AI Citations Have Three Clocks — Hours, Weeks, and Quarters

How long until AI citations show up? There is no single clock, and anyone selling a universal “37 days” or “one week” is collapsing three speeds into a marketing number. Live answer engines can surface a newly indexed, quotable page in hours to a few days. Crawl, snippet eligibility, and competitive fan-out often stretch into weeks. Model-memory residue — what a system still “knows” without a fresh fetch — can lag for months. Google does not publish a citation SLA. John Mueller has said new pages can take several hours to several weeks to index, with no guarantee and no absolute timeline. Plan ranges. Then do the work that actually moves each clock.

This spoke sits under the Answer Engine Optimization playbook. For the Google-specific eligibility track, pair it with how to get cited in AI Overviews.

The short answer

• Retrieval-backed answers (ChatGPT with search, Google AI Overviews and AI Mode) can cite within hours once the URL is fetchable, indexed where required, and extractable.
• Indexation, snippet eligibility, and competitive fan-out usually need weeks of crawl plus rewrite cycles — not a Friday-to-Monday miracle.
• Training or long-term memory residue updates on a quarter-scale, not a sprint-scale. There is no public “forget the old founding year” ship date.
• Day 30 is for eligibility and extractability. Day 90 is for citation-rate movement you can defend in a meeting.
• A missing citation on day 37 is often lag — or a Clock-2 bug you have not checked. It is not proof the channel is dead.

Why is there no single citation clock?

Because “cited” is not one event. A retrieval engine can quote a page it just fetched. An AI Overview can only use a page that is already in Google’s index and eligible to appear in Search with a snippet. A chat product can still recite a stale brand fact from compressed memory even when a better page exists. Those are three systems. They do not share a calendar.

Vendor samples disagree because they measure different things.

Source	What they measured	Headline number	Why it is not your SLA
Profound, via Known & Cited	~900 new marketing pages; ChatGPT and Claude; ~60 days, Mar–May 2026	Median 6.81 days to first cite; 90th percentile 37.10 days	Narrow content type, two products, first-cite only
Minty Orange	95 already-live articles starting at zero cites; ChatGPT, Perplexity, AI Overviews; Apr–Jul 2026	46% earned a first cite; median wait 36 days; range 5–109 days	Small cohort, mostly travel/food, optimization-after-publish
Semrush AI Overviews study	10M+ keywords, Jan–Nov 2025	AIO trigger share swung 6.49% → 24.61% → 15.69%	Feature prevalence moved. Your “citation rate” can move with it

Treat those as samples. Your category density, how answer-shaped the page is, and whether Clock 2 is even clear matter more than their averages.

A first cite is also not the outcome most executives think they bought.

Event	What it proves	What it does not prove
First linked citation	The URL is fetchable and a passage was extractable for that prompt	You will keep the slot, or win the money query
Mention without a link	The brand string is in the answer	The engine will send a click, or that the fact is accurate
Repeat cite on the same definition prompt	Clock 1 is stable on an easy query	Category recommendation has moved
GSC generative AI impression	A Google AI feature showed a link	ChatGPT search cited you, or Clock 3 cleared
Competitor disappeared for a week	Their extract lost a sampling round	You displaced them as the default recommendation

If you put one vendor median on an OKR, you will fire a working channel on a schedule.

Clock 1: how fast can live retrieval cite a new page?

When an engine fetches the open web for a prompt, your page can appear as soon as it is crawlable, indexed where that engine requires it, and quotable. That is the fast clock. It is also the noisy one.

Surface	Typical first-cite window	What actually gates speed
ChatGPT (search / browsing)	Days to ~2 weeks once discoverable	Whether search runs for that prompt; index coverage; corroboration; passage quality
Google AI Overviews / AI Mode	Days to several weeks	Index + snippet eligibility + extractable passages + fan-out fit
Other retrieval chat products	Hours to ~1 week	Fetchability, a clear answer block, and how crowded the source set is

OpenAI launched ChatGPT search in October 2024 and opened it to everyone in regions where ChatGPT is available by February 2025. Search-mode answers can include links. Default chat without a fresh fetch cannot. If your panel logs “ChatGPT” without noting whether search ran, you are mixing Clock 1 with Clock 3.

Google’s generative features are not a separate index. They retrieve from core Search ranking systems (RAG / grounding) and may fan out into related sub-queries. A page that ranks for the head term and misses the fan-out questions will look “slow” when the real miss is coverage.

