Grokipedia has not issued a single edit decision since April 25, and no article text on the site has changed since April 24, according to a Lawfare analysis published August 5 that PPC Land covered the following day. The queue holds 13,002 human-submitted corrections that nothing appears to be reviewing, while the site’s contents keep surfacing inside AI-generated answers elsewhere. If your job involves watching how a brand gets described inside AI answers, the exposure is plain enough: a reference the engines still cite has lost the ability to be corrected.
The analysis comes from two researchers: Ronald Robertson, a Stanford Cyber Policy Center research scientist, and Renée DiResta of Georgetown’s McCourt School of Public Policy, an associate research professor there who also edits for Lawfare. They built their sample using Grokipedia’s own search-suggestion feature, pulling 301,035 pages, 34,519 of which held at least one proposed correction, adding up to 225,496 requested changes. Two pools of stalled work resulted: 13,002 requests sit in review, and another 8,658 were approved but never written into the articles.
The pattern is not explained by topic sensitivity. Twelve edits filed against the ChatGPT page on April 24 were approved that same day, yet nothing submitted to that page in the three months since has cleared review. The authors also tested an unambiguous case. They submitted a correction noting SpaceX’s initial public offering, filed shortly after the June 12 IPO. The fact was well sourced and undisputed. It never appeared on the page.
The likelier explanation sits inside the system, not at its front door. Two accounts that submitted edits automatically, called grok and Grok Editor, generated 57.8 percent of all requests in the sample, and both stopped submitting between March and mid-April. Human contributions never stopped, running at roughly 216 corrections a week after April, arriving into a review process that produces no response. That timing points to an internal change rather than a capacity failure: the automated submission loop stopped before the human-facing queue visibly stalled.
A separate problem compounds the freeze. A mass rewrite around March 14 broke the links between suggested edits and the text they referenced, and Grokipedia’s logging system began retroactively marking earlier approved edits as rejected. The actress Prunella Scales offers one example: the Tow Center for Digital Journalism’s archive shows 10 of her page’s launch-week edits marked approved, but the live site now flags nine of them as rejected, even though the corrected text is still visible in the article.
The freeze registers commercially because Grokipedia is not a niche property. Similarweb estimated 6.7 million visits to the site in June 2026. Ahrefs, measuring separately in March 2026, counted Grokipedia’s citations across AI systems at roughly 356,000, concentrated in ChatGPT and Google’s AI Mode, its conversational search experience. That count is nearly five months old, and the Lawfare piece does not say whether it has grown, shrunk, or held steady since the freeze began. Wikipedia draws an estimated 25 million such citations by the same measure, so Grokipedia’s share is small. It is not zero.
That gap matters most for whoever’s entity page is wrong. Individuals and companies with inaccurate biographies on the site can submit a correction through Grokipedia’s interface, but the Lawfare authors find no response on the other end. Glenn Gabe, who consults on SEO and AI search at G-Squared Interactive, posted on August 6 that “Grokipedia just tanked again in Google,” pointing to the site’s lack of updates since April 24. The post had drawn 831 views at the time this reporting was captured.
A measurement framework the IAB released on August 4 happens to name exactly this failure mode. It sorts what an AI system gets wrong into two buckets: the model inventing something, and the model faithfully repeating a mistake it inherited. A page that has sat untouched since April falls squarely in the second bucket. Nothing is malfunctioning inside the model. The error is sitting in the source it reads, and the route that would normally remove it has stopped running.
Search teams practicing GEO, generative engine optimization for LLM-driven answers, should treat Grokipedia as a citation source they cannot influence through its own correction channel. The next ninety days call for monitoring entity pages there the way teams monitor an unresponsive vendor: flag errors publicly, because the interface will not resolve them.
PPC Land reported this story on August 6, 2026, based on Lawfare’s August 5, 2026 analysis by Renée DiResta and Ronald Robertson.
