Domain-level AI citations are remarkably stable: BrightEdge found that 96.8% of cited domains saw zero change in any given week. But URL-level citations are volatile: Writesonic tracked 23 million cited sources and found 44% of pages appeared exactly once before disappearing from AI answers entirely. Both findings are correct. They measure different layers of the same system — and the gap between them reveals the two-tier structure that determines which sources hold AI visibility and which cycle through it.
Last updated: July 10, 2026
The Paradox: 96.8% Stable and 44% Disposable #
The apparent contradiction between BrightEdge and Writesonic dissolves once you separate domain-level authority from page-level selection.
BrightEdge analyzed citation and mention behavior across ChatGPT, Gemini, Google AI Mode, Google AI Overviews, and Perplexity. Their finding: among domains that are part of the cited set for a given prompt, 96.8% stayed in that set from one week to the next. For brands in the top mention positions (#1 or #2), stability reached 99.4% — effectively cemented.
Writesonic, tracking 23 million sources across ChatGPT, Gemini, Perplexity, and Google AI Overviews from April to June 2026, measured the page layer underneath. Their finding: 44% of cited pages appeared once and never returned. The median citation lifespan for a specific URL was 11 to 15 days. Trakkr's 10-month longitudinal study across 10,000 brands and 7 AI models corroborated this: 73.4% of cited URLs appeared only once, with a 30-day peak-to-half decay.
The two layers explain each other. A domain like Forbes, G2, or Crunchbase remains in the cited set week after week (BrightEdge's 96.8%). But the specific Forbes article or specific G2 page an engine cites on a given query rotates frequently (Writesonic's 44% one-time). The domain is the structural anchor. The page is the replaceable unit.
This is why point-in-time citation counts can mislead. A domain cited 65 times in one 30-day window and 65 times in the next may have entirely different URLs behind both counts. The domain authority held. The pages churned.
Engine-Level Stability Differs Sharply #
Not all AI engines treat citation stability the same way. Profound's longitudinal tracking measured monthly citation drift — the share of cited domains that changed from one month to the next — and found wide variation:
| AI Engine | Monthly Citation Drift | Interpretation |
|---|---|---|
| Perplexity | ~40.5% | Most stable; retains sources roughly 2x longer than ChatGPT |
| ChatGPT | ~54.1% | Most volatile among conversational engines |
| Copilot | ~53.4% | Similar volatility to ChatGPT |
| Google AI Overviews | ~59.3% | Highest drift; reshuffles citation set most aggressively |
Perplexity's lower drift aligns with Writesonic's finding that Perplexity retains sources approximately twice as long as ChatGPT. The difference is architectural: Perplexity's retrieval pipeline indexes and re-retrieves from a crawled web corpus, creating more stable source relationships. ChatGPT's retrieval changes with model updates — after the GPT-5.5 update in May 2026, ChatGPT's share of citations going to third-party sites jumped to 96%, with the average vendor site capturing just 4.3%.
Google AI Overviews' high drift rate is consistent with its role as a surface embedded in traditional search, where query volume and diversity expose more borderline source relationships to rotation. The Gemini 3 rollout in January 2026 reshuffled approximately 42% of cited domains and increased sources-per-answer from roughly 11.5 to 15 — demonstrating that a single model release can rewrite the citation surface overnight.
Cross-platform overlap is low. Profound's analysis of 595 prompts found 76% of citations were unique to a single platform, and only 0.8% appeared across all four measured engines. A domain stable on Perplexity may be absent on ChatGPT for the same query. Per-engine measurement is not optional — aggregate citation tracking across engines masks the structural instability within each one.
The Core-Fringe Dynamic: Why Volume Predicts Churn #
BrightEdge's data reveals a counterintuitive pattern: domains with larger citation footprints are more likely to see week-over-week changes.
| Domain Tier | % That Changed | Typical Fringe Size |
|---|---|---|
| Top 50 domains by volume | 90% | ~5% of citation share |
| Domains with 100+ citations | 65.2% | ~17% of citation share |
| Top 10% by volume | 21.1% | Larger shifts |
| Bottom 50% by volume | 0.4% | Minimal |
The explanation is structural. Every domain's citation footprint operates in two zones:
The core — prompts where the domain is clearly the best source. Citation here is stable because no alternative source satisfies the query as well. A domain cited on a handful of highly specific prompts is almost certainly in its core zone for all of them.
The fringe — prompts where the domain is borderline relevant, ranked 8th or 9th among candidate sources. This is where rotation happens. Among the top 50 domains by volume, 90% have a fringe — but it typically represents only about 5% of their total citation share in play any given week.
The operating implication: a domain with 200 citations may see 10 change in a given week. The 190 stable citations are the core. The 10 that rotated were fringe. Measuring total citation change without separating core from fringe overstates volatility for high-authority domains and understates it for low-authority ones.