Log Clock 1 as a retrieval event, not as “we asked ChatGPT.” If you cannot say whether search ran, you cannot say which clock missed.

1. Freeze the prompt text. Do not rephrase it mid-week to “help it along.”
2. Note the product and mode (ChatGPT search on, AI Overview present or not, AI Mode thread).
3. Record mention, linked citation, citation URL, and whether the quoted passage matches your HTML.
4. If search did not run, tag the row Clock 3. Do not score it as a Clock-1 miss.
5. Re-run the same prompt 7 days later before you rewrite the page.

Panel mistake	What it does to the timeline
Mixing search-on and search-off chats	Inflates Clock-3 residue into a “citations are slow” story
Changing the prompt when you dislike the answer	You cannot tell lag from a different query
Scoring a mention as a citation	Mentions can move weeks before a linked source appears
Testing once on a laptop and shipping a deck	One-off chats are not a sample

Speed without substance is noise. A thin page that gets cited once and drops is not a win.

Clock 2: how long do crawl, index, and snippet eligibility take?

Most “we shipped Friday, why aren’t we cited Monday?” failures live here. Clock 1 cannot start if Clock 2 is blocked.

Google’s own ranges are already ranges. Mueller: several hours to several weeks to index, most good content often inside about a week, no guarantee it stays indexed or ranks. Search Console’s URL Inspection docs say requesting indexing typically takes a day or so, can take a week or two, and still does not guarantee the page enters the index. The Page indexing report says a new page or site can take a week or so before crawl even starts, and after discovery it can take up to a few weeks before Google crawls some or all of the site.

That is discovery and index — not citation.

Clock-2 stage	Honest range	What people confuse it with
Discovery (link, sitemap, or inspect)	Hours to a couple of weeks	“We hit Request indexing, so we are in”
Index selection	Hours to several weeks; not guaranteed	“Indexed” = “will be cited”
Snippet eligibility	Immediate once tags are clean; recrawl can take days to months	Preview-control changes apply on next successful crawl
AI Overview / AI Mode supporting link	Only after index + snippet eligibility	A special AEO markup track (there isn’t one)

Google is explicit: to appear as a supporting link in AI Overviews or AI Mode, the page must be indexed and eligible to be shown in Search with a snippet. There are no extra technical requirements. After you change preview controls, recrawl and processing can take several days to several months.

A sitemap helps discovery; it does not force crawl or index. If Clock 2 is red, no amount of “AEO content” moves Clock 1.

Request indexing is a queue ticket, not a citation. Use it as a scalpel.

Situation	Use URL Inspection / Request indexing?	Then wait
New revenue URL, Clock-2 checklist green	Yes, once	A day or so typical; up to a week or two; still not guaranteed
Same URL, still “Crawled – currently not indexed”	No — Google says there is no need to resubmit that status	Fix uniqueness, internal links, or quality; re-check the report
Template shipped nosnippet by accident	Fix the tag first, then request recrawl	Days to months for preview-control changes to clear
Twenty blog posts from Friday	No. Submit the sitemap; do not burn the daily inspect cap	Discovery on the sitemap’s schedule

If you are still clicking Request indexing every morning, you are managing anxiety, not a clock.

What actually gates Clock 2 on a live URL?

Run this before you rewrite tone.

• URL returns 200, not a soft-404
• Not noindex, not blocked to the relevant crawler in robots.txt or at the CDN
• Not nosnippet / max-snippet:0 — Google says those apply to AI Overviews and AI Mode and block the page as direct input
• Canonical points at the answer URL you want cited
• Sitemap includes the page; Search Console shows it indexed, not “Discovered – currently not indexed”
• Opening answer and key tables render in the HTML Google can use, not only after client JavaScript
• Site is still included in Search generative AI features in Search Console, if that control is available on the property
• Internal links from a crawled page reach the URL — orphan URLs wait longer

JavaScript is a common silent delay. Google crawls, then queues rendering, then indexes the rendered HTML. The render queue can be seconds or longer. If your 60-word answer block only exists after a client fetch, Clock 1 is waiting on a second pass you do not control. Server-render the answer unit.

If those checks fail, fix eligibility before you hire another writer.

Clock 3: how long does model-memory residue last?

Some answers still pull stale brand facts from older training or compressed memory even when a better page exists. That is why fixing a wrong founding year, a retired product name, or a competitor you no longer overlap with can take a long time to clear everywhere.

There is no public calendar for “the model forgot the old name.” Do not invent one.

