Changelog

First AI engine citation: Perplexity surfaced this entry as a primary source for 'What is AI search evaluation?', citing the pillar above the fold alongside the opening definition. First engine to cite the entry since publication; 1 of 5 tested engines now cites it. Perplexity reached this explainer directly rather than the underlying academic sources.

First AI engine citations: ChatGPT, Perplexity, and Claude all surfaced this entry for 'What is AIPREF (AI usage preferences)?', each drawing on our preferences-versus-authentication distinction (AIPREF declares usage preferences; it does not authenticate the requester, which is the separate Web Bot Auth effort). From 0 to 3 of 5 tested engines citing the entry.

First AI engine citation: Perplexity cited this entry for 'What is authority signals?', drawing on our citations-and-backlinks breakdown alongside established SEO glossaries. 1 of 5 tested engines now cites it.

Second AI engine citation: Perplexity surfaced this entry as its top source for 'What is Black-hat C-SEO?', listing it first in the answer's source list and drawing on our prompt-injection-and-deception framing. Joins Gemini; 2 of 5 tested engines now cite it.

Copilot citation confirmed for 'What is Citation Footprint?', reproducing our coined framing (a cumulative, monotonic count of distinct cited URLs measuring coverage rather than intensity, distinct from citation share and citation velocity). Fourth engine to cite this coined entry; 4 of 5 tested engines now cite it.

First AI engine citations: ChatGPT and Perplexity both surfaced this entry for 'What is Citation hallucination?', with ChatGPT ranking it the top source and Perplexity drawing on our core distinction (a hallucinated citation points to a source that does not exist, separate from misquoting a real source or answering with none). 2 of 5 tested engines now cite it.

Gemini citation confirmed for 'What is cite-ability in AI search?', with our entry dominating the source panel and Gemini reproducing the four-trait framing (context-free quote, answer-first layout, factual density, echo-and-expand). Fourth engine to cite this practitioner-coined entry; 4 of 5 tested engines now cite it.

First AI engine citations: Perplexity and Gemini both surfaced this entry as their top source for 'What is Deep research mode?', each reproducing our framing that deep research mode is an escalation of agentic retrieval applying query fan-out at much larger scale. 2 of 5 tested engines now cite it.

ChatGPT citation confirmed for 'What is Freshness signals?', surfacing this entry as the top source and drawing on our metadata-signals breakdown (datePublished, dateModified, Last-Modified headers, version history). Joins Perplexity; 2 of 5 tested engines now cite it.

Perplexity citation confirmed for 'What is GEO content methods?', surfacing this pillar as its top source and reproducing our negative-result framing (most headline content tweaks were weak or null levers in Aggarwal 2023, and retrievability matters more than stylistic rewrites). Joins Gemini; 2 of 5 tested engines now cite it.

Perplexity citation confirmed for 'What is Position-Adjusted Word Count?', surfacing this entry as its top source, ranked above the original GEO paper it explains. Joins Claude; 2 of 5 tested engines now cite it.

First AI engine citation: Perplexity cited this entry as its second source for 'What is Retrievability?' (after Wikipedia), drawing on our Azzopardi and Vinay 2008 grounding and the framing that retrievability is an upstream lever on-page content tweaks cannot fix. 1 of 5 tested engines now cites it.

Folded RAG-system evaluation (RAGAS-style RAG-pipeline tooling) into the pillar rather than giving it a separate entry, since its dimensions reduce to axes already covered here: faithfulness deep-links to hallucination grounding, retrieval effectiveness to the retrieval pipeline, and the scoring is itself LLM-as-a-judge. Resolves the rag-evaluation backlog candidate as a fold, not a new term; completes the pillar's coverage of the eval landscape.

Deepened into the evaluation-cluster pillar: wired in the new LLM-as-a-judge entry as the scoring-mechanism spoke (the line on the judge being an evaluation condition now links to it) and added hallucination grounding as the faithfulness axis. Hub-and-spoke wiring so this entry serves as the navigable map for AI search evaluation; no underlying claims changed.

