The short answer
To verify an AI-generated marketing report, run five checks before its numbers leave your hands. Confirm every source finished syncing. Make the agent paste the raw tool output. Recompute every percentage, rate and difference from that output. Trace each sentence back to a tool call. Then freeze the prompts, so next month's run asks the identical questions and a changed number means changed data.
- Copy counts straight from the tool table; redo anything divided, summed or subtracted
- Ask how many rows came back before you believe "it isn't on the list"
- Cut any sentence the agent cannot tie to a tool call
Five ways agents get marketing numbers wrong
Fund analysts, group marketing leads and agency account leads all get tidy tables fast. When the agent goes wrong, it is rarely a figure from thin air. It is a real number read over the wrong window or pushed through the wrong arithmetic. Each failure below comes from how DuckDive hands data to an agent: fixed windows, ranked lists with a row limit, and tools that return either a line over time or one total per window.
1. The wrong window
You write "October". DuckDive accepts only six preset spans, passed as last7days, last14days, last30days, last90days, last6months and last12months, and not one of them lines up with a calendar month. Run in mid-November, last30days straddles two months, and "this quarter" quietly turns into last90days, sometimes without a word from the agent. Write those parameter names into your prompts and the swap cannot happen silently.
GA4 offers a partial way back. get_ga4_overview returns a row per day, so an agent can add up October's rows out of a longer span; ask it to name the first and last date it added and how many rows that was (31, for October). Search Console is coarser. Its summary arrives as weekly points dated to the Monday that starts each week, so a week straddling two months counts wholly in one, and any October click total built from it is an estimate.
The newest days carry a second risk. Google's GA4 data freshness page says "Data processing can take 24-48 hours. During that time, data in your reports may change" (as of November 2026). Yesterday's figures can still move.
2. Averaging percentages
Nine companies, nine percentage changes. The agent adds them and divides by nine. A tiny site that jumped counts as much as your largest site that dipped. A mean of changes answers "how did the typical company do", not "how much did portfolio clicks grow". The error is a sentence that claims one while the math did the other. The guide to benchmarking conversion rates across clients covers the case where one vote per company is exactly what you want.
3. Mixing totals and rates
Conversion rate can rise while conversions fall, and an agent may write "conversions improved" when only the rate did. Rates also cannot be averaged across sites. Search Console defines CTR as "(clicks ÷ impressions)", so a portfolio CTR is total clicks over total impressions, never the mean of nine CTRs. Average position is already "averaged across all queries in which your property appeared", per the same page, so averaging it again across companies means little.
4. Subtracting nested windows
The landing page, query and page tools give back one total per window. Asked for "the 90 days before", an agent may take the last6months total and subtract the last90days total. The leftover only matches 90 days if last6months is exactly 180 days long, which DuckDive's documentation does not say, and a page that ranks in one list but not the other leaves a gap that is part real change, part missing row.
A series beats subtraction. The GA4 overview has one row per day, so both 90-day windows can be cut from a single longer span. The Search Console summary has weekly points, so the fairest cut there is the latest 13 weeks against the 13 before: 91 days each, and the memo should say 13 weeks rather than 90 days.
5. The row cap hiding a page
Ranked lists stop at 100 rows. The GA4 landing page tool returns 25 rows unless the agent asks for more, and the website list returns 50 unless it asks for more. An agent that reads absence as zero will report a page with no traffic, or total a portfolio without the sites past row 50.
| Error | How it gets in | Caught at |
|---|---|---|
| Wrong window | Calendar words mapped to fixed windows | Step two, the window column |
| Averaged percentages | Mean of company changes called a total | Step three, recompute from totals |
| Totals mixed with rates | Rate movement described as volume | Step three, then step four labels |
| Nested subtraction | Six months minus 90 days on list tools | Step two, days per window |
| Row cap | Default or maximum rows treated as everything | Step four, rows returned versus asked |
Step one: confirm every source is complete
A source that is still importing looks exactly like a bad month. So does one that stopped syncing two weeks ago. get_sync_status gives each source a state (ready, importing or failing) along with the date its data runs to. list_websites shows which sources each site has linked, which catches the quieter problem: a company with no Search Console linked simply drops out of a portfolio total.
Ask your agentCheck sync status for all nine Tern Harbor Partners companies and tell me if any source is behind before we start the monthly memo.
