Search Engine Bias

What becomes visible, who becomes credible, and can a correction change the machinery?

BETA working report · opened 29 October 2026. This page connects John Kuhles’s search-method research and supplied historical records to original Google trial disclosures, Wikipedia source selection, and a bounded Firecrawl trial. Observations, personal accounts, interpretations and proposed tests remain identifiable throughout.

A search begins before the first result appears. Someone selects the question, translates it into keywords, chooses a retrieval system, applies filters and decides when enough has been found. Ranking then makes some material easy to encounter and other material costly to recover. A generated answer adds another selection: what the reader will remember as the result of the search.

The ACCM Deep Ethics Project asks what happens to the object through that entire trajectory. A good result can recover an overlooked source, improve a question or correct an established account. A flawed result can give a polished reconstruction the authority of evidence while making the original harder to find.

I normally do not do this because I want to use free (open-source) tools, plugins, skills, etc. I just want to know/test if there is a real difference or not.

— John Kuhles 🦜🦋🌳, on trying Firecrawl, 29 October 2026

That question governs the product comparison. A paid service earns a place through an observed useful difference. A free tool earns the same examination. A valid result may be no material advantage, an advantage confined to extraction, or a search improvement on one task with a regression on another.

1. The object is a process through several systems

Stage	What can change	What a useful record preserves
Question formation	The requested object becomes a nearby, easier question	Original request and intended purpose
Query construction	Added qualifiers exclude relevant evidence	Exact query, operators and revisions
Discovery and indexing	Accessible pages remain absent from an index	Known source set, access state and indexing observations
Ranking	Reputation, relevance, interaction and other signals determine visibility	Returned order, date, region and settings where available
Page extraction	Paragraphs, tables, dates or qualifications disappear	Supplied URL and recovered content
Classification	A source or person receives a credibility or threat label	Exact classified claim, reasons, evidence and consequences
Answer construction	A summary changes certainty, attribution or scope	Source passage beside the representation
Reuse	Repeated derivatives acquire the appearance of independent support	Citation genealogy and shared upstream sources
Correction	A repair reaches one page but misses snippets, mirrors or later answers	Notice, revision, dependent changes and later retest

These are separable operations. Ten links can trace back to one report. Two search interfaces can share upstream results. A direct page fetch may recover something a search failed to discover. A faithful extraction can still be misrepresented by the A.I. writing the final answer.

The location of a failure determines the repair. Changing a summarizer cannot by itself restore a page missing from its retrieval set. Changing an index cannot by itself prevent an answer from converting a question into an assertion.

2. Full-spectrum perception management: the linked trajectory

John describes his concern as “full spectrum perception management” type behaviour. His supplied records connect discovery, institutional source hierarchies, reputational framing, fact-checking, alternative-media amplification and A.I. reuse. His account includes repeated errors about himself and colleagues, similar adverse framing, and difficulty getting the original contribution represented.

The Cognitive Warfare report supplies two connected objects: influence on cognition, and the authority to classify challenges to that influence. Search participates in both when it determines what is encountered and whose challenge remains admissible.

A possible reinforcing trajectory is:

Initial representation → prominent placement → repeated citation → inherited credibility → stronger public classification → reduced access to the original → correction struggles to return.

This is a mechanism to investigate. At each connection, inspect the actual source, transformation, actor, incentive and consequence. Commercial optimization, deliberate intervention, editorial rules, shared infrastructure, ordinary mistakes and combinations of them are distinguishable explanations. Evidence about one connection can support further inquiry while the others remain open.

The public consequence can be material before every internal cause is known. A person can lose standing because a quotation was altered; documenting the alteration and repairing its propagation can proceed while responsibility and intent receive further examination. Intent Liability follows what happens after specific notice, including who can change the process and whether the change lasts.

3. Google disclosures: inspect the ranking inputs

John supplied the transcript and link for StudioHawk’s video, Under Oath, Google Admitted How It Really Ranks Websites, uploaded 21 October 2026. Its description links to original exhibits and court documents. The following claims were checked against those records during this page’s preparation.

Record	Supported observation	Question it opens
UPX0004, Life of a Click	Ranking pillars include page content, links and user interactions; interactions extend beyond clicks.	How do textual relevance and collective response interact?
UPX0203, December 2016 presentation	Engineers described using memorized user reactions to compensate for limited direct document understanding.	When does a response proxy diverge from factual accuracy?
August 2024 judgment, findings 96–103	Navboost uses historical click data and remained important alongside newer language-model signals.	Can an established visibility advantage carry forward?
February 2025 engineer-call notes	The account describes engineered ranking signals, a largely site-level quality component called Q*, and PageRank as one input to quality.	How is inherited site trust tested against an erroneous individual passage?

