AI Citations Change Every Month -- Even When You Don't

Most AI visibility work is structured like a one-time project. Run an audit, identify the gaps, fix the directories and schema and NAP data, then assume the work holds. The logic feels right: you invested in your entity infrastructure, now it should show up.

The problem is that what AI platforms cite for identical queries changes substantially every month -- regardless of whether you or your competitors have done anything. A business that earned a solid citation position in April may not be cited in June, not because anything about the business changed, but because the citation pool itself shifted around it.

This is citation drift. It's widespread enough to affect how AI visibility audits should be structured -- and how often they need to run.

How Much the Citation Pool Actually Changes

In our June 23, 2026 investigation of source pool stability (`knowledge/citation-drift-and-source-stability.md`, session 56), we documented findings from Machine Relations' analysis of 240 million ChatGPT citations.

The numbers are larger than most practitioners expect. 40-60% of cited domains change month to month for identical queries. Over a six-month window, 70-90% of the citation pool is completely different from the starting set. Only 30% of brands remain visible from one AI answer to the next across that span.

These are not freshness effects -- the underlying content on cited pages isn't necessarily changing. The citation pool itself is unstable due to model updates, retrieval behavior shifts, source trust recalibrations, and the inherent randomness in how AI systems select from candidate sources.

Platform-Specific Retention Rates

The instability is not uniform across platforms. SISTRIX's April 2026 analysis -- which our June 23 citation stability investigation documented -- tracked how much of each platform's citation set for identical queries is retained over four weeks:

| Platform | 4-week citation retention | |---|---| | Perplexity | 44% | | Copilot | 34% | | ChatGPT | 31% | | Google AI Overviews | 27% | | Gemini | 11% |

The ordering follows from each platform's architecture.

Perplexity performs real-time web search for every query, following stable ranking logic -- freshness, Google organic position, domain trust. That determinism makes its citations more consistent over time even though the sources are live-retrieved. The businesses that appear in Perplexity citations tend to recur, because the ranking signals that surface them are themselves stable.

Gemini's 11% retention is the least stable of any platform measured. Gemini grounds its responses in the live Google index, Google Maps, and Knowledge Graph data -- all of which update continuously. A business that appears in a Gemini citation this week may not appear next week if Google's Maps data or index has updated. That volatility is a structural feature of real-time data grounding, not an anomaly.

ChatGPT sits in the middle at 31%, reflecting its hybrid architecture: primary reliance on training data (stable), combined with selective live web search retrieval (volatile).

The Starkest Examples of Overnight Source Shifts

Citation drift operates at a background rate, but it also happens in sudden events that expose how completely citation pools can move.

In our June 23, 2026 citation stability investigation, we documented two of the sharpest source shifts on record.

ChatGPT's source footnote volume collapsed between February and April 2026, triggered by two separate model rollouts. The week of March 8, GPT-5.3 became the default model and responses began citing approximately 20% fewer sources per response. Then on April 19, a second event hit: US zero-citation response rates spiked to 78%, UK and Germany above 80%, in a single week. By April 26, ChatGPT's explicit source footnote volume was down 86-94% from the February baseline across five markets tracked by SEOClarity.

Perplexity's source pool shifted just as quickly in October 2025. After Reddit sued Perplexity for unauthorized scraping, Reddit's share of Perplexity citations dropped approximately 86% within weeks -- documented in our June 2026 platform divergence analysis (`knowledge/platform-divergence-data.md`, session 42, 2026-06-09). Businesses and agencies that had built Reddit presence specifically to appear in Perplexity watched that lever disappear in a single news cycle.

Neither shift was recoverable through optimization. The underlying source pool moved. The businesses that remained visible were the ones with presence across multiple source types -- not the ones that had concentrated on the channel that collapsed.

What Content Survives Drift

Not all citation types are equally vulnerable. Our June 23 citation stability investigation documented a specific finding: only 1.4% of cited news articles remain in the citation set permanently. News citations appear and disappear as model training updates and retrieval patterns shift.

Evergreen content performs better. FAQPage-structured content, service description pages, and structured directory listings recur in citation sets at higher rates -- because the ranking signals that surface them are themselves durable. Directory presence and entity structure don't expire on a news cycle.

For businesses building AI citation infrastructure, this validates the standard approach. FAQPage schema and well-structured service pages outperform time-sensitive content over any horizon longer than a few weeks. A local press mention generates a brief citation window, then drifts out. A Yelp listing with reviews, a BBB profile, a well-structured service page with LocalBusiness schema -- these recur.

The 60-Day Re-Audit Standard

Citation drift has a direct implication for audit scheduling.

An audit captures which sources are cited for specific queries on the day it runs. Given 40-60% monthly drift in citation pools, an audit from 60 days ago no longer accurately represents current visibility. A 90-day audit cycle means roughly half the citation landscape has changed since the measurement was taken. The score is measuring a reality that no longer exists.

Our June 23 citation stability investigation supports a 60-day re-audit window for businesses that have made infrastructure changes and want to measure whether those changes held. Not because AI visibility is unpredictable -- it has consistent structural patterns per platform -- but because the specific citation pool that's confirming or denying a business on any given day shifts often enough that a longer cycle misses meaningful movement in both directions.

Platform Priority and Drift Durability

The SISTRIX retention data adds a practical layer to platform prioritization. Perplexity's 44% four-week retention is the highest of any major platform. A citation won through infrastructure improvements on Perplexity is more likely to still be there next month than the same citation won on Gemini.

Our session 42 (2026-06-09) platform divergence analysis found that Perplexity cites brands at a 13.05% rate per response compared to 0.59% on ChatGPT -- a 46x gap. Perplexity sits at the intersection of two advantages: highest citation frequency and most stable citation retention. For most local businesses, it is both the most achievable near-term citation target and the one where improvements are most likely to hold.

A Sourcepull Signal Check gives you a point-in-time read on where your business currently stands across platforms. Given how much citation pools shift monthly, running one every 60 days is the only reliable way to know whether infrastructure changes are actually holding in practice -- and whether you need to respond to something that changed.