Change type	Planning range	What actually moves it
New how-to page, retrieval engines	Hours–weeks once Clock 2 is clean	Fetch + extractable unit + some corroboration
AI Overview citation on a competitive query	Weeks; sometimes longer	Index + snippet + fan-out fit + competing extracts
Correcting a wrong “memory” fact	Weeks to quarters	On-site canonical fact + off-site repeats + time
Category recommendation displacement	Often a full quarter of corroboration	Other sources naming you for the same job

Clock 3 is why a CEO hears ChatGPT recite a 2023 competitor and assumes AEO “doesn’t work.” Retrieval may already be citing the new page on search-mode prompts while default chat still recites the old stack. Log those as different rows.

Do not promise that ChatGPT “will forget the competitor” in two sprints. Track accuracy separately from citation rate.

Clock-3 work is corroboration, not another homepage rewrite.

• One canonical fact block on-site (founding year, product names, what you do not sell)
• About, pricing, and product pages agree with each other — no leftover SKUs
• Directory and profile pages (the ones engines already fetch) match the same facts
• Two independent third-party pages state the corrected fact, not just your press room
• Old PDFs and “as featured in” pages that still publish the error are updated or noindexed
• Accuracy rows sit on the panel next to citation rows, with a date the error was last seen

Residue you are fighting	On-site fix	Off-site fix	Planning range
Wrong founding year	About page + JSON-LD foundingDate if you publish it	Wikipedia-class sources only if they are already wrong; else press and profiles	Weeks to a quarter
Retired product still recommended	Kill the URL or mark it discontinued in HTML	Update partner roundups that still list the SKU	Weeks; leftover roundups linger
Competitor named as “like you”	Clear positioning sentence on the offer URL	Stop repeating the comparison in your own guest posts	Often a quarter
Old price in default chat	Dated price block; remove undated “from $X”	Ask publishers who copied the number to refresh	Retrieval can update first; memory lags

If the wrong fact still lives on five directories, refreshing your hero line will not clear Clock 3. The residue is distributed.

Why does one page cite in a week and another take months?

Same brand, different clocks. The week-one cite is usually a low-competition definition. The months-long miss is usually a money query with ten roundups.

1. Query competitiveness — “what is X” with thin SERPs moves faster than “best X for Y.”
2. Extractability — a 60-word answer block plus a table beats a 2,000-word essay with no quotable unit.
3. Corroboration — engines prefer sources that other sources also name. A lone brand page is slower than a brand page plus two independent write-ups.
4. Fan-out — AI Overviews and AI Mode may issue related sub-queries. Your page can rank for the head term and miss the sub-questions.
5. Freshness vs uniqueness — a refresh of an already-trusted URL can beat a brand-new orphan URL.
6. Feature prevalence — if AI Overviews appear on fewer queries this month than last, your citation count can fall without your page getting worse. Semrush’s 2025 series showed that swing in public.

Fast-looking win	Slow-looking miss	What to tell the room
Definition page, week one	“Best [category] for [job]” still empty at day 60	Different query class. Do not average them.
Refresh of a trusted URL	New subdomain with no internal links	Clock 2, not “AI hates us.”
How-to with a table	Homepage manifesto	Extractability. Split the offer onto its own URL.
Mention without a link	Linked citation on a competitor	Mentions are Clock 1 adjacent. Citations need a passage.

If the week-one cite was a definition and the miss is a money query, that is normal — not a mystery.

What should you expect by day 30, 60, and 90?

Use horizons as checkpoints, not promises. The dates are when you inspect the system. They are not when the engine owes you a cite.

Horizon	Healthy signals	Panic signals (investigate)
Day 30	Indexed; snippet-eligible; prompt panel logged; 1–2 soft cites on easy prompts	Still Discovered – not indexed; nosnippet; zero crawl; answer unit only in JS
Day 60	Rising mention rate; a few citations on how-to / definition prompts	Mentions without any citations and no extractability fixes shipped
Day 90	Citation rate moving on priority prompts; fewer accuracy errors on Clock-3 facts	Still invisible on every engine with clean eligibility — then audit the strategy

Day 30 is an operations checkpoint. Day 90 is a results checkpoint. Mixing them up is how teams declare AEO “dead” at day 37.

Profound’s 90th percentile at 37.10 days is a first-cite statistic on a marketing-page sample in two chat products. Minty Orange’s median of 36 days is a first-cite statistic on 44 of 95 optimized articles. Neither is a Google AI Overview inclusion date. Neither is a category-recommendation date. Report them as samples if you report them at all.