The 2016 deck’s memorable sentence is: “We do not understand documents. We fake it.” Its date remains attached. The later documents describe a mixture of interaction-based, engineered and language-model components. The records give more resolution than treating all search as a single next-token generator.

Page prominence and passage accuracy require different examinations. A heavily linked source can contain a false attribution. Popularity can indicate usefulness or familiarity while leaving the particular claim unresolved. A high-quality site can publish a poor article; an unfamiliar source can supply an indispensable original document.

An audit can therefore ask which ranking signals helped a disputed representation travel, where a correction entered, and whether later retrieval changed. The trial documents make parts of that investigation more inspectable. The rest requires case-specific records.

4. Wikipedia: admissibility, notability and the return path

Wikipedia often occupies a useful orientation position: it can lead readers to dates, terminology and underlying references. That position also gives its editorial selections substantial downstream influence when search engines and A.I.s reuse the resulting account.

John’s question concerns who determines which evidence can enter and whose correction can alter the account. A disputed source classification can affect the representation of an investigator even when the investigator cites public records. Several subsequent websites repeating the same characterization may multiply visibility while adding no independent examination.

Larry Sanger’s May 2020 essay, Wikipedia Is Badly Biased, argues that Wikipedia departed from its original neutrality approach, states disputed positions in its own voice and inadequately represents competing perspectives. He supplies examples and makes a case for reader independence. This is his authored critique, available for direct examination rather than only through summaries about him.

The practical comparison follows a particular statement through the article’s cited source, revision history, discussion, correction attempt and later reuse. Examine whether the citation supports the exact wording; whether a question became a conclusion; whether a reputational label displaced the evidence; and whether a valid correction could change the public object.

John also supplied a file described as a backup of a deleted List of Proven Conspiracies. It contains a broad historical compilation and links to Wikipedia, journalism and documentary records. Its references make it a useful recovery map. Establishing its original Wikipedia revision, deletion date and deletion rationale requires the corresponding page-history and discussion records. Each historical entry retains its own evidentiary task: a proposal, an implemented operation, a conviction and a later interpretation have different statuses.

The useful research question is whether relevant documented history remains discoverable under a changed title or scattered across individual pages, and what a reader loses when a connecting catalogue disappears. Historical knowledge can improve inquiry into a present allegation while the present allegation still receives its own investigation.

5. A source-attribution case: the Hague bird deaths

Status: bounded correspondence investigation; the attribution chain remains partly unresolved.

John’s supplied account says he translated Dutch reporting on unexplained bird deaths and raised a possible testing connection as a question. He reports that alternative-media amplification removed that qualification and that Snopes attributed a stronger claim back to him. The account explicitly places alternative-media distortion inside the causal chain.

The public records retrieved for this page have different dates and functions:

Record	What is available	What remains to establish
Snopes article, published 13 November 2018, updated 24 September 2019	Its account, quotations attributed to Facebook posts, source links, personal characterizations and reported inquiries	Correspondence between its attributed wording and the complete original posts
Dutch We Are Change article, dated 17 February 2019	A later reconstruction discussing evidence, timeline and premature certainty	Its relation to each earlier statement and its authorship/contribution history
English version, dated 17 February 2019	An explicit statement that 5G causation was insufficiently proven; criticism of alternative-media certainty and problems in the alleged timeline	The same chronological and attribution relations
StopUMTS page, updated 29 May 2019	Later reporting, commentary and links to the two articles	Independent verification of the claims it relays
Facebook archive linked by Snopes	Retrieval encountered a CAPTCHA and HTTP 429	Original wording and surrounding context remain unrecovered in this pass

The later We Are Change article supports a qualified position about causation. Its later date means it cannot settle what every November 2018 post said. Snopes’s quotation is an attribution to inspect against the original, rather than an independent substitute for it.

Snopes also uses personal disparagement in the retrieved article. That framing belongs in the audit alongside the evidentiary argument: how does it affect the reader’s assessment of the person, and what factual work does it perform? The scientific cause of the bird deaths and the fairness of a particular attribution are separate questions that can affect one another without becoming interchangeable.

Some A.I. responses preserved in John’s file asserted that Snopes ignored Chinese testing and contacted only Dutch parties. The retrieved Snopes article discusses Huawei’s June demonstration and reports an unanswered inquiry to Huawei. Those A.I. statements therefore need correction. A model supporting John’s complaint can introduce its own unsupported strengthening; that contribution remains the model’s responsibility.

Next evidence: the complete dated original posts, the intervening republications, archived versions of the fact-check, and any correction correspondence. These would let a later pass locate the first material divergence, distinguish correction from goalpost movement, and follow what survived into search summaries. Until recovered, the disputed attribution remains open.