When Citations Change, They Disappear — Not Redistribute #
BrightEdge's most striking finding: among the roughly 3% of domains that did see citation changes in a given week, 87% of changes were losses. Only 13% were gains. And the losses were binary — domains went from cited to not cited on a given prompt, not from position 3 to position 5.
| Direction | % of Changes | Share of Citation Volume |
|---|---|---|
| Declining | 87% | 51.3% |
| Growing | 13% | 5.3% |
| No change | — | 43.5% |
Over 51% of all citation volume was associated with declining domains. Only about 5% was associated with growing ones. The losses were not redistributed to new entrants. Only 0.4% of all tracked domains gained new citations in a given week. AI engines are tightening their citation radius — becoming more selective about what they link to, not swapping one source for another.
Citation concentration reinforces this dynamic. BrightEdge found the top 1% of domains captured 64% of all citations. The top 10% captured 84%. 5W's State of AI Search 2026 annual report, synthesizing over 10,000 buyer-intent prompts across five engines, confirmed the same pattern: citation share is concentrating faster than traditional market share, and category leaders are extending their lead while challengers without active programs lose ground.
The door in is narrow and getting narrower. But for domains already inside the cited set, the position is durable.
Stability by Source Type: Institutional Sources Hold #
BrightEdge's data by website type reveals which source categories experience the most citation churn:
| Source Type | % of Domains That Changed | % of Changes That Were Declines |
|---|---|---|
| Finance | 51.1% | 91% |
| Review sites | 45.5% | 100% |
| News/media | 44.8% | 92% |
| Reference/encyclopedia | 38.5% | 80% |
| Health/medical | 34.2% | 100% |
| eCommerce/retail | 23.1% | 73% |
| Tech | 15.2% | 91% |
| Government/institutional | 3.6% | 77% |
Government and institutional sources were the most stable — under 4% changed, and when they did, 23% of changes were actually gains. Tech sources were second most stable at 15.2% change rate. Finance, review sites, and news/media were the most volatile, all above 44%.
The pattern maps to source function. Government and institutional sources serve reference queries where no alternative source provides the same data. Review sites and news/media serve competitive queries where multiple sources can satisfy the same need. The more replaceable a source is for a given query type, the more volatile its citation behavior.
BrightEdge also identified a growing divergence between citations (linked references) and mentions (unlinked name references). Social platforms saw citation declines of 34-45% alongside mention gains of 11-18%. Financial analysis sites saw citation declines of 35-57% alongside mention gains of 20-65%. AI engines are still referencing these sources by name but increasingly choosing not to link to them. Brand authority and citation authority are becoming distinct dimensions.
What Predicts Citation Durability #
Across BrightEdge, Writesonic, Trakkr, 5W, and Profound's independent measurements, the structural properties that predict citation durability converge:
1. Source-type exclusivity. Sources that answer specific query types where no alternative is equally valid hold citations longest. Government data, structured comparison databases, and methodology-specific reference sources exhibit the highest stability because their data is not available elsewhere. This is what BrightEdge's "core zone" captures — queries where the source is indispensable, not interchangeable.
2. Continuous freshness, not periodic campaigns. Writesonic's 11-15 day median citation lifespan and 5W's documentation of measurable citation decay when brands pause earned media investment both confirm the same finding: citation authority is an operating expense, not a project. Trakkr's 10-month study found that AI visibility, once earned, does not stay earned without sustained signal input. Sources with continuously updated structured data (market databases, review platforms with live ratings) maintain citation rates without deliberate freshness campaigns because their data updates structurally.
3. Cross-engine presence. With 76% of citations unique to a single platform, a source visible on only one engine has structurally fragile citation authority. Domains maintaining citation presence across multiple engines have independent retrieval pathways — meaning a model update on one engine does not eliminate their visibility. The 5W annual report found that each engine has distinct citation preferences: ChatGPT weights community content, Claude weights institutional sources, Perplexity prioritizes real-time data, and Google AI Overviews draws heavily from blogs. Multi-engine stability requires satisfying multiple retrieval architectures simultaneously.
4. Earned third-party validation. 5W's annual report identified earned media as the strongest single input to citation authority across every category tested. Ahrefs' 75,000-brand study found branded web mentions correlate with AI visibility at approximately 0.664 — roughly 3x stronger than backlinks. YouTube mentions showed the strongest correlation at approximately 0.737. The common thread: sources that other sources reference are structurally harder to rotate out of citation sets because the retrieval system encounters them through multiple independent pathways.
How This Connects to Machine Relations #
The Machine Relations Index measures citation rates per domain across AI answer engines — how frequently each source is cited when a relevant segment of queries is answered. The MRI's confidence tiers (A, B, C, or collecting) reflect measurement stability: an A-confidence rate is backed by dense evidence across many run dates, meaning the citation behavior is structurally stable. A collecting domain has insufficient evidence to produce a settled rate.
The third-party data synthesized here validates the measurement architecture the MRI is built on. BrightEdge's finding that 96.8% of domain-level citations are stable week-over-week confirms that citation rates measured at the domain level are meaningful signals, not noise. Writesonic's finding that individual URLs churn at 44% confirms that page-level citation tracking would produce unstable, unreliable measurements — which is why the MRI measures at the domain level rather than the URL level.
