The short answer
A live-API server sends your question to Google at the moment you ask, once per property or more, and spends that property's quota. A stored-history server imported the data earlier and reads its own copy. Live wins for the last hour and odd one-off reports. Storing wins when one question spans many sites and months.
- Count upstream requests per question before you pick a design, not after.
- Ask any server how it reports a site it could not fetch or a source that stopped syncing.
- Running both designs side by side in one agent is normal and often right.
Following one 30-site question through each design
Take one real question and trace it. You manage a 30-site roster, each site with its own GA4 property, and you want the year's trend.
Ask your agentFor all 30 websites, compare sessions for each of the last 12 months and list the five with the steepest decline.
Tools it calls list_websites get_ga4_overview
Under a live design, the server holds no data. The agent lists the properties it can reach, then the server sends a report request per property to the GA4 Data API while you wait: one per property with a month dimension, twelve if the agent asks month by month. Each draws on that property's token budget.
Under a stored design, the requests to Google happened earlier, when each site's sources were linked and imported. At question time the agent still calls a tool once per site, because DuckDive has no single portfolio endpoint. But each call reads DuckDive's copy and does not wait on a live fetch from Google.
| Step | Stored (DuckDive) | Live API server |
|---|---|---|
| Find the sites | 1 call to list_websites | 1 call to list accounts and properties |
| Data tool calls by the agent | 30, one get_ga4_overview per site | 30 with a month dimension, 360 month by month |
| Requests sent to Google while you wait | 0 | 30 to 360 |
| GA4 quota buckets drawn at question time | None | 30 properties' hourly and daily buckets |
| Rows handed back to the agent | About 365 daily rows per site, near 11,000 in all | 12 monthly rows per site, 360 in all |
| A follow-up on the same window | Reads the copy again or reuses rows in context | Another round of Google requests unless reused |
Note the row that favours live. Google's report API can group by month on its side and return 12 tidy rows. DuckDive's overview tool returns sessions per day, so the agent gets far more rows and sums the months itself. On a long sweep, tell the agent to reduce each site to monthly totals as it goes, or run the loop in a script from Claude Code so the daily rows never fill the chat.
For all 30 websites, compare sessions for each of the last 12 months and list the five with the steepest decline.
list_websitesreturns the 30 sites on the roster with their ids and linked sourcesget_ga4_overviewrun 30 times, once per site, for the last 12 months of daily sessions from stored history
I grouped each site's daily sessions into calendar months and dropped the partial months at both ends of the window. I then compared the average of the first three full months with the last three. These five fell furthest.
| Roster site | Avg sessions per month, first 3 months to last 3 | Change |
|---|---|---|
| Site 22 | 18,400 → 11,960 | -35% |
| Site 19 | 7,500 → 5,025 | -33% |
| Site 07 | 9,200 → 6,440 | -30% |
| Site 15 | 26,500 → 19,610 | -26% |
| Site 03 | 4,800 → 3,600 | -25% |
Site 28 just missed the list at -23%. All 30 sites returned data. Site 22 slid steadily rather than off a cliff, so a broken tag is unlikely.
Trace your own roster instead of a made-up one
Link your sites once, then ask the 12-month decline question and count the Google requests it skips.
Free for 2 sites · No credit card
Where per-property quotas come in
The call counts above only matter because Google rations them, and it rations them per property. As published on Google's Data API quotas page (checked November 2026), a standard GA4 property has 200,000 core tokens a day and 40,000 an hour, of which any one Cloud project may spend 14,000 per hour. The page lists Analytics 360 properties in a separate column with ten times the daily and hourly token figures.
There is no fixed price per request. In Google's words, "the number of tokens charged depends on the request's complexity", and the drivers it lists include date range length, number of rows, and the dimensions and metrics asked for. So the expensive part of our sample question is the part that makes it useful: twelve months, broken out by month. A one-week total for the same site costs less.
Run the roster through those rules and two details decide whether the sweep finishes cleanly:
- Burst size per property. The same page caps a standard property at 10 concurrent core requests. An agent that sends all twelve month-by-month requests for one site at once has asked for more than that property allows in flight. Asking for one request with a month dimension, or pacing the twelve, stays inside it.