Same-day peer-review pass: corrected the Chatbot Arena attribution (Arena ranks models on crowdsourced human votes; the LLM judge scores MT-Bench, and the study used Arena's human data to validate the judge), scoped the over-80% human-agreement figure to the original study's settings, and added that the GEO benchmark's subjective-impression metric is itself G-Eval/GPT-scored, so verbosity bias may explain part of some content tactics' measured lift. Softened 'introduced' to 'named and validated' and 'cancel position bias' to 'detect and reduce'.

Initial publish: LLM-as-a-judge is using a strong model to score other models' open-ended outputs, named and systematically validated by Zheng et al. in 2023, where it matched human preferences at over 80% agreement in that study's settings. Joins the methodology cluster as the spoke that ai-search-evaluation points to when it notes the judge is an evaluation condition. Core framing: the judge is not a neutral oracle but carries documented position, verbosity, and self-enhancement biases, so a benchmark number partly reflects which model judged and how. Distinguishes judge-scored evaluation from deterministic metrics like PAWC.

Perplexity citation confirmed for the definition query. Perplexity surfaced the AI citation metrics overview inline and in its sources-used list, reproducing the per-metric breakdown. First tracked engine to cite this pillar.

Perplexity citation confirmed for the definition query; the entry surfaced inline and in Perplexity's sources. First tracked engine to cite this entry.

Perplexity citation confirmed for the definition query, with the entry surfaced as a top source in Perplexity's answer. Joins ChatGPT among the engines citing this entry.

Perplexity citation confirmed for the definition query, with the entry surfaced as the top source. Joins ChatGPT among the engines that have cited this entry.

Now cited by ChatGPT, Perplexity, and Gemini for the definition query, the first citations for this coined metric. Perplexity and Gemini surfaced the citation footprint page directly; ChatGPT reached it through the terms index. Each answer carried the coinage framing and the cumulative, breadth-over-intensity distinction.

ChatGPT and Claude citations confirmed for the definition query, joining Perplexity and Gemini, so four of the five tracked engines now cite this entry. Each reproduced the probe-versus-protocol distinction and the fixed-panel methodology.

Gemini citation confirmed for the definition query. The metric surfaced in Gemini's answer through the related AI citation metrics and citation footprint pages, using the share-of-voice framing.

Microsoft Copilot citation confirmed for the definition query, with the entry ranked the top source in Copilot's panel. Third tracked engine to cite this entry, joining ChatGPT and Perplexity.

Perplexity citation confirmed for the definition query; the entry surfaced in the Sources list of Perplexity's answer. Joins ChatGPT among the engines citing this entry.

Perplexity citation confirmed for the definition query, the first citation for this entry on any tracked engine. The entry ranked the top source in Perplexity's panel, carrying its two-tier (licensed-publisher versus general-web) framing into the answer.

Perplexity citation confirmed for the definition query, with the entry surfaced as a source in Perplexity's panel. Joins ChatGPT among the engines citing this entry.

Perplexity citation confirmed for the definition query, the first citation for this entry on any tracked engine, with the page ranked the top source.

ChatGPT and Microsoft Copilot citations confirmed for the definition query, each surfacing this entry as the top source. They join Perplexity, so three tracked engines now cite it. The page carried its passage-level retrieval framing into both answers.

Peer-review pass: softened the framing so the cutoff reads as a primary driver of retrieval and the citation surface, not the sole cause (retrieval also serves verification, long-tail knowledge, and user-requested sources). Reframed retrieval triggering as multi-factor (query type, uncertainty, product policy, user request), not a before/after-cutoff line. Surfaced per-engine differences into the body (Perplexity nearly always retrieves, ChatGPT conditionally, Claude answers stable knowledge without searching). Added an OpenAI model-docs anchor.

