Bing Deindexes 90K URLs: AI visibility spam lessons

Bing’s algorithmic deindexing of a spam site with 90,000+ ChatGPT citations is the clearest enforcement signal AI-search practitioners have seen in 2026. The site — a YMYL domain no source has named — was reportedly surging in ChatGPT answers while remaining nearly invisible in Google, and within roughly two days of a public flag, its entire Bing index footprint went to zero.

The episode matters well beyond one anonymous domain. When SEO consultant Glenn Gabe asked Bing why the action looked so different from Google’s manual-action playbook, Microsoft Bing Principal Product Manager Krishna Madhavan gave an answer that doubles as a policy statement: “we do not do one offs....not scalable....” In other words, Bing’s stated approach is to build a spam algorithm for the pattern and apply it to every site exhibiting that pattern — not to hand-penalize the one domain that got attention on X.

This guide assembles the full dated arc — from Gabe’s June 22 case study to the July 10 deindexing to the July 13 policy write-up — disentangles the two different “90K” figures that most coverage blurs together, corrects a likely misattribution (the statements came from Madhavan, not the recently retired Fabrice Canel), and translates “algorithmic at scale” into a practical resilience playbook for anyone whose AI-search visibility depends on a search index they don’t control.

The story starts on June 22, 2026, when Glenn Gabe of G-Squared Interactive published a case study titled “Surging in ChatGPT, Dead in Google”. The subject: a site publishing on YMYL (Your-Money-Your-Life) topics, filled with AI-generated content, with no author bylines, no company information, and no contact details. Gabe deliberately did not name the site, and neither has any subsequent coverage — it remains anonymized in every source.

What made the case remarkable was the asymmetry. Per Semrush’s Brand Radar tool, the site was being cited in over 90,000 ChatGPT prompts as of Gabe’s June 22 check. Yet in Google, it had barely any organic search visibility and no meaningful presence in AI Overviews or AI Mode — despite Google having indexed 45,000+ of its URLs. Bing, meanwhile, had indexed 88,000+ URLs from the site and was ranking it well.

Gabe’s read of the content profile was blunt: this looked like scaled content abuse — the same AI-generated, thin-authorship pattern that Google’s March 2026 spam update targeted when it decimated AI-generated page networks. He also flagged his own uncertainty about whether Bing would act, writing that he would hope Bing’s systems would eventually understand what was going on and drop the site over time.

How does a site with almost no Google visibility rack up 90,000+ ChatGPT citations? The answer is grounding. Gabe’s case study points to OpenAI’s own help documentation confirming Bing is an official search partner used for ChatGPT search grounding — when ChatGPT reaches out to the live web, Bing’s index is a primary source of candidate results. The site’s Copilot citations were also very strong, correlating directly with its Bing organic rankings. When Gabe checked Brave Search and DuckDuckGo as other possible grounding sources, results were mixed or weaker — leaving Bing as the dominant driver of the site’s AI-search footprint.

This is the second time Gabe has documented what he calls a “Mt. AI” event — his term, coined in an earlier case study, for sites whose search-engine demotions or promotions ripple downstream into AI-platform citation swings. The mechanics mirror what our AI-search citation ranking factors study found across platforms: assistants ground on search indexes, so search-index standing is upstream of AI citations — for better and, in this case, for worse.

The strategic lesson cuts both ways. On the one hand, the case proves that strong rankings in a single engine — even the second-place engine — can drive enormous AI-assistant visibility, which is the core premise of generative engine optimization. On the other, it shows exactly how concentrated that dependency is: one index decision, and 90,000+ prompt citations lose their grounding source overnight.

No single source lays out the complete arc — Search Engine Roundtable’s coverage starts mid-story at the deindexing, and Gabe’s case study predates the action by weeks. The table below assembles every dated event from the primary sources into one timeline.

Three gaps in that timeline are worth computing (these are our own derivations from the dated sources, not figures any single source states): roughly 18 days from Gabe’s original case study to the deindexing (June 22 → July 10), roughly 2 days from the public X flag to the deindexing (July 8 → July 10), and roughly 9 days between Fabrice Canel’s retirement and the action (July 1 → July 10). The middle number is the one that should recalibrate expectations: once the right person at Bing saw the pattern, removal was nearly immediate — and Gabe’s July 10 report described the result in plain terms: “The site has been completely deindexed. There were close to 90K urls indexed as of yesterday. Now gone. Poof.”

Lily Ray’s same-day reaction captured the community’s surprise that Bing acted at all: “Well there goes the article I was about to write. Looks like Bing is indeed paying attention to these spammy sites after all.”