- Who else is drinking from the bucket. Thirty properties are thirty separate buckets, so your sweep alone is unlikely to drain any of them. But every request against a property counts, whoever sends it. A Looker Studio report refreshing all morning, a nightly export and your agent's follow-ups all draw on one pool, and once "any quota bucket reaches 0 remaining", Google answers that property with
429 RESOURCE_EXHAUSTEDuntil it refills.
Search Console is rationed per site and per user instead of in tokens. Google's Search Console API usage limits page lists 1,200 queries per minute for each, plus load quotas counted in 10-minute and 1-day chunks. Sixty requests will not trouble the per-minute figure; repeated 12-month pulls are what the load quotas are for. Google's own advice on that page is to "Avoid requerying the same data", and the example it gives, pulling last month's data over and over, describes an agent's third and fourth follow-up on a live server. Not asking Google twice is the job a stored copy does.
Where the quota goes
A stored design does not escape quotas. It spends them during import and sync instead of at question time. The difference is that a quota error at import shows up as a failing sync you can see, not as a hole in an answer you are about to forward.
Freshness: the case for live calls
Live calls return whatever Google has processed at that moment, with no sync schedule in between. For some questions, that settles it.
Ask your agentWhat happened on my site in the last hour?
Tools it calls run_realtime_report on Google's Analytics MCP server, not DuckDive
DuckDive cannot answer this one. It has no realtime data, and its shortest range is the last 7 days. An agent connected only to DuckDive should say so rather than guess. Google's own Analytics MCP server has a run_realtime_report tool for exactly this.
Even live has a limit here, and few people check it. Google's realtime reporting guide says a realtime query covers "the present moment to 30 minutes ago (up to 60 minutes for Google Analytics 360 properties)". On a standard property, "the last hour" is really the last half hour.
Yesterday is where live looks fresher than it is. Per the GA4 data freshness article, Google may need 24 to 48 hours to finish processing a day, and the figures can shift while it does. On the search side, the Search Console Performance report help warns that the most recent days are preliminary. Both designs start out holding draft numbers for those days. The difference: a live server picks up the revised figure on its next call, while a copy keeps the draft until something re-pulls it.
That lag is the stored design's weak spot, so judge any such server by how it catches up. DuckDive gives you a lever and a gauge. The lever, refresh_website_data, re-pulls a span of recent days you choose (anywhere from 1 to 30) for a single source or for every source on a site, and it accepts one request per site per quarter hour. The gauge, get_sync_status
Ask your agentRe-import the last 3 days of GA4 for the sites you just flagged as stale, wait until their sync shows ready, then rerun the 12-month comparison.
Tools it calls refresh_website_data get_sync_status get_ga4_overview
The re-import is an action, so it needs an admin account and a connection that kept "Also allow actions".
History depth: the case for storing
A live server can only show what the source still keeps. Search Console is the hard case, because its window is fixed by Google rather than by your plan. When Google launched the new Search Console in January 2018, its announcement said "you'll have 16 months of data" in the Performance report. That is the figure as of the 2018 post; the current Performance report help page we checked does not restate a number, so before you plan a year-over-year study, have the agent request the oldest date range Google will serve for one of your properties. Whatever the window is on the day you ask, days that have aged out of it are gone for every tool that queries Google live. Only a copy made before they aged out can still show them.
Twelve months is still enough for the questions that hurt most under live calls, such as finding each site's low point across the year.
Ask your agentSplit the past 12 months into quarters and tell me each site's worst one.
Tools it calls list_websites get_ga4_overview get_search_console_summary
Split the past 12 months into quarters and tell me each site's worst one for sessions. Did search clicks bottom out in that same quarter?
list_websitesthe same 30 sitesget_ga4_overview30 reads with the last12months preset, each site's daily sessions bucketed by quarterget_search_console_summaryonce per site, last 12 months, weekly click points summed into the same four quarters (a week that straddles two quarters is counted in the one it starts in)
I split the window into four quarters, oldest first, and found the quarter with the fewest sessions for each site. Most sites dipped in the third quarter, which looks seasonal. The seven whose low point is the most recent quarter matter more.
| Quarter | Sites at their session low | Clicks also lowest |
|---|---|---|
| Q1 (oldest) | 6 | 5 |
| Q2 | 4 | 3 |
| Q3 | 13 | 11 |
| Q4 (latest) | 7 | 2 |
| All sites | 30 | 21 |
Of the seven sites at their low point now, only two lost search clicks in step. For the other five, organic search held up, so the drop came from another channel. I can break those five down by channel next.