A worked 90-day read, without inventing a booked cite:

Week	What you inspect	What you may report	What you may not report
0	Panel frozen; Clock-2 checklist on five URLs	“Baseline captured. Eligibility work started.”	“We will be cited by day 21.”
4	Index + snippet + first easy-prompt cites	“3/5 URLs indexed and snippet-eligible. 1 soft cite on a definition prompt.”	“AEO is working” from one cite
8	Mentions vs linked citations; extractability diffs	“Mentions up on how-tos. Money queries still empty. Two passages rewritten.”	“We are behind the 6.81-day median.”
12	Citation rate on the priority set; Clock-3 accuracy	“Citation rate moved on 4 of 12 priority prompts. Two stale facts remain in default chat.”	“Q2 AI Overview inclusion is done.”

If week 4 is still “Discovered – not indexed,” you do not have a content problem yet. You have a Clock-2 problem. Keep the slide that honest.

What actually moves the clock?

Not volume. Not a new AEO SaaS login. Not an llms.txt file — Google says it does not use special AI text files or extra markup for generative Search features.

Lever	Clock it moves	How you know it worked
200 + index + snippet-eligible HTML	2, then 1	URL Inspection + live SERP snippet
Server-rendered answer block + table	1	Panel shows a quote that matches your passage
Internal links + accurate sitemap lastmod	2	Discovery lag shrinks on new URLs
Off-site corroboration of the same fact	1 and 3	Independent sources start repeating your wording
Frozen 25–40 prompt panel, same engines	Measurement	You can tell lag from a drop
Refresh of a trusted URL when facts change	1 and 2	Recrawl, then a cite you used to win returns
Buying “AI citation” directories	None	You wasted a sprint

Do this in order:

1. Clear Clock-2 blockers on the five revenue URLs.
2. Add one quotable answer and one table to each.
3. Log a prompt panel before you publish more.
4. Re-test the same prompts on a fixed cadence.
5. Only then spend words on net-new URLs.

Semrush helps with SERP and competitor context around the prompts. It does not replace the multi-engine panel, and it does not replace Search Console’s generative AI performance report for Google impressions.

Run a quote test before you call a page “AEO-ready.” If an engine cannot lift a passage, Clock 1 has nothing to fetch.

1. Open the live URL in a logged-out browser. Copy the first 80–120 words that answer the prompt.
2. View source or use URL Inspection’s rendered HTML. Confirm those words exist without waiting on a client fetch.
3. Check that a table or numbered list sits next to the claim you want quoted.
4. Search the exact claim in a retrieval product. If a competitor is quoted, diff their passage against yours — length, specificity, and whether they name the constraint you buried.
5. Ship one tighter passage. Re-test the same prompt in 7 days. Do not ship six new posts in the meantime.

Quote-test result	Clock	Fix
Passage missing from rendered HTML	2	Server-render the answer; kill the app-shell gap
Passage present, never quoted, competitors are	1	Sharpen the unit; add the missing constraint or number
Quoted once, gone next week	1	Defend the passage; check whether the query cooled
Quoted in search mode, wrong in default chat	3	Corroborate the fact off-site; keep measuring accuracy

GSC’s generative report is impressions in AI Overviews and AI Mode, rolled out as a dedicated view in June 2026. It is not a citation-rate dashboard for ChatGPT. Do not brief it as “our AEO number.” Brief it as Google-surface visibility while the panel covers the rest.

When is a missing citation a bug vs normal lag?

Treat it as a pipeline bug when:

• The page is not indexed, is blocked, or returns a soft-404
• Snippet controls prevent quotation (nosnippet, max-snippet:0)
• The answer is buried under hero fluff with no standalone passage
• The answer unit exists only after client JavaScript
• Competitors are cited with near-identical claims you never published clearly
• Search Console still shows the URL as discovered-not-indexed after several weeks on an established site

Treat it as normal lag when:

• Eligibility is clean and the page is new (under 2–4 weeks)
• The query is crowded with strong roundups you have not countered
• Mentions are rising even if linked citations are not yet
• AI Overviews are not even triggering on that query this week
• You are judging a money query against a definition-page sample

Evidence you have	Call it	Next action
noindex / nosnippet / blocked crawl	Bug	Remove the block; wait for recrawl (days–months)
Indexed, snippet live, page <14 days old	Lag	Re-test weekly; do not rewrite the strategy
Indexed, no answer unit, competitors quoted	Bug (extractability)	Ship the passage and table this week
Clean page, competitive “best of” query, day 40	Lag	Counter the roundup; do not pause the channel
Search-mode cite, default-chat stale fact	Clock 3	Fix on-site + off-site facts; measure accuracy

Re-test on a fixed panel weekly. One-off chats in a browser are not a timeline.