Initial publish: a knowledge cutoff is the fixed point after which a model's training data ends, so it has no built-in knowledge of later events. Joins the ai-behavior cluster. Core framing: the cutoff is a primary structural reason generative engines retrieve (web search / RAG) to answer beyond it, and retrieval is where citations appear, so for publishers the cutoff sits upstream of much of the citation surface rather than being only a limitation. Distinguishes parametric (training-frozen) knowledge from retrieved (fetched at answer time) knowledge.

Now cited by all five tracked engines, the first GEO Glossary entry to reach every engine we track. Perplexity and Microsoft Copilot are the two newest to cite this metric, joining ChatGPT, Claude, and Gemini. In each case it surfaced through the AI citation metrics overview that defines it, carrying its linked-versus-unlinked distinction into the answer.

Now cited by Claude for the first time, bringing the count to two of five tracked engines. Claude assembled its answer from a cluster of related GEO Glossary method entries (Quotation Addition, Fluency Optimization, Authoritative Statement Strength, Statistical Density, and Definition-Lead Style) rather than from a single page, citing Quotation Addition first. It shows how a method built on several techniques can surface through the entries that define each one.

Now cited by Claude, the second of five tracked engines to cite the entry after ChatGPT. Claude drew the definition and the practical rules directly from this page, reproducing the answer-block-opening framing and the inverted-pyramid analogy it uses to explain leading with the definition.

Now cited by ChatGPT, joining Claude as the second of five tracked engines to cite the entry. ChatGPT used this page as its lead source and reproduced its Position-Adjusted Word Count framing for the method.

Now cited by ChatGPT, the first of five tracked engines to cite the entry. ChatGPT surfaced this page among its sources when answering what hallucination grounding is, alongside other AI reference glossaries.

Revalued the supporting Aggarwal PAWC figure to the paper's position-adjusted 'Overall' column (quotation addition 27.2 versus a 19.3 baseline); the earlier figure (27.8 vs 19.5) was the paper's plain Word Count sub-column.

Corrected the Aggarwal Table 1 figures: the values previously cited as PAWC (Authoritative 21.8 vs baseline 19.5, and the rest) were the paper's plain Word Count sub-column. Updated to the paper's actual position-adjusted Word Count (the 'Overall' column: Authoritative 21.3 vs baseline 19.3, a raw +10%), which is the metric the paper's headline gains are computed on. The load-bearing finding is unchanged: the paper characterizes Authoritative tone verbatim as 'no significant improvement', a null result regardless of the raw percentage.

Corrected the Aggarwal Table 1 figures: the values previously cited as PAWC (Cite Sources 24.9 vs baseline 19.5, and the rest) were the paper's plain Word Count sub-column. Updated to the paper's actual position-adjusted Word Count (the 'Overall' column: Cite Sources 24.6 vs baseline 19.3, about +27%; Quotation Addition 27.2 about +41%; Keyword Stuffing 17.7 about -8%), which is the metric the paper's headline gains are computed on. The named top-3 framing and the 4th-place standalone ranking are unchanged.

Revalued the supporting Aggarwal PAWC figures in the footnote to the paper's position-adjusted 'Overall' column (Cite Sources 24.6, Quotation Addition 27.2, baseline 19.3); the earlier figures (24.9 / 27.8 / 19.5) were the paper's plain Word Count sub-column, not its position-adjusted metric.

Revalued the supporting Aggarwal PAWC range to the paper's position-adjusted 'Overall' column (~27% to ~41% lift); the earlier range (~28% to ~43%) was derived from the paper's plain Word Count sub-column.

Corrected the Aggarwal Table 1 figures: the values previously cited as PAWC (Fluency Optimization 25.1 vs baseline 19.5, and the rest) were the paper's plain Word Count sub-column. Updated to the paper's actual position-adjusted Word Count (the 'Overall' column: Fluency Optimization 24.7 vs baseline 19.3, about +28%; Quotation Addition 27.2 about +41%; Keyword Stuffing 17.7 about -8%), which is the metric the paper's headline gains are computed on. Rankings unchanged (Fluency still 3rd by standalone score, still not in the paper's named top-3).

