract from a page, approximating the extractive summarization used in Google's AI Mode and Gemini retrieval. Working from a faithful reconstruction of that snippet, the process can test changes against what the model actually sees rather than against the whole page.
The optimization cycle
The cycle runs the model backward from a desired outcome to find which words best produce it. Each candidate word carries a mathematical fingerprint that can be scored against the target ranking. A first pass shortlists candidate tokens filtered for natural readability; a second pass refines them against the model, balancing ranking impact against text that still reads naturally. Repeated across each position, this yields a short phrase that improves selection without reading as manipulation.
Related concepts
Relevance Engineering
Relevance Engineering is the practice of deliberately building a page's relevance to a query, using the same semantic machinery that search and AI systems use to judge it. Where traditional SEO tuned keywords, titles, and links, relevance engineering works with meaning directly: topics, pages, and queries are turned into embeddings, and relevance is measured as how close those vectors sit together.
The shift in the name is the point. The discipline treats search visibility as an engineering problem rather than an optimization exercise. You build toward a measurable target, semantic closeness, instead of nudging signals and hoping. The term was coined by Mike King of iPullRank.
How it works
- Meaning, not keywords. Content and queries are represented as vectors that capture meaning, following the move from lexical to semantic search.
- A center for each topic. Each core topic is represented by an average, central vector, so any page can be scored against it.
- Closeness as a score. The distance between a page's vector and its target topic gives a relevance score, rather than a guess.
- Decisions follow the score. Those numbers drive what to write, what to cut, how to cluster pages, and how to link them. Expertise can be measured the same way, by averaging the vectors of everything an author or site has published.
Relevance Engineering is the technical method beneath AI Visibility: making a page genuinely, measurably relevant is how it earns a place in the answers AI systems generate.
Related concepts
GEO (Generative Engine Optimization)
GEO, or Generative Engine Optimization, is the practice of optimizing for generative engines — the LLM-powered search and chat tools such as ChatGPT, Gemini, Perplexity, and AI Overviews. The goal is for your brand, pages, and ideas to appear within the text these systems generate.
GEO is one of the names for the work behind AI Visibility, measured through how often and how prominently you are mentioned and cited in AI answers.
Also known as
Related concepts
AI SEO
AI SEO is the practice of optimizing your content and brand so that AI systems surface them when they answer people's questions. It is the broadest of the names for this work: where traditional SEO chased rankings in a list of links, AI SEO targets presence inside the generated answer itself.
The outcome AI SEO works toward is AI Visibility — being seen, cited, and recommended across AI answers, measured through mentions and citations.
Also known as
Related concepts
AIO (AI Optimization)
AIO, short for AI Optimization, is the practice of shaping your content and online presence so AI systems favor them when generating responses. The name is deliberately broad, covering any optimization aimed at AI outputs rather than classic search rankings.
AIO is one route to AI Visibility, the outcome of being named and cited by AI systems when they answer questions.
Also known as
Related concepts
AEO (Answer Engine Optimization)
AEO, or Answer Engine Optimization, is the practice of optimizing so your content becomes the answer an answer engine gives. Answer engines — AI assistants, voice search, and featured answers — return a direct response rather than a page of links, so AEO focuses on being the quoted or cited source for a question.
AEO contributes to AI Visibility, the broader outcome of showing up prominently across AI-generated answers.
Also known as
Related concepts
The Open Knowledge Format (OKF)
The Open Knowledge Format, or OKF, is a new open standard from Google Cloud for packaging the knowledge an AI system needs so that any model or agent can read it. It takes the informal habit of keeping an "AI wiki" next to your work and turns it into a portable, vendor-neutral format. The announcement and the full specification are linked at the end.
The problem it solves
Foundation models are only as good as the context they are given, and in most organisations that context is scattered. Table schemas, metric definitions, runbooks, join paths, and the reasoning behind past decisions live in catalogs, wikis, shared drives, code comments, and the heads of a few senior people. Each tool stores this knowledge in its own shape behind its own API, so it does not travel. Every team that builds an agent re-solves the same job of gathering context, and every catalog vendor reinvents the same data model.