How do you brief leadership without inventing a date?

Executives want a date. Give them a range per clock instead of a fake day count. I have been SEO-certified since 2021 and I still will not put “AI Overview by Q2” on a slide as a booked outcome. Inclusion is earned. Eligibility work is scheduled.

Stakeholder ask	Honest reply
“When will ChatGPT recommend us?”	“Retrieval cites can start in weeks if eligibility is clean; category recommendations often need a quarter of corroboration.”
“Why did competitor X show up in five days?”	“Usually a low-competition prompt or an already-trusted URL — not proof their agency owns a faster API.”
“Can we guarantee AI Overview inclusion by Q2?”	“No. We can guarantee eligibility work, extractability rewrites, and a measurement ritual. Inclusion is not a line item.”
“Is day 37 a fail?”	“Only if Clock 2 is still broken. Otherwise it is early for Clock 1 on hard queries and irrelevant for Clock 3.”
“Which vendor number do we use?”	“Neither as a target. Use 6.81 and 36 as samples that prove the spread. We report three clocks.”

Put the three-clock table in the deck. Remove the single OKR date. Teams that keep one number will keep declaring AEO dead on a schedule.

A slide outline that survives a board meeting:

1. Title: “Three clocks, not one date” — hours–days / weeks / quarters.
2. Eligibility: five URLs, Clock-2 status, date last inspected.
3. Panel: 25–40 prompts, engines, last run date. No live demo from someone’s laptop.
4. Movement: mentions vs linked citations vs accuracy errors, split by query class.
5. Ask: the next Clock-2 fix or extractability rewrite — not budget for a citation guarantee.

Slide anti-pattern	What it trains the room to do
A single “days to cite” KPI	Fire the channel when a sample median is missed
Screenshot of one lucky chat	Confuse a bonus with a system
Competitor cited in 5 days, no query class	Assume they bought a faster pipe
“Guarantee AI Overview by Q2”	You will eat that sentence in the Q3 retro

If they still want a number, give the range and the work: “Retrieval can start in weeks on clean URLs; we will not call the program successful or failed before day 90 on the money prompts.” That is a management decision, not a physics constant.

Should you pause publishing if nothing moved in 37 days?

No — unless eligibility is broken. Pausing “because a vendor median said 36 days” is cargo-cult measurement.

Minty Orange’s own spread ran 5 to 109 days, with 80% of the eventual cites inside two months. Profound’s 90th percentile was 37.10 days on a different sample. Those two sentences cannot both be a universal fail line. They can both be true as samples.

Keep shipping answer-shaped updates on the pages that already pass the fetch checks. Pause volume only when you have no measurement ritual. Then the fix is a dated baseline, not silence. The visibility lane is built for that: /visibility.

If this is true	Do this	Do not do this
Clock 2 still red	Stop net-new posts; fix crawl/index/snippet	Brief that “AEO failed”
Clock 2 clean, day 37, hard queries	Keep improving the five URLs; re-measure at 60/90	Pause because 36 or 37 appeared in a blog post
No prompt panel exists	Freeze 25–40 prompts this week	Ship twelve more undifferentiated posts
Cites appear then vanish	Defend the winning passage; log query type	Assume the channel is random and quit

What should you not optimize while you wait?

While clocks run, do not burn the sprint on vanity work that does not move eligibility or extractability.

1. Mass blog posts with no answer unit
2. Buying “AI citation” directories that look like link farms
3. Rewriting brand voice into identical FAQ sludge on every URL
4. Chasing every new AEO dashboard before you have a prompt panel
5. Blocking training bots and assuming that alone explains missing Search cites
6. Special AI files, forced chunking, or schema invented only for generative Search — Google lists those as things you can ignore

Waiting is not the same as idling. Ship Clock-2 fixes and quote tests. Skip the costume changes.

What re-test cadence matches the three clocks?

Change the panel only when the business changes. Moving the goalposts every week is how you fake progress.

