How fast can AI visibility change?

How fast can AI visibility change? In one measured case, within two weeks. An open-source developer-tools company shipped a targeted fix, about two person-days of work, and its own-domain AI citation rate rose from 1.25% to 5.8%, a 4.7x lift, confirmed independently by two separate measurement tools. The movement stepped the week the work landed and held. That speed is the useful fact here, and it deserves a careful reading: what moved, why it moved that fast, and what one case does and does not prove.

What actually moved in two weeks?

The case, as measured in July 2026: 14 README pull requests plus identity metadata standardized across 16 repositories. No paid placement, no site redesign, no new content program elsewhere. Within two weeks:

• Own-domain citation rate: 1.25% → 5.8% (the share of AI answers citing the product’s own domain rather than a third party describing it)
• Category rank: outside the top 25 → #10
• Visibility score: 22.7 → a range of 26.3 to 30.94 across follow-up measurements

Two independent tools confirmed the movement: the company’s own analytics platform and Collimer’s probes, run separately against the same before/after window, with different methodologies and neither tuned to the other. The full write-up, confounders included, is published as a receipt on the Sandcastle Labs site. The mechanism has its own guide: your GitHub README is a GEO surface.

Why can AI visibility move faster than SEO rankings?

Traditional SEO waits on a pipeline: crawl, index, rank, and the slow accumulation of authority signals. Answer engines shortcut part of that. Retrieval-grounded assistants fetch sources at answer time, and retrieval bots are distinct from training crawlers: a page fixed today can be retrieved, in its fixed form, the next time an engine grounds an answer on it. There is no waiting for a model retrain, because grounded answers are assembled from what the retriever reads now.

That is why the movement in the measured case was a step, not a climb. The fix changed what the engines’ retrieval layer found, and the citation rate moved as soon as answers were grounded on the new text. Content levers of this kind are established beyond this one case: controlled work on GEO content changes reports visibility lifts in the 30 to 41% range from wording and structure changes alone, applied at retrieval time, not over an aging period.

What does one fast case not prove?

Read the speed honestly, the way you would want a measurement instrument to report it:

• One case is strong evidence, not a law. The client’s site team shipped internal linking work in the same window, so some share of the movement may not belong to the README fix. We publish that confounder rather than rounding it away.
• Two weeks was this case’s window, not a guarantee. The product had real adoption and a heavy GitHub footprint; the fix removed a bottleneck that was already cheap to remove. A brand with no retrievable sources at all is solving a different, slower problem: an unresolved identity, not a low score.
• Scores drift. Model updates and panel changes move visibility numbers without any action on your side, which is why a single-shot before/after needs a confidence interval to distinguish signal from noise. The case above cleared that bar: a 4.7x step, held for two weeks, on two instruments.

What makes visibility work move fast?

The pattern from the measured case generalizes into a test you can apply to any proposed GEO work, as of August 2026:

1. It targets a surface engines actually retrieve. READMEs, product pages, comparison tables, FAQs. The levers that move citation are signals in the text engines ground on, not site-wide authority campaigns.
2. It changes what a retriever reads, not what a crawler eventually indexes. Fixes that alter retrievable text can surface in answers within days to weeks. Work that waits on link equity, like classic backlink building, moves the weaker lever.
3. It is measured on a frozen question set, before and after. Speed claims without a fixed panel are noise. The two-week number above exists because the same questions were probed before and after, on two tools.

If the work in front of you fails all three tests, expect months. If it passes them, weeks are on the table, and the way to know is to measure your own baseline first.

The measured case is from July 2026, the most current we have as of August 2026. We will update this guide as more before/after windows accumulate; a pattern from one case is evidence, and a pattern from ten is a benchmark.

See where you stand first. Run a free scan to get your baseline: your citation rate, your score with its interval, and the fixes the measurement says are cheapest to move.