Tools it calls list_websites get_sync_status
| Company | GA4 | Search Console |
|---|---|---|
| Kelpline Water Treatment | Importing, linked this morning | Ready, to Nov 14 |
| Sedgecomb Signage | Ready, to Nov 14 | Failing, last data Oct 29 |
| The other seven | Ready, to Nov 14 | Ready, to Nov 14 |
An importing source gets time; imports usually take a few minutes, longer for a long history. A failing source needs a fix first. refresh_website_data lets an admin pull between 1 and 30 recent days again, one request per website per quarter hour. If it still fails, check that your Google login can still see that property, and keep the company out of the totals with a footnote. A stale company must never sit silently inside a portfolio sum.
Step two: make the agent show the raw tool output
Prose hides arithmetic; a table of what each tool returned does not. Anthropic's guide to reducing hallucinations recommends making a response auditable by having the model cite a source for each claim, and says that if it cannot find support, "it must retract the claim." With marketing data, the source is the tool result. So ask for the raw table first and the story second.
Google Cloud's grounding overview defines grounding as "the ability to connect model output to verifiable sources of information." Your raw table is that connection, made visible.
Ask your agentFor each company, give me last 90 days against the 90 before for sessions, conversions and search clicks, as a table with the raw tool numbers.
Tools it calls list_websites get_ga4_overview get_search_console_summary
A raw table worth reviewing shows:
- The tool name behind every column.
- The window passed, plus its first and last date.
- The method for the earlier period: cut from a series, or one window minus another.
- Rows requested and rows returned for any ranked list.
- Unrounded counts, rounded once after every division.
The same Anthropic page suggests explicit permission to say "I don't know". So add: if a company has no data for a window, write "no data" in that cell.
Step three: recompute every derived number
Split the draft's numbers into two piles. Raw numbers you check against the step two table. Derived numbers you rebuild. Paste the raw table into a spreadsheet and redo each one, then have the agent do the same pass.
- Raw number
- A value exactly as a tool returned it: sessions, conversions, clicks or impressions for one site and one window.
- Derived number
- Any value the agent computed from raw numbers: a rate, a change, a share, a sum across companies.
- Pooled change
- Total across all companies now, against the total before. Answers "how did the portfolio move".
- Mean of changes
- Each company's percentage change, averaged with equal weight. Answers "how did the typical company move".
Ask your agentRecompute every percentage in your draft memo from the raw numbers and list any that do not match.
Tools it calls get_ga4_overview get_search_console_summary
Agreement is not proof
An agent checking its own memo can repeat its own mistake. Anthropic suggests running a prompt several times and treating disagreement as a warning sign. Treat that as an alarm; your spreadsheet pass is the sign-off.
Worked example: a memo with one wrong percentage
The analyst at Tern Harbor Partners, a fictional fund, has an agent-written search paragraph for the partners' memo. All nine companies passed the sync check once Sedgecomb's re-import finished. One figure is wrong.
Before this goes to the partners, rebuild the search paragraph from the raw clicks and tell me what doesn't hold: "Organic search clicks across the portfolio rose 9.4% over the last 90 days, led by Wherrybank Boatworks (+40%) and Tidewick Candle Co. (+36%). Our two largest sites, Foxwhistle Pet Supply and Saltfinch Valve Co., slipped about 2.5%."
list_websitesfound the nine Tern Harbor Partners company sitesget_search_console_summaryeach company, last12months, weekly points: the latest 13 weeks against the 13 before
Three of the four figures hold. The 9.4% does not. It is the average of the nine companies' percentage changes. Summed clicks went from 82,800 to 83,560, which is +0.9%. One label is also off: Search Console reports in weeks, so these windows are 13 weeks (91 days) each, and the memo should say 13 weeks, not 90 days.
| Company | Before | Latest | +/- |
|---|---|---|---|
| Foxwhistle Pet Supply | 22,900 | 22,310 | -2.6% |
| Saltfinch Valve Co. | 18,400 | 17,950 | -2.4% |
| Mossgate HVAC Supply | 14,300 | 13,870 | -3.0% |
| Pennyrock Insurance Services | 9,620 | 9,880 | +2.7% |
| Kelpline Water Treatment | 6,050 | 6,410 | +6.0% |
| Sedgecomb Signage | 4,780 | 4,920 | +2.9% |
| Hollowbeam Timber Frames | 3,360 | 3,520 | +4.8% |
| Wherrybank Boatworks | 2,150 | 3,010 | +40.0% |
| Tidewick Candle Co. | 1,240 | 1,690 | +36.3% |
| Portfolio, summed | 82,800 | 83,560 | +0.9% |
Wherrybank and Tidewick earn under 6% of portfolio clicks, yet their jumps drive most of the 9.4%. In clicks, they added 1,310, the three largest sites lost 1,470, and the other four added 920.
