56% of AI Citations Rotate Every Week. Here's What Stays.

Most AI visibility work is framed as a correction problem. Find the gaps, fix the infrastructure, let the citations accumulate. The logic assumes that once you're in the AI citation pool, you stay there.

The actual citation pool doesn't work that way. Most of it is replaced every week -- not because anything changed about the businesses being cited, but because the underlying source pools that AI platforms draw from are continuously rotating. A business cited by Google AI Mode this week has a roughly one-in-three chance of not appearing in that same source pool next week, for the same query, with no changes made by anyone.

Understanding this doesn't make AI visibility impossible. It changes what kind of work actually compounds versus what kind gets washed away in the next rotation cycle.

The Weekly Rotation Numbers

In our 2026-07-30 methodology review of SISTRIX's longitudinal citation study, we documented the most granular public dataset on AI citation stability: 82,619 prompts, 1,548,213 snapshots, tracked across 17 consecutive weeks and three platforms.

The weekly rotation rates:

- **Google AI Mode:** 56% of the source pool changes weekly - **ChatGPT Search:** 74% of the source pool changes weekly - **Co-citations** (sources cited alongside a primary brand): 89% weekly rotation -- near-complete turnover

These aren't aberrations from a short tracking period. The 17-week window in the SISTRIX study is long enough to rule out measurement noise.

Model updates compound the rotation. GPT-5.5's launch shifted 47% of ChatGPT citations within 48 hours. Gemini 3 replaced 42% of AI Overviews cited domains overnight. Model update events don't override the steady-state rotation -- they stack on top of it.

One finding from the same study stands out against all this volatility: **brand domains hold at 43% stable presence across all 17 weeks.** Not the majority of the pool. Not even close. But a definable, measurable stable layer exists, and brand domains are the primary occupants of it.

What "Stable" Actually Means

The 43% anchor figure describes brand domains -- the primary website of a business, not third-party references to it. News article citations, trending content pages, and co-citations rotate at 74-89% weekly. Brand domains rotate at a materially lower rate.

This distinction changes how to think about co-citation strategies. If your business appears in AI citations primarily because you're mentioned alongside a larger brand in a review article or comparison page, you're in the 89% rotation tier. That citation likely won't survive next week's refresh cycle. The source that cited you alongside the larger brand will be replaced, and so will the association.

Anchor status requires your brand domain itself to be a trusted source -- not your mentions elsewhere. Getting into the 43% stable layer means becoming the kind of entity that AI platforms return to on their own, for the queries relevant to your business, week after week.

How Each Platform Establishes Entity Trust

What makes a brand domain "anchor-stable" is not identical across platforms. Our 2026-08-01 analysis of citation source architecture (Scout's `platform-citation-behaviors.md`, built from Yext's 2026 AI Visibility study of 17.2 million citations) documents the specific signals each platform uses to establish entity trust.

For ChatGPT, third-party directories account for 49% of citations. ChatGPT resolves entity identity through its training corpus, and that corpus is heavily weighted toward directory listings -- Yelp, BBB, G2, Capterra, industry-specific registries. A brand that appears consistently across three or more of those directories gives ChatGPT a stable entity record to return to. A brand with a strong brand site but thin directory coverage gives ChatGPT weak entity resolution material -- and ends up in the volatile 51%.

For Gemini, brand-owned websites account for 52% of citations. Gemini is more Google-native: it weights domain authority, structured data, and Google Business Profile data. A brand whose site has well-implemented schema, consistent GBP data, and E-E-A-T signals is the entity Gemini trusts most. Directory listings matter, but less so -- Gemini's primary anchor signal is the brand domain itself.

For Perplexity, neither of those fully applies. Perplexity retrieves live pages per query and favors niche expert directories specific to the business category over general-purpose platforms. A plumber needs to appear in contractor trade directories. A law firm needs to appear in legal directories. The general directory infrastructure that anchors ChatGPT results is necessary but not sufficient for Perplexity.

The practical implication: true cross-platform anchor status requires different infrastructure for each platform. Businesses that focus on one channel -- building directory presence for ChatGPT, or improving the brand site for Gemini, or targeting niche registries for Perplexity -- are partial anchors. They hold on one platform while rotating on others.

The Entity Cluster Density Mechanism

Our 2026-06-06 research into schema markup effects (Scout's `schema-markup-effects.md`, Fischman correlational study, session 39) identified the underlying mechanism that drives anchor stability: entity cluster density.

The finding, from Fischman's cross-platform study of 730 AI citations: sites implementing attribute-rich, domain-specific schema are cited at a 54.2% rate, compared to 31.8% for generic CMS-default schema, among lower-authority domains. The gap isn't the JSON-LD property itself -- it's the verified entries at the URLs the schema points to.

A `sameAs` link in your Organization schema pointing to a Wikidata Q-item, a Crunchbase profile, and your Yelp profile creates three separate traversal paths that AI systems can use to resolve your entity. One traversal path (brand site only) produces weak entity resolution. Six traversal paths -- site plus Wikidata plus three directories plus GBP -- produce a dense entity cluster that AI systems return to with confidence.

Entity cluster density is what distinguishes brand domains in the stable 43% from brand domains that rotate with the rest of the pool. Both have a brand site. The stable ones have a web of consistent, verified signals pointing to and from that site. The rotating ones don't.

Fix Timelines in a Weekly-Rotation World

Our 2026-07-30 methodology recommendation on this data includes a specific guidance update on fix timelines: **4-12 weeks before consistent AI Mode citation appears after infrastructure fixes.** The 56% weekly rotation explains why.

A newly added directory listing or an updated GBP profile needs to survive multiple rotation cycles before it registers as a stable anchor signal. One rotation cycle equals one week. Becoming a consistent presence in the stable anchor pool means surviving four to eight rotation cycles in a row -- each of which is independently probabilistic.

This is why evaluating fix effectiveness at two weeks produces misleading results. It's also why clients who say "we fixed our directories and nothing changed in Perplexity" after a week are not wrong about what they observed -- they're just looking at a window too short to see signal propagation. The 60-90 day re-audit cadence we recommend isn't arbitrary; it maps to the number of rotation cycles required for stable anchor presence to establish.

Infrastructure fixes that compound do so because they shift you from a probabilistic citation (showing up randomly when the pool happens to include you) to structural citation (showing up consistently because your entity cluster is dense enough that the rotation filter can't easily exclude you).

What to Build

The path to the stable 43% is the same across every business category, even if the specific directories differ:

Build directory presence in the platforms your category's AI citations actually draw from -- not just the most popular directories. Run a search on each platform for the category query most relevant to your business and look at which domains appear as sources. Those are the directories you need to be in, not the ones that rank highest in a directory popularity list.

Make your entity cluster dense. Your brand site, your Wikidata entry, your GBP profile, and your directory listings should all use identical name, address, and contact information. Every variation creates a separate entity that competes with the one you want AI to cite. Every consistency reinforces the same entity cluster.

Set a 60-90 day re-audit window. The rotation math is not forgiving of shorter evaluation periods.

If you want to know your current citation status -- which platforms are citing you, which sources they're drawing from, and where your entity cluster has gaps -- a Sourcepull Signal Check gives you a per-platform breakdown in a few minutes. The rotation data makes clear that this is not a one-time measurement; it's an ongoing one.