From Keywords to Semantics to Epistemics: The Third Era of Search Optimisation

The Three Eras of Search Optimisation

The first era of keywords asked "Does this match?", evaluating whether strings on a page matched strings in a query. Optimisation meant exact-match placement, keyword density, and anchor text. The second era of semantics asked "What does this mean?", using embedding models to represent content as meaning so conceptually similar texts could be matched without shared words. The third era of epistemics asks "Is this true, is it new, and can I trust the source?" That three-part question defines the third era.

The semantic era has a failure mode called semantic saturation. When everyone optimises for meaning, and when AI writing tools produce semantically competent coverage of any topic in minutes, the top of every results set converges on content that means the same thing. Similarity cannot distinguish between ten pages that all say the same true things, and it cannot distinguish a verified claim from a fluent hallucination. A new signal is needed.

The Evidence Base: Four Documented Sources

Evidence 1 (information gain patent): Google's patent US11354342B2, "Contextual estimation of link information gain", describes assigning documents an information gain score measuring how much additional information a document contains beyond documents a user has already viewed on the same topic. The score is relational (a page cannot be scored in isolation) and claim-sensitive (what counts is the information added, not length or comprehensiveness). A 3,000-word article restating the top ten results adds nothing and scores accordingly.

Evidence 2 (generative summaries patent): Google's patent US11769017B1 describes using large language models to generate natural-language summaries of search results with the generated statements corroborated by links to source documents. The epistemically significant detail is the corroboration step: sources are selected because they can verify the statements the model wants to make, not only because they are relevant. A page can be perfectly relevant and still be useless for corroboration if its assertions are too vague to support anything.

Evidence 3 (Check Grounding API): Google sells claim-level verification as a commercial Vertex AI product. The Check Grounding API takes a piece of text and a set of source documents and returns a support score from 0 to 1 per sentence. Four documented properties matter: the unit of verification is the sentence, partial correctness fails entirely (one wrong detail sinks the whole sentence), it operates in real time under 500 milliseconds, and an experimental feature returns contradicting citations. This demonstrates sentence-level claim verification is productised infrastructure, not a patent abstraction.

Evidence 4 (official guidance): Google's documentation states its generative AI features rely on retrieval-augmented generation to retrieve relevant pages from the Search index. Retrieval remains semantic and rank-driven. A verification layer sits on top. You need the semantic layer to be retrieved. You need the epistemic layer to be cited.

The Five Layers of Epistemic Optimisation

• Claim Density: What proportion of your sentences are discrete, checkable assertions? Verification systems treat the sentence as the unit of analysis. Sentences carrying no claim are invisible to them. Target at least 40 percent factual claims per article.
• Groundability: Could an independent system verify your claims from your page as evidence? Vague claims cannot be entailed by anything. Compound claims fail if any part is weak. "Results are generally positive" cannot ground anything. "In a 2024 study of 1,200 Australian retailers, 68% reported increased organic traffic within six months of structured data implementation" can ground multiple things.
• Information Gain: What do you say that the rest of the candidate set does not? This is your citation inventory. First-party data, coined terms, and documented experience with named outcomes are the three highest-value claim types because they are the only kinds competitors cannot replicate.
• Provenance: Is each significant claim attributable to an identifiable entity with standing? Named authors, primary-source citations, and regulatory references (ACCC, AHPRA, ASIC) carry higher provenance than anonymous "experts" or generic "studies".
• Consistency: Do your claims agree with each other across the page and site, and do they position correctly against topic consensus? An unacknowledged inconsistency with authoritative sources is a grounding failure. An acknowledged, evidenced disagreement is a feature.

The Four-Step Claim Optimisation Process

Step 1: Extract and classify every sentence using an LLM into FACTUAL CLAIM, OPINION, HEDGE/FILLER, or NAVIGATION. Build a spreadsheet of the output as your work surface. Step 2: Triage every row with one of five actions: KEEP (already sourced), SPECIFY (true but vague), ANCHOR (attach existing source inline), REFRAME (convert unverifiable assertion to labelled interpretation), or CUT (filler with no claim or transition value). Step 3: Apply four rewrite patterns to SPECIFY, ANCHOR, and REFRAME rows: vague-to-specific (add number, timeframe, condition), split compound claims (one verifiable assertion per sentence), anchor attribution inline (not in a references section), and reframe-not-overclaim (separate documented fact from interpretation with explicit labelling). Step 4: Run the information gain delta by comparing your claim set against the top 8 to 10 competing pages to find unique claims that form your citation inventory and consensus claims you are missing.

Target pages that already rank in positions 3 to 15 but earn no AI citations. For those pages, retrieval is solved. Only the epistemic layer is failing. Improving claim specificity, inline attribution, and cross-page consistency on those pages will have the greatest measurable impact.