That answer took 61 tool calls: one list and two reads per site. Under a live design, each of the 60 reads becomes a Google request. The fields behind the clicks column, and the weekly points they arrive in, are listed on the Search Console data source page.
What vendors say about their own design
Few products are purely one design, and the ones below say so themselves. This table is narrowed to what each vendor publishes about the two things this page has traced: where a question goes, and what happens when the source does not answer.
| Product | Design | What its own pages say |
|---|---|---|
| Supermetrics | Live for the MCP; optional storage elsewhere | MCP: "If a query fails or data is unavailable, the server tells you". Pricing page: "You can choose to store all of your data, some of your data or none at all." |
| Google Analytics MCP | Live | Runs core and realtime reports against the Data API on the property you name |
| Windsor.ai | Live, with an optional data store | Holds a query's response for 6 hours; also keeps a "Cache Data Store" for "large or long-range datasets" |
| Funnel | Stored | Starter plan: "Connect, store and prepare your marketing data." |
| DuckDive | Stored | Pulls each source's back history at link time; agent tools query that copy, and get_sync_status dates the newest day held per source |
As published on each vendor's page, checked November 2026: Supermetrics MCP page, Supermetrics pricing, Google Analytics MCP repository, Windsor.ai cache documentation, Funnel pricing.
Supermetrics deserves credit on the failure question. Its MCP page promises the loud failure this page keeps asking for: when a query fails, the server says so rather than filling the gap. That is what makes the four missing sites in the next section visible, provided the agent passes the message on instead of ranking around it. The "live" label also covers only its MCP. On its pricing page, Supermetrics sells optional storage to "protect your reports from unexpected API outages", which is the stored design's argument made by a vendor whose agent connector calls live.
Windsor.ai publishes both halves of a hybrid. Its documentation says "Windsor temporarily stores the response for 6 hours for that exact Connector URL", so an identical query inside that window does not go back to the source, and a manual clear is allowed "only once every two hours". Separately, it describes the Cache Data Store as "an internal Windsor.ai database that stores large or long-range datasets". For a 12-month question across a whole roster, that data store is the part to ask Windsor about before you count its requests as live.
A design label is not a verdict on fit. Supermetrics reaches paid social and CRM sources DuckDive does not read, and so does much of the connector field; our Supermetrics comparison weighs that trade. Whatagraph, Porter and the rest of the hosted field are sorted by design in the marketing data MCP server roundup.
Failure modes you will meet with each
Every design fails. What matters is whether the failure is loud or silent. Here is the live design's most common one, on the same roster.
| Rank | First answer (26 sites) | After retrying the 4 |
|---|---|---|
| 1 | Site 22, -35% | Site 22, -35% |
| 2 | Site 07, -30% | Site 19, -33% |
| 3 | Site 15, -26% | Site 07, -30% |
| 4 | Site 03, -25% | Site 15, -26% |
| 5 | Site 28, -23% | Site 03, -25% |
In this scenario, sites 05, 11, 19 and 24 came back with 429 RESOURCE_EXHAUSTED because a dashboard had drained their hourly buckets. A careful agent opens with "26 of 30 sites returned data". A careless one ranks what it got: Site 19, the second steepest decline, vanishes, Site 28 takes its place, and nothing looks wrong.
Live design: what goes wrong
- Partial results ranked as if complete, as above.
- Hourly or daily token buckets drained by other tools reading the same property.
- Long date ranges costing more tokens per request, so the expensive questions fail first.
- Follow-ups that refetch the same window instead of reusing it.
Stored design: what goes wrong
- A source whose sync is failing, so one site's numbers stop at an old date while the rest move on.
- Recent days copied before Google finished processing them.