Cadence	What you run	Why
Weekly	10–15 money prompts across 2–3 engines; note whether search/retrieval ran	Catch retrieval wins and losses early
Biweekly	Index + snippet eligibility on top URLs in Search Console	Catch template regressions (nosnippet in a new layout)
Monthly	Full 25–40 panel + accuracy review; export GSC generative AI impressions	Trend citation rate without trusting a single chat
Quarterly	Memory/accuracy deep dive + off-site facts	Clock-3 residue and PR gaps

Log each row with engine, date, mention, citation URL, accuracy flag, and clock tag (retrieval / eligibility / memory). If you cannot say which clock a miss belongs to, you are not measuring yet.

Bad measurement creates fake clocks. Kill these before you add another tool.

Failure mode	What it costs	What you do instead
New prompts every week	You cannot tell improvement from a easier question	Freeze the panel; add prompts only when the offer changes
One engine, one intern, Fridays	Variance looks like strategy	Two or three engines, same operator, same day
Dashboard without Clock tags	Every miss becomes a content rewrite	Tag retrieval / eligibility / memory on the row
Counting impressions as citations	GSC gen-AI impressions are not ChatGPT links	Keep Google and chat panels on separate slides
“We asked it once and it knew us”	Anecdote becomes the QBR	Require two consecutive panel runs before you brief a win

A 30-day kit that matches this cadence:

• Freeze a 25–40 prompt panel before more publishing
• Baseline citation / mention / accuracy across ChatGPT search, AI Overviews, and one other retrieval surface
• Clear Clock-2 blockers on the five revenue pages
• Add one quotable answer block + one table to each
• Schedule week-4 and week-8 re-runs (same prompts, same engines)
• Brief leadership on three clocks — not one OKR date

How do seasonal and multi-product pages distort the timeline?

If your cite depended on a trending news hook, disappearance in two weeks can be normal — the query cooled, not your AEO. Log query type (evergreen vs news) next to each prompt. Evergreen how-to and definition prompts are the timeline you manage. Newsjack cites are bonuses you do not put on the OKR.

If one SKU cites in a week and the flagship offer takes a quarter, check whether the fast win was a definition page with thin competition while the flagship sits behind a vague homepage. Split clocks by URL and offer, not by brand average. A blended “citations in 30 days” KPI hides the page that actually funds payroll.

Distortion	What it looks like	How you correct the report
Newsjack	Cite in 48 hours, gone in 14 days	Tag as news. Do not use it as the Clock-1 benchmark.
Seasonality	Cite rate drops in the off-month	Compare to the same month last year, or to query volume
Multi-SKU blend	“We get cited in a week”	Split the dashboard by URL. Keep the slow flagship visible.
Feature swing	AIO cites fall while eligibility is clean	Check whether the Overview still triggers on that query
Template regression	Every URL loses snippets the same week	Diff the head tags, not the copy

Do not average a newsjack win with a flagship miss and call it “the AEO timeline.” That number is a costume.

FAQ

Can citations appear in under a week?

Yes — especially on low-competition how-to prompts when the URL is already indexed and the opening answer is extractable. Competitive recommendation prompts and Google AI Overviews rarely move that fast. Under-a-week cites are a bonus, not the plan, and they are not proof the next URL will repeat them.

Why do citations disappear after they appear?

Engines re-sample sources as competitors refresh, SERPs shift, feature prevalence changes, or your passage stops being the cleanest extract. Treat citations as rental, not ownership. Re-run the same panel monthly and defend the passages that won instead of publishing a cousin URL.

Does freshness change the timeline?

It can shorten Clock 1 and Clock 2 for trusted URLs when you fix facts and answer blocks. It does not instantly rewrite model memory. Pair refreshes with off-site corroboration when the error is a brand fact, and refresh when facts, prices, or steps change — not on a costume cadence.

How long for model-memory / training residue to update?

Often weeks to quarters, and not uniformly across products. There is no public ship date. Correct the canonical facts on-site, align directories and press, and keep measuring accuracy. Do not promise a hard date for “ChatGPT forgot the old name.”

When is a missing citation a pipeline bug vs normal lag?

Bug if crawl, index, snippet eligibility, or extractability is broken. Lag if those are clean, the page is new, and the query is competitive. Log evidence either way so you are not arguing from vibes, and tag the miss to a clock before you rewrite the strategy.

Should I pause content if nothing moved in 37 days?

Do not pause because a vendor median was 36 days or a 90th percentile was 37. Pause net-new volume only if you lack a prompt panel or still have eligibility failures. Otherwise keep improving the five pages that should win, and re-measure at day 60 and 90.

CTA

Stop buying a single timeline. Instrument three clocks, then decide what to fix first.

Lane overview: /visibility. Book a visibility audit if you want a dated baseline and a 30/60/90 that matches how engines actually update.