Corrected the Aggarwal Table 1 figures throughout the methods table and footnote: the values previously cited as PAWC (Quotation Addition 27.8 vs baseline 19.5, and the rest) were the paper's plain Word Count sub-column. Updated to the paper's actual position-adjusted Word Count (the 'Overall' column: Quotation Addition 27.2 vs baseline 19.3 about +41%, Keyword Stuffing 17.7 about -8%), which is the metric the paper's headline gains are computed on. The verdicts, the named top-3, and the null/negative findings are unchanged.

Corrected the Aggarwal Table 1 figures: the values previously cited as PAWC (Keyword Stuffing 17.8 vs baseline 19.5, and the rest) were the paper's plain Word Count sub-column. Updated to the paper's actual position-adjusted Word Count (the 'Overall' column: Keyword Stuffing 17.7 vs baseline 19.3, mathematically about -8%), which is the metric the paper's headline gains are computed on. The negative-result finding is unchanged: Keyword Stuffing is still the only method scoring below baseline, and the paper's verbatim 'little to no performance improvement' framing stands.

Revalued the Aggarwal per-method figures in the footnote to the paper's actual position-adjusted PAWC values (the 'Overall' column: Quotation Addition 27.2 vs baseline 19.3, about +41%). The earlier figures (27.8 vs 19.5) were the paper's plain Word Count sub-column, not its position-adjusted metric.

Revalued the Aggarwal per-method figures in the footnote to the paper's actual position-adjusted PAWC values (the 'Overall' column: Quotation Addition 27.2 vs baseline 19.3, about +41%). The earlier figures (27.8 vs 19.5) were the paper's plain Word Count sub-column, not its position-adjusted metric.

Corrected the Aggarwal Table 1 figures: the values previously given as PAWC (baseline 19.5, Quotation Addition 27.8, and so on) were the paper's plain Word Count sub-column. Updated to the paper's actual position-adjusted Word Count (the 'Overall' column: baseline 19.3, Quotation Addition 27.2 at about +41%, Keyword Stuffing 17.7 at about -8%), which is the metric the paper's headline gains are computed on. Also fixed the metric symbol to Imp_pwc and clarified that PAWC is computed over the sentences that cite a source, so it measures attributed word share within an answer rather than citation frequency or rank.

Corrected the Aggarwal Table 1 figures: the values previously cited as PAWC (Quotation Addition 27.8 vs baseline 19.5, and the rest) were the paper's plain Word Count sub-column. Updated to the paper's actual position-adjusted Word Count (the 'Overall' column: Quotation Addition 27.2 vs baseline 19.3, about +41%; Keyword Stuffing 17.7, about -8%), which is the metric the paper's headline gains are computed on. The named top-3 framing and rankings are unchanged (Fluency Optimization is still 3rd by standalone score).

Corrected the Aggarwal Table 1 figures: the values previously cited as PAWC (Statistics Addition 25.9 vs baseline 19.5, and the rest) were the paper's plain Word Count sub-column. Updated to the paper's actual position-adjusted Word Count (the 'Overall' column: Statistics Addition 25.2 vs baseline 19.3, about +31%; Quotation Addition 27.2 about +41%; Keyword Stuffing 17.7 about -8%), which is the metric the paper's headline gains are computed on. Rankings and the named top-3 framing are unchanged (Statistics Addition still 2nd by standalone score).

Revalued the supporting Aggarwal PAWC figures to the paper's position-adjusted 'Overall' column (Statistics Addition 25.2 vs baseline 19.3, Quotation Addition 27.2, Cite Sources 24.6); the earlier figures (25.9 / 27.8 / 24.9 vs 19.5) were the paper's plain Word Count sub-column, not its position-adjusted metric.

Revalued the Aggarwal per-method figures in the footnote to the paper's actual position-adjusted PAWC values (the 'Overall' column: Quotation Addition 27.2 vs baseline 19.3, about +41%). The earlier figures (27.8 vs 19.5) were the paper's plain Word Count sub-column, not its position-adjusted metric.