- Questions the copy cannot shape: in DuckDive, fixed ranges only, no device or country split, and at most 100 rows per call.
- Large raw payloads, such as daily rows for every site, that the agent must sum itself.
Both lists share a cure: make the agent state its coverage before its conclusion.
Ask your agentBefore you rank anything, check sync status on every website and list any source that is failing, still importing, or whose newest data is more than two days old.
Tools it calls list_websites get_sync_status
For a full routine on checking an agent's numbers against the source, see the guide to verifying your agent's marketing numbers.
A decision guide by type of question
Pick by the questions you will actually ask. Most teams end up with a mix.
| Type of question | Stored history (DuckDive) | Live API calls |
|---|---|---|
| The last 30 minutes on one site | No, no realtime data | Yes, realtime report |
| One property, custom dimension or funnel | No, fixed metrics only | Yes, full report options |
| Exact dates, such as a launch week | No, fixed ranges | Yes, any date range |
| Twelve months across 30 sites | Yes, 30 reads of the copy | Yes, at 30 to 360 Google requests per ask |
| Weekly check across every site | Yes, no quota spent asking | Yes, but it shares quota with dashboards |
| GA4 and Search Console side by side | Yes, same website id | Yes, often across two servers |
| Search Console days Google has already dropped | No, agent presets top out at last12months | No, Google no longer serves them |
Sources: Google Analytics MCP repository for realtime and funnel tools; Google's 2018 Search Console launch post for the Performance report window; checked November 2026.
Read the table by row, not by column count. Live calls can answer every multi-site row; the price is paid in requests, quota and the chance of a partial set. The stored copy loses every row that needs today, exact dates or a custom report. Its sweet spot is a question you will ask again next month.
Ask your agentFor every site, compare average Search Console position in the first and second halves of the last 6 months, and count how many improved and how many slipped.
Tools it calls list_websites get_search_console_summary
Questions to put to any MCP server before you commit
- Where does a question go? To the source each time, to a cache, or to a stored copy.
- How many upstream requests per site? One per property, one per month, or none at question time.
- Whose quota is spent? Your property's buckets, the vendor's project, or neither while you ask.
- How are missing sites reported? Look for a count of sites that answered.
- How fresh is the copy? Look for a per-source date and a way to force a refresh.
- What shape comes back? Daily rows, monthly rows or totals, and the row cap per call.
Where DuckDive is the wrong choice
Picture a week where nearly every question concerns a single GA4 property, today's traffic, and a custom dimension or a funnel step. None of that gains anything from a stored copy, so a live server is the better tool. Google's Analytics MCP is the natural pick, and we set it beside DuckDive in a separate head-to-head. The same applies when a roster's spend sits in Meta, LinkedIn or a CRM: DuckDive's copy holds four Google sources (GA4, Search Console, Google Ads, Local Services Ads) plus Yext and Local Falcon, and nothing beyond them.
Also skip it if you need campaign or keyword detail from Google Ads (DuckDive returns the account summary), exact custom date ranges, or more than the last 12 months through an agent today. And if you need minute-by-minute monitoring, no stored design is the right tool.
Running the 30-site question on DuckDive
To test the stored design on your own roster, the one address you need is https://api.duckdive.ai/mcp. Clients reach it over streamable HTTP and authenticate through OAuth, so there is no key to paste anywhere. DuckDive files everything by website: a site owns its linked sources and its imported history, and each data tool is pointed at one site id per call, which is why the sweep above loops. The DuckDive overview covers the product; below is the order of operations for loading a roster.
-
Register the server in your client
For Claude Code that is a single command (
claude mcp add --transport http duckdive https://api.duckdive.ai/mcp), with sign-in waiting in the/mcpmenu. Chat apps want the address pasted into their list of remote servers, under whatever label that app and version gives it. -
Sign in, and decide on actions
A DuckDive page opens, where Continue with Google either signs you in or creates the account. Next comes an approval screen with an "Also allow actions" box. That box separates an agent that can only read from one that can register sites and link sources; loading a roster needs the second kind, so leave it ticked, and reconnect without it later for a read-only agent. The client is issued a DuckDive token, and your Google password never passes through it.