The core-fringe dynamic BrightEdge identified maps directly to the MRI's evidence floor and confidence system. A domain with A-confidence in the MRI has been measured across enough observations and distinct dates that its citation rate is in the core zone — stable, reproducible, and structurally anchored. A domain still collecting data may be in the fringe zone — present in some measurement windows but not consistently enough to produce a settled rate. The confidence tier is the MRI's answer to the stability question: not all citation rates are equally durable, and the tier tells you how much trust to place in the number.
For practitioners building citation architecture, the stability evidence points to a specific strategy: optimize for the properties that put your domain in the core zone (source-type exclusivity, structured data, cross-engine presence, earned validation) rather than optimizing for the properties that increase raw citation counts. A domain with 50 stable core citations has more durable AI visibility than a domain with 150 citations, 100 of which are in the fringe zone and rotate weekly.
FAQ #
How stable are AI search citations week to week? #
At the domain level, extremely stable. BrightEdge found 96.8% of cited domains unchanged week-over-week, with top-ranked brands (#1 or #2 position) stable at 99.4%. At the page level, volatile — Writesonic found 44% of cited URLs appeared exactly once before vanishing, with a median lifespan of 11 to 15 days. The two layers measure different things: domain authority holds while specific pages rotate beneath it.
Which AI engine has the most stable citations? #
Perplexity shows the lowest monthly citation drift at approximately 40.5%, retaining sources roughly twice as long as ChatGPT (~54.1% drift). Google AI Overviews has the highest drift at approximately 59.3%. These figures come from Profound's longitudinal tracking across 2026. Stability also varies by source type — government and institutional sources are stable across all engines, while news/media and review sites show high volatility.
How long does an AI citation last? #
The median citation lifespan is 11 to 15 days according to Writesonic's study of 23 million cited sources. Trakkr's 10-month study found a 30-day peak-to-half decay rate and that 73.4% of URLs were cited only once. However, these figures describe page-level citation. Domain-level citation authority persists much longer — BrightEdge's data shows 96.8% week-over-week stability for domains already in the cited set.
What makes AI citations more durable? #
Four properties predict citation durability across independent measurement systems: source-type exclusivity (answering queries where no alternative is equally valid), continuous freshness (not periodic campaigns), cross-engine presence (cited on multiple AI platforms, reducing dependency on any single engine's retrieval changes), and earned third-party validation (branded mentions and independent references from other authoritative sources).
Sources: BrightEdge AI Catalyst™ weekly analysis of citation and mention behavior across 5 AI engines. Writesonic study of 23 million cited sources across ChatGPT, Gemini, Perplexity, and Google AI Overviews (April-June 2026), reported via Foundation Inc. Trakkr 10-month longitudinal study across 10,000 brands and 7 AI models. 5W State of AI Search 2026 annual report: 10,000+ buyer-intent prompts across 5 engines. Profound/LumenGEO citation drift and overlap data. Ahrefs 75,000-brand correlation study (published Dec 2025, re-promoted May 2026). Machine Relations Index v2 methodology: What is Share of Citation.
Last updated: July 10, 2026
Additional source context #
- This guide explains the "rank and tank" pattern in LLM citations, why it happens, and what you can do to hold your position. (AirOps Rank and Tank: Why AI-Generated Content Sometimes Drops Out of LLM Citations After Initial Gains – Surferstack (s, 2026).
- KPIs, tracker comparisons, how broad my citation footprint is. (AI Citations Have a Half-Life. I Tracked Mine for 9 Weeks and Watched Them Decay. - DEV Community (dev.to), 2026).
- (Our Data) | FogTrail Free Visibility Scan Back to blog # How Often Do AI Search Engines Update Their Citations? (How Often Do AI Search Engines Update Their Citations? (Our Data) | FogTrail (fogtrail.ai), 2026).
- AI Citation Drift: What the Data Really Shows About LLM Source Stability | SEO Francisco SEO # AI Citation Drift: What the Data Really Shows About LLM Source Stability AI citation drift is real. (AI Citation Drift: What the Data Really Shows About LLM Source Stability | SEO Francisco (seofrancisco.com), 2026).
- AI Search Keeps Changing: Why Your Visibility Won't Hold Free AI Check Loudmink has tracked 25 B2B SaaS brands across five AI search engines every week since spring 2026, and engine behavior keeps moving in both directions. (AI Search Keeps Changing: Why Your Visibility Won't Hold (loudmink.ai), 2026).
- Domains cited frequently experience 0.7% weekly volatility, while those cited sporadically swing 50%+—a 70x difference that holds across every platform. (AI Search Engine Citation Volatility Analysis: How Authority Creates 70x More Stability (brightedge.com)).
- AI Citation Volatility Benchmark 2026: Why 50% Decay in 13 Weeks Is the New Normal | Searchless.ai Blog - AI Visibility provides external context for how stable are AI search citations week to week.