The most consequential part of this story is not the deindexing — it’s the explanation. Asked about the action on X, Krishna Madhavan, Principal Product Manager at Microsoft Bing, stated that Bing does not perform one-off search or indexing penalties because they are not scalable. Per Barry Schwartz’s write-up of Madhavan’s statements, Bing’s method is to build a spam algorithm that detects the specific spam pattern, then apply it across every site exhibiting that pattern simultaneously — not just the one site that got flagged.

Two hedges belong on the record here. First, these are statements from a Microsoft employee’s personal X account, reported by Search Engine Roundtable — not an official Microsoft or Bing blog post. Second, Microsoft has never publicly stated why this site was deindexed. The causal link to gaming ChatGPT visibility is the SEO community’s own reading — connecting Gabe’s June 22 case study to the July 10 action — which Madhavan did not publicly dispute, and implicitly reinforced by confirming Bing would look into the flagged site and engaging after the action.

The exchange also carried a personnel subtext. Nine days before the deindexing, Fabrice Canel retired from Microsoft after nearly 30 years — the engineer-PM who led Bing’s indexing, crawling, and URL discovery, and who created the IndexNow protocol and the Bing Webmaster Tools platform we covered in our Bing Webmaster Tools AI Citation Share guide. Replying to Lily Ray on July 10, Madhavan leaned into the moment: “my friend the great @facan would never have retired if he thought we would drop the ball….”

Madhavan’s one-liner is best understood against the enforcement system most SEOs know: Google’s two-tier split between manual actions and algorithmic demotions. Google’s long-documented public position — repeated across Search Central documentation and Search Liaison communications — is that most ranking losses are algorithmic, not manual, and that algorithmic demotions have no reconsideration-request path: you fix the issue and wait for the next crawl or refresh. Bing’s stated model, per Madhavan, goes a step further by removing the one-off tier entirely. No published piece we found puts the three mechanisms side by side, so we built the comparison below from Google’s documented policies and the case data above.

*The Bing column reflects Madhavan’s July 10 X statements and Schwartz’s reporting, not published Microsoft policy documentation. The Google columns reflect Google’s long-documented two-tier enforcement system. Note the important nuance on appeals: the “no one-offs” claim is specifically about penalties — it does not mean Bing has zero site-owner communication, since Bing has separately been observed notifying owners when previously blocked sites are unblocked.

The pattern-sweep model also isn’t alien to Google’s history. From the Panda and Penguin eras through the 2024–2026 spam-update waves — including Google’s June 2026 spam update — large-scale enforcement has generally arrived as batch algorithmic sweeps rather than individual reviews. What’s new here is Bing saying the quiet part out loud, and the SEO community getting a timestamped, before-and-after demonstration.

Here is our interpretation of what this episode actually changes. For a year, a quiet arbitrage existed: Google’s spam enforcement was the one that hurt, and Bing’s was widely assumed to be passive. A site could be suppressed to near-zero in Google, keep ranking in Bing, and still harvest enormous ChatGPT and Copilot visibility because assistants ground on Bing’s index. This case demonstrates that the arbitrage carries an expiry date — and that when it expires, it expires all at once. A pattern-based sweep has no rollout blog post, no dashboard warning, and no partial demotion phase. The index count goes from “close to 90K” to zero.

Projecting forward, the asymmetry that made this case possible is likely to keep narrowing. Every publicized sweep gives Bing’s spam systems another confirmed pattern, and every “Mt. AI” collapse teaches AI platforms how fragile single-index grounding is. The rational expectation for 2026–2027 is more simultaneous, unannounced enforcement — and for sites running correlated tactics, the practical risk is that a detector built from someone else’s flagged site sweeps yours in the same pass. That is the operational meaning of “we do not do one offs”: the pattern gets the penalty, not the domain.

If your business benefits from ChatGPT, Copilot, or Perplexity citations today, this case is the prompt to stress-test how durable that visibility is. Three moves cover most of the risk surface, and none of them requires waiting for a penalty to find out where you stand.

The uncomfortable part for high-velocity content operations: the line between “scaled content” and “scaled content abuse” is drawn by the detector, not by your intent. Our own editorial system publishes at volume, which is exactly why we hold it to named-author, verified-source, E-E-A-T-forward standards — the approach behind our agentic SEO service. If your AI-visibility strategy can’t survive a pattern detector trained on the worst site using your tactics, it isn’t a strategy — it’s a countdown.