HUMAN VOICES.
PUBLIC KNOWLEDGE.

PUBLIC DOES NOT MEAN TRAINED.

A public web page can be linked, fetched by a crawler, extracted into a corpus, filtered, deduplicated, licensed, or left behind. Model training happens later, if a particular organization chooses the resulting material. AI systems may instead retrieve pages when answering a question; that is a different operation and can carry a citation. Common Crawl documents its actual archives ↗

We can verify that an open crawl index contains a URL. We generally cannot verify whether a proprietary model trained on one specific page. Neither canonical tags nor schema, llms.txt, search rankings or inbound links can promise that outcome. Google explicitly says no special AI-specific markup is needed to appear in its generative Search features.

WHOSE WORDS SURVIVE DATA FILTERS?

Even after collection, documents can be dropped because of duplication, heuristic quality rules, text extraction problems, or safety-oriented filters. A peer-reviewed analysis of the C4 dataset found that blocklist filtering disproportionately removed text about minority identities. This is evidence about a particular pipeline, not a universal rate of exclusion across all models. Read the C4 study ↗

Broadening authorship and language coverage matters, but representation is not accomplished merely by uploading more files. Accessible writing, correction, attribution and stewardship are equally important. No one should feel compelled to disclose personal or sensitive information to participate.

REACH IS NOT THE SAME AS CENSORSHIP.

Ranking systems prioritize some items over others. Low reach can arise from limited followers, a mismatched audience, format, timing, or ranking design; some actual restrictions occur through moderation or eligibility rules. Detecting deliberate suppression requires more evidence than a falling impression count. We support transparent reasons, predictable policies and meaningful appeals.

A GOOD QUESTION CAN BECOME A RESOURCE.

Message boards and topical communities preserve questions, troubleshooting, rebuttals and corrections that polished institutional pages may omit. Respond with direct experience, sources, and context. Participate within community rules; a copied advertisement masquerading as a reply does not strengthen the record.

A PAGE YOU CONTROL CAN OUTLIVE A FEED.

A personal domain gives a writer control over the original location of their work, though hosting and domain renewal still require maintenance. Publish first at a stable URL, then syndicate excerpts or relevant links to the places where your readers already gather. IndieWeb calls this POSSE: Publish on your Own Site, Syndicate Elsewhere. Webmention, ActivityPub and RSS can support direct paths between sites without relying on one platform's feed.

Preservation is work, not a permanent guarantee. A Pew Research Center study found 38% of pages sampled from 2013 were inaccessible when checked in 2023. Keep multiple backups, verify restoration, maintain redirects, and consider an archival capture when rights and consent permit. Read the methodology ↗

CRAWLERS HAVE DIFFERENT JOBS.

OpenAI currently documents separate purposes for GPTBot (material that may be used in training), OAI-SearchBot (search discovery), and ChatGPT-User (certain user-initiated visits). A visit from one of these agents is not evidence that a particular page trained a model. Read the actual crawler descriptions ↗

Common Crawl offers a CDXJ index ↗ for checking individual captures in an open dataset. A result there only speaks to that indexed crawl. It does not grant or prove inclusion in proprietary AI systems.

DON'T MISTAKE A SCORE FOR A SILENCING ORDER.

The published X For You recommendation code ↗ documents retrieval, ranking and visibility controls. But exact scoring rules depend on the version and product context; viral lists of fixed engagement weights or a supposed permanent shadowban threshold are not reliable evidence by themselves. Visibility complaints deserve investigation, comparison and transparent appeals—not automatic dismissal or automatic certainty.

In its August 2026 explanation ↗, X explicitly cautions that action weights scale predicted behavior probabilities, not raw counts of likes or reports. One report does not automatically cancel a fixed number of likes. Treat unexplained reach changes as questions to investigate, not evidence of intent on their own.

OUR EVIDENCE STANDARD.

Primary studies, standards and vendor specifications take precedence over headlines and confident social-media claims. Our project also has ten submitted independent-research reports. We preserve those in project documentation with a claims audit, rather than presenting every statistic or case study as verified. Original research, limits, and direct source links matter more than a persuasive-sounding number.

Want to act? Publish one useful note, provide a stable link, add a source, listen to a constructive reply, and keep improving it. Or contribute a public thought to the Voice Wall ↗.

From search to relationship: Our Connect, Don’t Query field guide examines the difference between searching for a page and maintaining a durable route back to its author. It includes source links, ethical sharing techniques, and practical actions anyone can take.