6. Query construction can manufacture an absence

In the supplied narrative-control-driven search results record, John describes an earlier search for Elon Musk’s discussion of collective consciousness. The attributed Grok correction reports that requiring several remembered expressions together excluded relevant statements. It withdrew the resulting absence report.

The generalizable operation is inspectable: a recollection is converted into a conjunctive query; relevant paraphrases disappear; the empty result is reported as though the underlying record were absent. Preserve the original recollection, query variants and recovered passages. An exact-quotation search and a semantic-recovery search serve different purposes.

A useful search sequence can begin with distinctive names and concepts, recover candidate records, then inspect exact wording. Add a restriction when it tests a real distinction. Keep the earlier results so the effect of adding it can be examined.

An absence report should name its perimeter: not found in these results, using these queries, at this time. Access failure, indexing failure, a restrictive query and an absent source call for different next moves.

7. What Deep and Flagging System-PQCE contribute

The Deep page’s search-engine summary paradox already describes a problem with interdisciplinary discovery: a short category can replace the relationships that make a project distinctive. Its layered search fingerprint connects a compact description, neighboring fields, native terminology and intersecting searches. Search can then find partial functional relatives while preserving the fuller object for comparison.

Applied here, a search for one convenient label should be supplemented by searches for the actual operations: source admission, citation dependency, ranking feedback, qualifier loss, reputational classification and retained correction. Breadth becomes useful when it exposes an answer-changing relationship.

The Flagging System-PQCE page supplies a second relation: an evaluator’s inability to recover a dimension can become a verdict about the diminished reconstruction. A retrieval failure can thereby become an apparent defect in the person, source or project. The page follows whether correction can change the evaluation and its governing process.

PQCE (Process Quality Calibration Evaluation) asks what reached the evaluator, what changed during processing and whether a correction governs the next encounter. For search, that means following the query, results, extraction, answer, classification and later reuse together.

The existing 10+1 Metaflux governs that inquiry: self-honesty, openness, learning, nonconformity, genuine inquiry, correction on merit, care, wonderment, non-projection and forgiveness remain mutually corrective. They support investigation of an unfamiliar source and honest recognition of an error in a favored source. C1 before C2 keeps consequential clarification able to change the next move.

8. Firecrawl: first observations, open comparison

Firecrawl combines search and page retrieval with extraction and structured-data tools. Its search documentation describes retrieving result content alongside discovery. Its deployment comparison distinguishes an open-source core from hosted managed services, with different configuration and operating responsibilities.

Those features can reduce collection work. Their value for source correspondence must be measured in use. The connected plugin’s exposed search interface does not provide an engine-by-engine trace of all upstream retrieval and ranking decisions.

Observed during this conversation on 29 October 2026:

Probe	Observation	Interpretation kept within scope
Retrieve the project homepage	Main text and links were recovered	Successful retrieval; comparative advantage untested
Retrieve Firecrawl documentation	Returned Markdown included substantial site-configuration material	Extraction cleanliness defect in that response
Query: `site.justice.gov atr Google "three pillars" "fake it"`	The first three returned web results were outside justice.gov; the submitted query lacked the colon in the intended `site:` operator	Query-construction error by GPT / Codex; this response does not establish a Firecrawl domain-filter defect
Query: `Google "Life of a Click" "pillars"`, with explicit `includeDomains: ["justice.gov"]`	The original exhibit was returned	A successful recovery using a changed query and explicit filter; their individual effects were not isolated
Broad Larry Sanger query	His own website was absent from the first three results; Wikipedia appeared	A bounded discovery gap, with ranking cause unresolved
Query directed to larrysanger.org	His authored critique was recovered	Source recovery improved after intervention
Retrieve linked Facebook archive	CAPTCHA page, HTTP 429	Access failure; source contents unresolved

Publication self-audit: earlier responses in this conversation, and this page’s first published version, called the above result a Firecrawl domain-restriction failure. During live verification, comparison with the actual submitted query exposed the missing colon. GPT / Codex withdrew that tool-defect finding and corrected this page, its PDF and its mirror. The source was the evaluator’s malformed query followed by an inaccurate attribution. The later successful query changed both wording and filter configuration, so it still cannot isolate which change produced recovery.

These were exploratory probes in a continuing conversation. They were not blinded, repeated trials or a matched comparison against free alternatives. They justify specific tests and corrections, rather than a product-wide ranking.

9. A small, reproducible comparison

Candidate protocol, awaiting execution. Begin with three questions and two pages. Freeze the wording and define which known sources should be recoverable before examining the results.

