Your AI visibility can move even when you have not changed a word. AI engines quietly update how they turn one question into many background searches, and which sources they reach for, so the same pages can rise or fall on their own. A mid-2026 study by Peec AI, a GEO analytics firm, tracked ChatGPT's query fan-outs and retrieved sources over several months and caught the system shifting under its users, repeatedly. This piece lays out what it found, attributed and hedged, and what to do when the ground moves and your content didn't.
It is the temporal companion to how query fan-out works (the mechanism) and why your AI share-of-voice score swings (the day-to-day noise). This is the slower story: directional drift across weeks and model versions.
Why does visibility change when my content doesn't?
Because the retrieval system is not fixed. When your brand gets mentioned less, or your domain gets cited less, the reflex is to assume something on your page needs fixing. Often the real cause is upstream: the engine changed how it fans a prompt into sub-queries, which sources it trusts, or how many retrieval loops it performs. None of that is visible in the final answer, and none of it is under your control.
Peec AI frames the two metrics that matter as brand visibility (how often you are named in answers) and retrieved sources (how often your domain is pulled in as a source). Its mid-2026 research shows both can move purely because the model updated. The examples below are single-vendor, self-reported observations of mainly ChatGPT, so treat the exact numbers as directional; the pattern is the point.
The default assumption when visibility drops is that you need to fix something on your side. Often the system simply changed which sources and sub-queries it prefers, with no change to your content at all.
What shifts did the study actually catch?
Four documented shifts, all attributed to Peec AI's mid-2026 tracking of ChatGPT unless noted:
| Shift | What Peec AI observed | Read as |
|---|---|---|
| Reddit spike (ChatGPT only) | From ~end of May 2026, fan-outs increasingly appended "reddit" (e.g. "best lawn mower 2026" → "…reviews reddit"); Reddit's share of retrieved domains rose from just under 2% to about 5% (~150% relative), spiked for ~a month, then eased but settled at a new, higher baseline | A brief change can leave a lasting shift |
| "Official" climb (ChatGPT only) | The term "official" rose in fan-outs across languages (German offiziell, Spanish oficial, French officiel), in N-grams like "official site" and "official pricing"; retrieval impact was smaller because "official" is spread across many domains, with a slight lift for consumer-protection and government sites | The engine leaning toward authoritative sources |
| arXiv overtakes Wikipedia (ChatGPT 5.5/5.6) | arXiv's share of the reference source category (Wikipedia, encyclopedias, official docs, glossaries; ~7% of all citations) climbed from just under 1% in April to about 4% in July (~fourfold); Wikipedia's share of that category fell from ~40% in early March to about 7% by July; arXiv passed Wikipedia around end of June | Source preferences reorder over months |
| Luna's precision turn (ChatGPT 5.6 Luna) | Across ~5,000 prompts, Luna's fan-outs used the site: operator ~43% of the time (vs ~0.004% in 5.5), "official" ~22% (vs 1.4%), and year modifiers ~26% (vs 6.2%); it also performed more multi-step retrieval |
Newer models search more like power-users |
Source: Peec AI, mid-2026, primarily observing ChatGPT; single-vendor and self-reported, so directional. ChatGPT 5.6 (the Sol/Terra/Luna tier family) was released by OpenAI on 9 July 2026, with Luna as the efficient tier, which independent coverage corroborates.
Why did arXiv start winning quietly?
Because it became a silent source: pulled in during retrieval but not always named or linked in the final answer. Peec AI reports that in a query for trail-suitable shoes for wide feet, arXiv was referenced six times during retrieval even though a science repository is not the obvious source, and that ChatGPT reads only the paper's abstract page, not the full document.
The move is not driven by a fan-out wording change; it reflects a broader tilt toward scientific reference material. The lesson for measurement: if you only watch the visible citations, you miss sources doing heavy lifting behind the scenes, and you miss the reordering entirely. Watching retrieved sources, not just named ones, is how you see it. It is the retrieval-side cousin of the dark library effect, where pages get read far more than they get clicked.
What does ChatGPT 5.6 Luna's precision turn mean for me?
It rewards exact, structured, first-party facts. Peec AI reports Luna does more iterative retrieval: an initial broad fan-out, then targeted follow-up loops where the site: operator jumps from roughly 20% of the initial fan-out to over 80% of the first iteration. It increasingly pairs site: with quoted attributes and negative operators, essentially Boolean-precise searches for a specific value:
site:<brand>.com "<exact phone number>"to find a service contactsite:<retailer>/productpage "organic cotton" "straight jeans" "black"to match precise product attributes"beige" "recycled polyester" "one-piece" women -aliexpress -instagramto locate an exact item while excluding noise
The examples above are generalised from patterns Peec AI published, not any single brand's data. The takeaway is durable: as retrieval shifts from broad discovery toward precision retrieval, pages that state exact attributes, specifications, pricing, and service details, cleanly and machine-legibly, have a better chance of being the page a targeted search matches. This is consistent with our standing advice to prioritise your structured data and prepare your product catalogue for AI agents; the Luna data raises the payoff.
Isn't this just the same as answer volatility?
No, and keeping them separate matters. Answer volatility is short-term noise: ask the same question twice today and the brand list and citations wobble because the model samples differently each time. The drift Peec AI documents is directional and slower: a sustained change over weeks, driven by a model update, that shifts the baseline itself.
The practical difference is what you do about it. You ride out volatility by sampling many prompts and reading the trend. You respond to drift by noticing the baseline has moved, checking whether it moved for everyone or just you, and adjusting only if it is real and sustained. Confuse the two and you will either over-react to noise or miss a genuine change. This is why a single citation-rate snapshot is unreliable and why measuring AI visibility has to be continuous.
What holds still while the system moves?
The durable levers. Model versions, preferred domains, and fan-out habits will keep changing; what keeps paying off across updates is narrower and more boring:
- Be reachable and parseable. Server-rendered HTML, crawlable by AI bots, in your sitemap. A source the engine can't fetch can't be retrieved no matter how it fans out.
- State exact, structured facts. Numeric, named, dated attributes in clean lists and tables are what a precision search can match; vague prose is not.
- Build entity strength. Being a corroborated, recognised entity in your category is the slow lever that survives whichever retrieval behaviour ships next.
- Measure the retrieval side, over time. Track mentions, retrieved sources, and fan-out behaviour as trends per engine, not one reading, so you can tell a model shift from a your-side regression.
The honest read for 2026: the biggest challenge is no longer just optimising content for AI, but understanding how the AI systems themselves evolve, so a shift you didn't cause doesn't get mistaken for a failure you have to fix. Watching mentions, cited sources, and fan-out behaviour drift across ChatGPT, Gemini, Claude, and Perplexity, over time, is exactly what Buffy Intel is built to do.