Two speeds inside AI search: why the summary moves slower than the citations
Published on 11 September 2026, OtterlyAI released the results of a 40- to 91-day monitoring run across seven AI engines: ChatGPT, Perplexity, Gemini, Google AI Mode, Google AI Overviews, Microsoft Copilot and Claude. 520 prompts, run once per engine per day, produced 252,407 answers. The headline most useful to anyone trying to steer their organisation's AI visibility is not a ranking of engines. It is the gap between how fast two different layers of the same system move.
What churns, what holds
The citation layer is noisy. In a 30-day window, the share of cited sources that appeared on only one day ranged from 26% up to 67%, depending on the engine (Google AI Mode sat at the high end). Day-to-day source reuse varied just as sharply: ChatGPT and Google AI Mode retained roughly a quarter of yesterday's cited URLs, while Perplexity retained about three quarters. Across the full dataset, 46,009 distinct URLs were cited in that month. 48% of them were cited only once. The top 1% of URLs accounted for 44% of all citations.
The brand layer behaves differently. In the same window, only 12% to 19% of brands mentioned for a prompt appeared on just one day. OtterlyAI's own average brand position across six engines (Perplexity was excluded, because 82% of its answers contained no brand mention) drifted between 0.09 and 0.15 places per day across the month, with no clear trend.
Two speeds, inside the same answer. The URL the model points to may change every day; the brand it names tends to stay named.
Why the asymmetry exists
The citation layer behaves like a weekly auction. Which page gets quoted depends on crawl freshness, re-indexing, the exact phrasing of the prompt, which snippets happen to match, and the retrieval choices each engine's pipeline makes that day. A single new article can replace yesterday's winner. That is why 48% of URLs in the dataset were used once and then dropped.
The brand layer is built from something slower. By the time a model is willing to name a company in response to a prompt, it has absorbed enough overlapping references, consistent descriptions and category associations to treat that name as a safe answer. That store does not reshape itself on a weekly cadence. New citations may feed it over time, but one new page does not rewrite the summary the model has already settled on. OtterlyAI's observation that brand position moves only around a tenth of a place per day fits that picture: the baseline moves, but slowly.
What this means for brand teams
Most AI-visibility work today optimises the fast layer. Teams track citations week by week, publish to chase whichever page is being quoted, and report on movement that is largely noise in a window where half the cited URLs will never be cited again. The signal they are trying to influence is real, but it sits in the thin, replaceable layer of the system.
The compounding asset sits in the layer underneath. It is written in the sentences a model is prepared to produce about the organisation when no one has prompted for a source link: the responsibilities it attributes, the positions it quotes, the category it places the company in, the comparators it names alongside it. Those sentences are built from repetition across controlled surfaces (the company's own site, executive bios, pillar essays) and from compatible phrasing in third-party coverage. They change on a quarterly, not weekly, rhythm.
For a UK brand team planning the next two quarters, the practical consequence is a reallocation rather than a new tool. Short-term tactics such as page-level optimisation, citation tracking and prompt-by-prompt audits belong in the operational layer and should be budgeted as such: cheap, continuous, not expected to move the baseline. The budget worth defending sits with the slow layer: an accurate, specific description of what the organisation does, written in sentences that other people can quote verbatim; a small number of pillar pieces that state the leader's positions in full sentences; and a handful of external placements that use compatible phrasing. The repetition across controlled surfaces is what gets summarised back later.
Caveats worth stating
OtterlyAI publishes AI-visibility software, so the study is primary data from an interested party. All runs were collected in the United States, with one run per prompt per engine per day. That measures day-to-day variation, not intra-day volatility, and it does not describe UK-specific behaviour. The authors are explicit that the ranges above are not fixed benchmarks and that each brand should build its own baseline.
With those caveats, the structural finding holds across every engine the study covered: the brand layer moves roughly an order of magnitude more slowly than the citation layer. That is a design feature of how these systems retrieve and summarise, not a quirk of one model.
The decision this forces
The question for a brand team is not whether to track AI citations. It is which layer deserves the next hire, the next retainer, the next quarter of planning. A citation dashboard will move every week by design. A baseline summary, once accurate, is harder to displace, and it quietly determines which citations the model reaches for in the first place. Over the horizon that matters to a brand's economic value, the slower layer is the one that compounds.
Source: OtterlyAI, "AI search visibility stability", 11 September 2026.
Source: https://otterly.ai/blog/ai-search-visibility-stability/