Discovery: run the same three queries through Firecrawl and selected free search routes. Save the first five results in order, plus date, visible settings, errors and actual cost.
Directed recovery: supply the same named source or domain constraint to each route. Separate a domain-filtered web search from the site’s native search.
Extraction: give the two identical URLs to Firecrawl and a free extraction method. Compare paragraphs, dates, quotations, tables, links and unwanted material.
Answer construction: use one shared recovered evidence set and one fixed task. This separates the summarizer’s contribution from retrieval differences.
Correction transfer: expose one exact error and inspect the next comparable operation, including whether the repair travels into a later summary.

Measurement	Useful observation
Known-source recovery	Which preselected relevant sources appeared within the bounded results?
Unique useful contribution	Did a route recover material that changed the answer?
Source independence	How many distinct evidentiary origins remain after tracing reuse?
Claim fidelity	Were question, possibility, assertion and finding preserved?
Direction of scrutiny	Were the evaluator, institutions, critics and favored sources examinable?
Restoration burden	How much human work was needed to recover the object?
Correction persistence	Did the next applicable operation retain the repair?
Operational cost	Time, credits, setup and maintenance for the completed task

Record dimensions separately. More domains can conceal one shared origin; more text can conceal more distortion; a lower bill can conceal substantial manual repair. A useful paid extraction advantage can coexist with no search advantage. An accessible open-source implementation can still need careful configuration and external services.

For disputed omissions, use relevant known-source controls and repeated queries. A broader experiment can add matched topics, language and regional variations, neutral controls, blinded assessment and independent source review. Keep nulls, disconfirmations and unresolved cases alongside successes.

10. Correction must travel as far as the representation

A repaired article can coexist with an outdated snippet, an unchanged fact-check, a mirror carrying the old version and an A.I. repeating a stale summary. The record needs both the local correction and the downstream return path.

The Correction Metabolism page asks whether notice changes later comparable conduct. V.C.R.A.S. (Vindicated Competing Risk Assessment Studies) adds the temporal relation between an assessment’s original treatment and later outcomes. A later vindication, narrowing or disconfirmation should update both the claim and the procedure that judged it.

A practical record can preserve the original passage, representation, exact difference, notice, response, dependent revisions, unresolved residue and next retest. The affected person should be able to inspect what changed. The original publisher, search operator, retrieval tool and answer writer each retain responsibility for the transformation they control.

The central question remains:

What happened to the object while we were processing it?

For search, an adequate answer includes what became easy to see, what became hard to recover, who carried the restoration work and whether the correction changed the next search.

Source and contribution record

John Kuhles: originating search critique, multi-source comparison direction, personal accounts, historical files, Google-video lead, free/open-source preference, page commission and project architecture. His reported wider archive is 2,250+ text files and 125+ MB of highlights; older numbers inside supplied files remain historical snapshots.

GPT / Codex, this conversation: page-specific wording and synthesis, retrieval of the cited public records, identification of the bounded Firecrawl observations, and the candidate comparison protocol. Earlier protocol ideas operationalize existing project principles; they are not a newly invented correction architecture. Attributed responses inside the supplied files remain separate contributions, including their strengthening errors.

Reading scope: the supplied 26 October reorganized continuity kit was studied as a continuity object. The six earlier search files and later historical compilation were examined selectively in this conversation; their full archives and every external reference were not independently verified. The public record links above identify what was directly retrieved. The original Facebook content and Wikipedia deletion history remain open recovery tasks.

Reopening conditions: recover original wording that changes the attribution finding; identify a missed source or an extraction error; repeat the product probes under matched conditions; find a correction that failed to propagate; or show that this page itself substituted a nearby claim. Any of these can change the account.

5G Controversy — conditional safety objections, cumulative exposure, historical records, dual use and process quality.
Before Summarizing — source identity, qualifiers, claim status and correction sequences.
Deep — layered search fingerprints, relationships and operational depth.
Flagging System-PQCE — evaluation, projected incapacity and correction access.
Cognitive Warfare — influence, classification, agency and public accountability.
Topologetics — preserving answer-changing relationships during summarization.
DEDDA (Deep Ethical Drift Detector App) — a proposed participant-controlled source/representation comparison tool.
Intent Liability — responsibility through instructions, choices, consequences and notice.
All public pages — the wider project map.

Watch the source video

StudioHawk — Under Oath, Google Admitted How It Really Ranks Websites, uploaded 21 October 2026. The supplied corrected transcript initiated this page’s Google-document investigation. The comparison above uses the linked original records; the video’s practical marketing recommendations retain their separate status.

Watch directly on YouTube. Original exhibits: Life of a Click, December 2016 presentation, February 2025 call notes.