Here is a question most brands never test: when you delete a page, does it truly disappear from AI Search? We ran the experiment. One brand deliberately removed 15 comparison pages on known dates, and we tracked their citations across seven AI Search Engines every day, before and after each takedown.
The pages kept getting cited. Not for a day or two, but for weeks, and for one older batch, for over 8 months. We call this a dead citation: an AI answer built on a source that was live when it was first cited, then removed, archived, or left to go stale, that still lives in the engine’s memory and keeps getting cited.
This is the first of two reports. Here we watch dead citations form in real time through a controlled takedown. Our companion large-scale study measures how common they are across 20 million cited URLs.
Key Findings
The pages never truly reached 0% AI citations even months after deletion, these dead pages still live on in the memory of the AI Search Platforms.
8 months later, AI still continued citing the deleted pages
Removing the pages barely dented citations. The freshly deleted batch even increased by +27% citations two days after it was deleted.
93% of citation volume survived the first week after removal. The decline only started in the second week after page deletion.
Claude still cited 78% of the pages weeks after removal, the most persistent of the seven engines. Indicating slower memory renewal.
Gemini produced 31% of all dead citations, the largest volume, on an average persistence rate.
Why It Matters for AI Search
AI Search Engines answer from what they have saved, not only from what is live now. A page you took down, archived, or let go stale keeps shaping answers after it stops being true.
For brand and legal teams, taking a page down does not take it out of AI answers. About 35% of the pages’ citations came after they were deleted.
Citations still point to pages that are gone, so your citation count looks higher than your real visibility, and each AI Search Engine inflates it by a different amount.
For users, answers arrive built on deleted, archived, or years-old sources, presented as if they were current.
Scope of Study
This article covers one live takedown experiment: a single brand, its comparison pages, and what happened to their AI Search citations when the pages were removed. It reports what we observed in monitored prompts over a set window, not a universal rule for how every engine behaves at all times.
Citation counts vary by prompt, country, and engine, so treat these as directional patterns in one brand’s data. Everything is anonymized per the brand’s agreement, and all figures are percentages. For the web-wide picture, see the companion large-scale study.
Disclaimer: The participating brand agreed to publish this case study in exchange for anonymity, for its own brand and domain and for those of its competitors. Brand names are replaced with letters and URL slugs with placeholders, for example best-A-alternatives. Dates are shown as day numbers relative to each removal. The citation data is unchanged.
Methodology
The brand published 15 comparison pages targeting ten competitors, in three batches: four older versus pages, for example brand-vs-A, and two newer sets of five alternatives pages each, for example best-A-alternatives.
Each batch was removed on a known date, which we set as its Day 0. The four versus pages were removed roughly eight months before this window; the two alternatives batches were removed two weeks apart inside it.
We tracked the brand’s prompts in OtterlyAI across all seven AI Search Engines it monitors, ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Gemini, Microsoft Copilot, and Claude, in the US market.
We recorded daily citation counts for every page across a 37-day window that covers both fresh removals and the tail of the older batch.
We aligned each batch on its own removal day and indexed daily citations to that batch’s pre-removal average, where 100% is a normal pre-removal day, so the batches are comparable on one axis.
We split every citation by engine to see which engines kept citing the removed pages, and for how long.
Results at a Glance
What we measured
Result
The memory lives on
The page citations never reached 0%. Even 8 months after deletion.
Citations two days after removal
+27% vs the pre-removal average
Citation volume kept in week one
93%
Citations that remained after removal
About 35% (21% – 66% per deleted page)
Most persistent engine
Claude, 78% of its citations came after removal
Largest dead volume
Gemini, 31% of all dead citations
Longest tail
An older batch still cited 8 months after removal (9-15% citations remained)
The memory lives on. Removing the pages did not remove the citations.
Both takedowns produced the same shape. The first alternatives batch peaked two days after removal at +27% of its pre-removal average, higher than any day the pages were live. It held 93% of its volume through the first week, and the decline only began around day nine. The second alternatives batch repeated the pattern two weeks later, topping its pre-removal average on day three and keeping 79% of its volume through week one.
What this means for your strategy: A takedown does not take effect in AI Search for weeks. If a page has to come down, plan for a multi-week window where it keeps getting cited, and decide in advance where you want that demand to land.
The fade is slow, not clean. Three weeks after removal, the first batch was still earning 15% of its pre-removal volume, every citation of it pointing at pages that no longer exist.Across all ten removed alternatives pages, about 35% of total citations arrived after removal. Per page, the share ranged from 21% to 66%.
Batch 2:
Removed page
Citations remaining after removal
/best-A-alternatives/
66%
/best-B-alternatives/
42%
/best-C-alternatives/
37%
/best-D-alternatives/
36%
/best-E-alternatives/
25%
Batch 3:
Removed page
Citations remaining after removal
/A-versus-X/
49%
/B-versus-X/
33%
/C-versus-X/
28%
/D-versus-X/
22%
/E-versus-X/
21%
This fits index lag. AI Search Engines answer from stored snapshots that refresh on their own schedule, so a removed page keeps appearing until each engine recrawls it and updates its store. The schedule varies by engine and page, which the per-engine split below makes visible.