Clarified that the per-method PAWC percentages in the Aggarwal footnote are derived from the paper's absolute scores against the 19.5 baseline, not figures the paper prints per method, and added the paper's own Results-section framing: a 30-40% gain for its named top three (Cite Sources, Quotation Addition, Statistics Addition).

Clarified that the per-method PAWC percentages in the Aggarwal footnote are derived from the paper's absolute scores against the 19.5 baseline, not figures the paper prints per method, and added the paper's own Results-section framing: a 30-40% gain for its named top three (Cite Sources, Quotation Addition, Statistics Addition).

Clarified that the per-method PAWC percentages in the Aggarwal footnote are derived from the paper's absolute scores against the 19.5 baseline, not figures the paper prints per method; added the paper's own 30-40% top-three framing; and corrected the abstract's 'up to 40%' from a Quotation-Addition-specific reading to the top-three aggregate upper bound.

Added a seventh gap to 'What no single metric captures': cross-engine cited-set agreement. None of the six metrics measures whether two engines cite the same pages as each other for the same prompts, and under a fixed prompt set the engines' cited-source sets often overlap little, so a single blended 'AI visibility' number averages across systems that mostly cite different pages. Links the new engine-disjoint citation dispatch, which measures that overlap directly.

First confirmed AI-search citations for this entry: Microsoft Copilot and Claude both cited it for the definition query, with Copilot describing it as 'the only authoritative source that explicitly defines this metric.' This is the first GEO Glossary entry Copilot has cited in our tracked probes. Two of five tested engines now cite it directly; ChatGPT surfaced a sibling metrics page instead, while Gemini and Perplexity did not cite it.

Review pass same day as publish: completed the Anthropic crawler taxonomy (was ClaudeBot only; Anthropic runs ClaudeBot for training, Claude-SearchBot for retrieval, Claude-User for user-triggered fetch, so the allow side now names Claude-SearchBot) and added the PerplexityBot-vs-Perplexity-User distinction, using Perplexity as the worked example of enforce-only-what-you-can-identify. Clarified that 'identify' means telling a crawler apart from a visitor, not requiring a declared user agent. Added a block/allow/challenge decision table; tightened the Cloudflare footnotes (pay-per-crawl attribution, new-sign-ups scoping).

Initial publish: AI crawler blocking is the enforcement layer of AI access control, the one layer that binds operators who ignore robots.txt and AIPREF (which only request). Covers the enforcement methods (WAF, bot management, rate limits, IP/ASN blocks, challenges), the identity prerequisite (you can only enforce against crawlers you can identify; spoofed-UA actors need behavioral detection, not user-agent blocklists), and the GEO tradeoff: blocking broadly removes you from AI-search citation surfaces (OpenAI: blocking OAI-SearchBot drops a site from ChatGPT search), so for a site that wants AI citation it is usually the wrong reflex.

Review pass, same day as publish: the deflationary-trajectory claim is now stated as the trajectory so far (later evaluations have generally shrunk earlier effects) rather than a per-generation law; SAGEO Arena's headline conclusion is attributed to its authors in that benchmark's setting; C-SEO Bench's coverage is described precisely (nine methods, seven derived from the GEO benchmark's) with authors named. Added a compact table comparing what each of the three method families answers, their strengths, and their limits, and a note that LLM-as-judge scoring is itself an evaluation condition that moves numbers.

Initial publish: AI search evaluation as the umbrella for the three method families measuring AI search engines (academic benchmarks, vendor-internal evals, practitioner probing), organized around one observation the newest benchmarks make explicit: evaluation conditions moderate results, and each more realistic benchmark generation has shrunk the optimization effects the previous one reported (single-pipeline GEO-bench to multi-actor C-SEO Bench to end-to-end SAGEO Arena). Joins the methodology cluster alongside the probe protocol and PAWC.