OKF answers this with a format that anyone can produce without an SDK, anyone can consume without an integration, and that survives being moved between systems. It lives in version control beside the code it describes, and the same file is readable by a person and parseable by an agent.
How it works
An OKF bundle is simply a directory of markdown files. The rules are deliberately small enough to fit on a single page.
- One concept per file. A concept is anything worth capturing: a table, a dataset, a metric, a playbook. The file's path is its identity, so
tables/orders.mdis the concepttables/orders. - Markdown with YAML frontmatter. Each file opens with a small block of structured fields. The only required field is
type. Recommended fields aretitle,description,resource(a link to the underlying asset),tags, andtimestamp. You may add any other fields you like. - Links make a graph. Concepts reference each other with ordinary markdown links, which turns the folder into a graph of relationships rather than a plain tree.
- Two reserved files. An
index.mdgives a directory listing so an agent can see what is available before opening files, and alog.mdrecords changes over time, newest first.
Because it is just markdown and files, a bundle renders on GitHub, opens in any editor, ships as a tarball, and is indexed by any search tool. If you can read a file with cat, you can read OKF.
Three principles
- Minimally opinionated. Exactly one field is mandatory. The spec defines how systems interoperate, not what you must write.
- Producer and consumer are independent. A human can author a bundle that an agent consumes, or a pipeline can generate one that a person browses. The format is the contract, and the tools at each end are swappable.
- A format, not a platform. It is tied to no cloud, database, model, or agent framework, and never requires an account to read or write.
What comes with it
Google shipped working proofs alongside the spec: a reference agent that drafts a bundle from a BigQuery dataset and then enriches it by crawling authoritative documentation, a self-contained HTML visualiser that renders any bundle as an interactive graph, and three sample bundles built from public datasets.
Why it matters for AI visibility
This is the same idea that sits under AI visibility, seen from the supply side. If you want models to represent your work accurately, the knowledge they rely on has to be legible to them. OKF is a clean way to make that knowledge portable and machine-readable, so the systems that answer questions about your domain can ground themselves in what you actually said rather than guessing.
Read more
How the Open Knowledge Format can improve data sharing is the announcement. The specification and reference code live in GoogleCloudPlatform/knowledge-catalog.
Related concepts
AI Visibility
AI Visibility is the broad outcome that AI-centric SEO work aims to achieve. It means being seen, cited, and recommended by AI systems when they answer people's questions. Traditional SEO aimed at ranking in a list of links. AI Visibility aims at presence inside the generated answer itself: whether your brand, your pages, and your ideas show up when a language model responds, and how prominently they do.
It is a desired outcome rather than a single number. Underneath it sits a family of measurable signals that together tell you how visible you are across AI answers.
How it is measured
People track AI Visibility through a set of related metrics. These divide into two kinds of presence: mentions, where your brand is named in the answer, and citations, where your pages are linked or referenced as a source. Each kind is measured by share, by absolute count, by how often it happens, and by how prominently it appears.
- Share of voice. Your overall slice of the AI answer space for a topic, relative to competitors.
- Share of citations. The proportion of cited sources in answers that are yours.
- Share of mentions. The proportion of brand mentions in answers that are yours.
- Number of citations. The absolute count of times your pages are cited.
- Number of mentions. The absolute count of times your brand is named.
- Frequency. How often citations and mentions occur across a body of queries and over time.
- Rank of citations and mentions. How prominently each appears, for example whether your source is cited first or your brand named first, rather than sitting lower in the answer.
Read together, these tell a full story. Share metrics show your standing against competitors, counts show absolute reach, frequency shows consistency, and rank shows prominence. Strong AI Visibility means appearing often, as a meaningful share of the answer, and near the top of what the model surfaces.
Also known as
AI Visibility is the outcome. These are common names for the practice of working toward it: