A team rewrites the opening answers and updates the schema on twelve product pages on a Monday. By Friday ChatGPT still cites a competitor, so the team rolls the change back and ships the next idea. Nothing was actually tested. Citation movement in AI answers usually shows up two to six weeks after a change, so any AEO testing cadence shorter than that window is measuring noise. The fix is to schedule releases around when crawlers actually come back.
Key Takeaways
- Citation changes trail releases because a known page has to be recrawled, reindexed and then retrieved for a live prompt. Figures Google has shown put a typical refresh of a known URL at about 30 days.
- If you ship changes faster than the read window, their effects overlap on the same pages, and no single change can be credited or blamed for what follows.
- Start each test’s clock at the first verified recrawl in your server logs. Run changes in staggered page cohorts against an unchanged holdout, and judge results at six weeks.
Where the Two to Six Weeks Actually Goes
The delay is made of stages that run one after another. Google’s documentation on recrawl requests says crawling “can take anywhere from a few days to a few weeks,” and adds that a request does not guarantee the page will reach results at all.
In October, Google’s Gary Illyes gave more detail at Search Central Live. Search Engine Roundtable and Search Engine Journal both reported his figures. Typical discovery of a new URL took about 20 hours. Typical refresh of a URL Google already knew took about 30 days, and the slowest refreshes took “weeks to never.” Sitemap processing usually took about 24 hours but could take up to 14 days. Structured data updates took anywhere from hours to two weeks. Search Engine Journal’s write-up made the key point: a page has to be crawled before it can be indexed, so the delays add up.
The AI layer adds more steps. After the index updates, the engine still has to retrieve your page for a given prompt, and then pick it as a source over the alternatives. Answers change from one run to the next, so you need several weekly readings before you can call something a trend. That is where the range comes from. Two weeks is a fast recrawl plus a couple of tracking cycles. Six weeks is a roughly 30-day refresh plus indexing plus enough readings to trust the direction.
This window only covers retrieval presence. Training-data presence builds slowly through wide coverage over many months, and a six-week test cannot measure it. Keep the two layers in separate experiments.
Why Shipping Every Week Breaks Attribution
Suppose a page cluster gets a change every week and each change takes up to six weeks to show. That means up to six changes can be in flight on the same URLs at once. When Share of Citation rises in week five, you have no honest way to say which change caused it. The early rollback in the opening example makes things worse. The crawler may pick up the original change on one visit and the reversal on the next, and both show up in your tracking as unexplained swings.
A test that ends before the crawler returns has only measured the old page.
You can measure as often as you like. Judgement has to wait for the window to close. Our NewDay USA engagement worked this way: we tracked results weekly on five engines and judged them over six months.
growth in AI brand presence for NewDay USA, tracked weekly across five engines over six months.See the case study.
Start the Clock at the Recrawl, Not the Release
The deploy date is your date. What matters for the test is the first time each crawler fetches the changed URL. Filter your server logs by user agent (Googlebot, Bingbot, OAI-SearchBot, PerplexityBot, ClaudeBot) and verify the requests against each operator’s published IP ranges. Record the first fetch after the deploy for each URL. For Google, the URL Inspection tool in Search Console also shows the last crawl date.
You can shorten the wait, within limits. Bing says it fetches a sitemap as soon as it is submitted, then rereads it at least once a day. It uses the lastmod value to decide which URLs to recrawl and which to skip. That only works if lastmod reflects a real content change. If you stamp the sitemap’s generation time on every URL, the signal stops meaning anything. Bing ignores changefreq and priority entirely. Bing also recommends IndexNow, which tells participating engines the moment a URL is added, updated or removed.
On Google, request indexing once per changed URL and then leave it alone. Google’s guidance says repeated requests for the same URL will not get it crawled any faster.
How Do You Build a Release Schedule Around Real Crawl Behaviour?
Start from the logs, not the content calendar. Over the past 30 days, work out how often each crawler fetched each URL, and group pages by that rhythm. Pages that crawlers visit every few days can support quick tests. Pages they reach once a month need the full window, or they need a crawl nudge before the test counts.
Then stagger the work so you can keep shipping without the windows overlapping:
- Week 0: ship one change type to cohort A. Leave a comparable holdout cohort untouched.
- Week 2: ship a different change to cohort B. Cohort B must not share pages or prompts with cohort A.
- Week 4: ship to cohort C, then read cohort A six weeks after its first verified recrawl, which may be later than week 6.
- Repeat on that rhythm. You release every two weeks, and every cohort still gets its own clean window.
Read every cohort against the same fixed prompt set, run several times per engine each week. Engines behave differently, so read each one separately. Perplexity cites its sources inline, which makes retrieval changes easier to see there. Gemini cross-references Search, YouTube and Scholar, so a page-level change can be outweighed by signals from elsewhere. Write the decision rule before you release, including the size of change in Share of Citation that will count as a result. Otherwise any movement will look like proof.
What to Do Before Your Next Release
Pull 30 days of server logs and calculate the median number of days between fetches for each URL and each AI crawler. That number replaces the guess you are currently planning around.
Next, audit your sitemap’s lastmod values so they change only when the content changes, and turn on IndexNow if your platform supports it. Freeze a prompt set for the pages you plan to test. Then choose one change type, two cohorts and a holdout, and put the six-week read dates on the calendar, counted from verified recrawl. Agree in writing that nobody rolls anything back before those dates unless the change broke something.
Use the results to decide your next test. A cohort that moved within two weeks tells you which pages can support a faster cadence. A cohort that showed nothing after a verified recrawl and six weekly readings is a real negative result. In both cases you have learned something, which the Monday-to-Friday test could not give you.
FAQs
How Long Does It Take for an AEO Change to Show Up in AI Answers?
Expect two to six weeks for changes that affect retrieval. The page has to be recrawled, reindexed and then selected for live prompts. Figures Google has shown put a typical refresh of a known URL at about 30 days, so the slower end of the range is common.
Should I Request Indexing After Every AEO Change?
Request it once per changed URL in Search Console, then wait. Google says repeated requests for the same URL do not speed up crawling. For Bing and other participating engines, use IndexNow and keep sitemap lastmod values accurate.
Can I Run AEO Tests on a Weekly Cadence?
You can measure weekly, but you should not judge weekly. Release to separate page cohorts every two weeks, and read each cohort six weeks after its first verified recrawl, so that no two changes overlap on the same pages.
Does the Two to Six Week Window Apply to Training Data?
No. The window covers retrieval presence, which depends on crawling and indexing. Training-data presence builds slowly through wide coverage over many months and needs a separate, longer measurement plan.