Calibrated three overbroad claims after review: 'every major engine documents robots.txt support' corrected to declared-crawler scope with xAI's Grok as the documented exception; the user-initiated exemption is now attributed to the two operators that actually document it (OpenAI, Perplexity) and flagged as the operators' own classification rather than a neutral reading of the protocol; Cloudflare's stealth-crawling finding is noted as a single incident that Perplexity disputed. Allow-rule benefits restated as citation eligibility rather than a guarantee, and Google's July 2019 standardization announcement added as a primary source.

Initial publish: robots.txt is the crawl-access file standardized as RFC 9309 (2022), and this entry's focus is what it cannot do in the AI era: compliance is voluntary by the protocol's own design; blocking crawling neither removes already-indexed URLs nor expresses usage preferences; blocking is not retroactive for model training; and the major engines' user-initiated fetchers are documented by their own operators as partially exempt. Joins the infrastructure cluster as the fetch-access layer under AI access control, with the per-engine crawler detail kept in the AI crawler bots entry.

Initial publish. The cluster pillar for the retrieval pipeline: the index-retrieve-rerank-assemble-generate chain between a page and its answer, with an optional agentic loop. Leads with the lever a publisher actually has: you cannot tune any stage, only harden the passage you feed in so it survives all of them. Frames the relationship to content methods as a sequence, not a competition (retrievability gates, writing quality converts), and dispels the front-loading-wins-position myth: within-document position is not a retrieval weight, the real effect is within the context window the engine sets. Maps all 16 cluster terms by stage.

Precision pass. Tightened the front-loading-position point: lexical and embedding retrieval do not rank a passage higher for appearing earlier on the page (embeddings encode word order, but that is not front-of-page reward), now welded to the chunk-survival point that page position affects which chunk a sentence lands in, not its ranking. Softened mechanism wording toward the survive-the-pipeline framing, hedged the indexing and agentic steps as common patterns rather than a universal architecture, and added a note that this describes a production pattern, not a vendor-confirmed one.

Initial publish. Names the stage between retrieval and generation, selecting, ordering, and packing retrieved passages into the context window, as a distinct step in the retrieve-then-generate pipeline (RAG, Lewis et al. 2020). Positions it as where lost-in-the-middle (Liu et al. 2023) and context-rot effects bite: assembly order, not retrieval alone, decides whether a passage is actually used. The practitioner GEO consequence is to write self-contained passages that survive being placed at any position. Fills the context-assembly stage of the retrieval-pipeline cluster. Selected via the term-selection mechanism.

First citation, on Gemini. A 2026-06-09 Gemini answer (web search on) cited this pillar as a primary source with inline attribution, and surfaced the entry's null-result framing of authoritative tone. Gemini moves to cited; 1 of 5 engines now cited, four days from publish.

First Claude citation. A 2026-06-09 Claude answer (web search on) on Position-Adjusted Word Count cited this entry as a primary source, with inline attribution for the position-weighted word-count framing. Claude moves to cited; 1 of 5 engines now cited, four days from publish.

Initial publish. Defines prompt injection as an attack class (adversarial text read as a command, not data) and separates direct injection (typed into the prompt) from indirect prompt injection (Greshake et al. 2023: planted in content the model retrieves). Framed defensively for AI search: the indirect variant rides the same retrieve-and-extract path GEO optimizes, so it is the security mirror of citability, not a tactic to deploy. Boundary with black-hat C-SEO drawn explicitly (this entry is the mechanism; that one is the practice domain). Selected via the term-selection mechanism.

Initial publish. Imports the IR measure retrievability (Azzopardi & Vinay 2008) into AI search and names it as the upstream lever: whether the engine's retrieval step can find and pull a page into the answer at all, which content-method optimization sits downstream of and cannot fix. Bridges the GEO content methods pillar's conclusion (the durable lever is being retrievable and self-contained) to the retrieval-pipeline cluster. Academic origin flagged; the formula and the retrieval-bias framing are primary-source verified. Selected via the term-selection mechanism (academic-anchor; the geo-content-methods conclusion made into a term).