Batch 1: The eight-month dead citation
The older batch shows how long the tail can run. Its four versus pages were removed about eight months before this window. They were still being cited on the first day of tracking, and they kept earning citations months after they were gone.
Removing these pages on day one did not prevent months of dead citations. They accounted for 5% of everything the brand’s comparison pages earned in this window, without existing for a single day of it.
What this means for your strategy: Dead citations are not only a short-term recrawl lag. Old, long-deleted pages can keep surfacing for months, so a page you removed last year may still be shaping answers about your brand today.
Which AI Search Engines keep citing the dead pages?
The engines split three ways.
AI Search Engine
Share of its citations that were dead
Share of all dead citations
Claude
78%
22%
Google AI Mode
38%
11%
Google AI Overviews
32%
13%
ChatGPT
30%
16%
Gemini
29%
31%
Perplexity
27%
7%
Claude is the outlier. 78% of its citations to these pages arrived after the pages were removed, more than double any other engine, and it kept citing them steadily through the end of the window.
Gemini sits at the other end of the same problem. Its persistence rate is average at 29%, but it cites these pages in such volume that it produced 31% of all dead citations, the largest share of any engine. Microsoft Copilot never cited these pages at all, live or removed, the only engine with a clean record here by absence.
What this means for your strategy: There is no single engine to watch. Claude will keep citing a removed page the longest, and Gemini will produce the most dead citations by volume, so a removal you consider done may still be active on both.
So which AI Search Engine has the most dead citations?
There are three answers, depending on what you measure.
Claude, by persistence. 78% of its citations to the removed pages arrived after removal, and it was still citing them three weeks later, more than double any other engine in this experiment.
Gemini, by volume. Its persistence rate is average, but it cites these pages so heavily that it generated 31% of all dead citations, the largest share of any engine.
Google AI Mode, by rate at scale. In our companion large-scale study, 33% of the URLs it cites are no longer live, the highest of the seven engines across 20 million checked URLs.
Microsoft Copilot avoided the experiment’s pages entirely. If your removed or outdated content is being kept alive in AI answers, the engine doing it depends on the failure mode you care about: persistence, volume, or rate.
The Dead Citations Playbook for GEOs
Audit your own inaccurate dead content. Lost, outdated, or forgotten PDFs, reports, help docs, and comparison pages may still feed AI search engines the wrong information about your brand. Correcting dead content after it lives inside an AI answer is difficult (as seen from this example, dead citations may live on up to 8 months in the memory on an AI), especially when the original source is hard to find or no longer maintained. Catch inaccurate citations early in OtterlyAI before they mislead your target audience, distort brand perception, or affect ROI.
Word of caution: Gray hat tactics – it works until it doesn’t. Some AI Search experts may use dead citations as a gray hat GEO tactic, publishing content in places like Wikipedia, Reddit, or comparison pages with the expectation that it may later be deleted, moderated, or taken offline.
The tactic is simple in theory: inject brand, competitor, or category information into sources AI search engines may crawl, summarize, or cite, then rely on the fact that AI systems can retain or repeat parts of that information after the original page disappears.
This can be tempting for brands without strong authority signals. Even if a brand does not meet Wikipedia notability requirements, a temporary page or discussion may still create an entity trace that influences how AI search engines describe the brand.
Another risky use case is competitor-brand mentions. Some teams may remove comparison pages or sensitive claims from their own website while hoping the language continues to live inside AI answers. But removal does not guarantee safety, accuracy, or reduced liability.
OtterlyAI recommends treading carefully. Dead citations are unstable, engine-specific, and highly dependent on crawl behavior, retrieval systems, and model updates. What creates visibility today may disappear, backfire, or become detectable tomorrow. Algorithms often change or get corrected.
The safer GEO playbook is to monitor, not manipulate. Track removed URLs, stale pages, dead citations, and competitor citation gaps in OtterlyAI. Separate live Link Citations from dead citations, because visibility pointing to deleted or outdated sources should be counted as risk, not a win.
What Comes Next: Largescale Dead Citation Study
This experiment shows dead citations forming in one brand’s pages. The obvious next question is how common they are everywhere else on the internet?
Our companion report, the Dead Citations Study, checks the live status of 20 million URLs cited across the seven AI Search Engines. The short version: about 23% already point to a page that is gone, moved, or unreachable, and that is only the part a status check can see. We will link it here when it is published.
Got a GEO experiment idea? Let’s collaborate
At OtterlyAI, we don’t write about AI Search from the sidelines. We test it. With the community, we run and update a public GEO Experimentation Tracker showing what we are running and what moved AI Search. If you are running your own GEO experiment, email rick.tousseyn@otterly.ai and we may feature it with your name attributed.
👉 Want to measure your AI Visibility? Start with OtterlyAI and our GEO Guide.