-
Hand the roster to the agent
Setup prompt for your agentSet up DuckDive for my website. Follow https://analytics.duckdive.ai/
agents.md -
Let the copy fill, then audit it
This is the step a live server never has. The moment a source is linked, DuckDive starts copying its back history: minutes for most sites, longer for ones with many years on record. Before any sweep, have the agent poll
get_sync_statusand refuse to rank until every source on every site is marked ready. -
Run the year-long sweep
Paste the decline prompt from the top of the page, and tell the agent to open its reply with how many of the 30 sites came back.
Writing the loop in code instead of chat? Scopes and each tool's inputs and outputs are on the developer page. If the roster belongs to clients, the agency page is the better starting point.
Architecture questions people ask
Does a stored-history server still call Google?
Yes, at link time and on each sync after that. What changes is timing: your question reads the stored copy, so it does not wait on a live fetch or spend the property's quota while you ask.
Can DuckDive show what is happening on a site right now?
No. Everything it returns comes from its imported copy, and the narrowest preset is a week. For the past half hour, point a live tool such as Google's Analytics MCP server at that property. If the copy seems to lag, an admin's agent can ask DuckDive to re-pull the most recent days.
Is there one call that returns all 30 sites at once?
No. Every DuckDive data tool reads one website, so the agent lists your sites and calls the tool once per site. Each of those calls reads stored data rather than sending a request to Google, which is what keeps a 30-site question practical.
Can one agent use a live server and DuckDive together?
Yes. MCP clients can hold several servers, and the agent picks tools by their descriptions. A common split: Google's Analytics server for realtime and funnels on one property, DuckDive for questions spanning many sites or sources.
What happens when a live call hits the GA4 quota?
Google returns a 429 RESOURCE_EXHAUSTED error for that property until its bucket refills. Your agent then holds a partial set of sites, so ask it to name every failed site and retry before it ranks anything.
What is the oldest day my agent can reach in DuckDive?
Twelve months back is the ceiling through the tools today. Ranges are presets (a week, two weeks, 30 days, 90 days, half a year or a year), never a start and end date you type. The copy itself begins with whatever Google still served on the day you linked the source; Google's 2018 Search Console launch post put the Performance report at 16 months of data.
Who can reach the stored copy, how long does it last, and how do I revoke access?
Reach follows your login: the agent gets the same websites your DuckDive account does, under your role, and nothing in the connector can delete data. Retention is set out in DuckDive's privacy policy, which keeps analytics data "for as long as your account is active or as needed to provide our services" and clears personal data within 30 days once an account is deleted, legal holds aside. Revoking the connection, from either DuckDive or the AI app, ends access at once. Tool results that reach your AI provider fall under that provider's terms, not DuckDive's.
Sources
- Limits and Quotas on API Requests (Google Analytics Data API docs): token buckets per property, the 10-request concurrency cap and the 429 behaviour. Read November 16, 2026.
- Usage Limits (Search Console API docs): the 1,200 queries per minute figures, load quotas and the advice against requerying. Read November 16, 2026.
- The google-analytics-mcp repository on GitHub: the live server's realtime and core report tools. Read November 16, 2026.
- Realtime Report Basics (Google Analytics Data API docs): the 30 minute realtime window, 60 for 360 properties. Read November 16, 2026.
- Data freshness (Google Analytics help): the 24 to 48 hour processing delay. Read November 16, 2026.
- Search Console's overview of the Performance report: why the latest days count as provisional. Read November 16, 2026.
- Google's January 8, 2018 announcement of the new Search Console: the 16 months of Performance data quoted above. Read November 16, 2026.
- Supermetrics MCP product page: its statement on failed queries. Read November 16, 2026.
- Supermetrics pricing, storage section: optional storage and the outage protection line. Read November 16, 2026.
- Windsor.ai documentation on its Cache Data Store and Query Cache: the 6 hour response cache and the two hour clearing limit. Read November 16, 2026.
- Funnel's plans page: the Starter plan's "store and prepare" wording. Read November 16, 2026.
- DuckDive app privacy policy (last updated October 2026): the retention wording in the FAQ.
Last reviewed . Numbers in EXAMPLE DATA blocks are fictional.