Suggested line: "Organic search clicks across the portfolio rose 0.9% over the latest 13 weeks, to 83,560. Wherrybank Boatworks (+40%) and Tidewick Candle Co. (+36%) grew fastest from small bases, while our three largest sites each slipped 2% to 3%."
Every raw number was right. The error lived in one division, the window carried the wrong name, and the draft also missed that Mossgate fell. Operating partners, the readers of the private equity portfolio analytics page, remember a headline figure long after any correction.
Put the raw table beside every draft
Link each company's GA4 and Search Console once, and your agent can show the counts behind any memo line before you send it.
Free for 2 sites · No credit card
Step four: trace each sentence to a tool call
Recomputing misses a sentence with no number behind it: a cause, a reason, a remark from the model's general knowledge. For that, make the agent sort its own output.
Ask your agentWhich numbers in your last answer came from a tool, and which did you calculate? Label each one.
Tools it calls none (it rereads the calls already in this chat)
Ask for three labels. Tool: copied from a call. Calculated: built from tool values, with the formula shown. Neither: an inference, a guess at a cause, or background from memory. Neither items get cut, or rewritten as an opinion with your name on it.
| Phrase in the draft | Label | Trace |
|---|---|---|
| "to 83,560" | Calculated | Nine companies' clicks, latest 13 weekly points each, summed |
| "Wherrybank Boatworks (+40%)" | Calculated | 3,010 against 2,150, same tool |
| "grew fastest from small bases" | Calculated | Ranked changes; bases from the table |
| "helped by boat show season" | Neither | Model's guess; no tool reads events |
That phrase crept into a later draft. It may be true, but no DuckDive tool knows when a boat show ran, so the reviewer confirms it with the company or deletes it. Plausible sentences are harder to spot than wrong numbers.
Worked example: a top-100 cap that hid a page
The group marketing lead asks about one page by name. The agent says it barely registered. The step four trace shows why that reply failed.
You said Foxwhistle's new /wholesale-signup page got no meaningful traffic in the last 30 days. Show me the rows you read. Is that a fact or a gap?
list_websitesfound Foxwhistle Pet Supplyget_ga4_landing_pageslast 30 days, the default 25 rows (the first answer)get_ga4_landing_pageslast 30 days, rerun at the 100-row maximum
It was a gap. My first answer read 25 rows and wrongly treated absence as no traffic. Rerun at 100 rows, the list came back full and /wholesale-signup is still not in it.
| Call | Rows back | Last row sessions | Found? |
|---|---|---|---|
| First answer | 25 | 1,140 | No |
| Rerun | 100 | 238 | No |
Both lists were full, so the page sits below row 100: fewer than 238 sessions, possibly none. I cannot tell which from DuckDive. For the exact figure, filter GA4's landing page report on the path.
Checked in GA4 by the reviewer: /wholesale-signup had 212 sessions and 31 conversions in the same 30 days, a 14.6% conversion rate, more than five times the 2.8% Foxwhistle's whole site converted at over that window. A ranked list sorted by sessions never showed it.
When rows returned equal rows requested, the list is cut off, and absence means "below the cut". The same holds one level up: a total across all clients starts from the website list, which stops at 50 unless the agent asks for more. An account lead with sixty clients, the reader of the QBR prep page for full-service agencies, should check the site count in every roster total.
Ask your agentBefore you tell me a page or a site isn't there: how many rows did you ask for, how many came back, and what was on the last row? If the list came back full, say "below the cut", not "no traffic".
Tools it calls list_websites get_ga4_landing_pages
Step five: freeze the prompts for next month
Let the agent rephrase a request and it may pick another window, another row limit or the mean instead of the pooled figure. The number moves while nothing in the business did. Keep a file of the exact prompts you ran and paste them each month, house rules first, before the agent touches a tool:
Ask your agentRules for this month's memo. The current window is last90days. For the period before it, never subtract one window's total from another's; cut it from a longer series. GA4 has daily rows. Search Console has weekly points, so compare the latest 13 weeks with the 13 before and call them 13 weeks. If a series does not reach back far enough, say so and mark that period approximate. Tell me the first and last date of each window. Ask for 100 rows on every list. Portfolio figures are summed totals, never the mean of company changes. Round to one decimal only after dividing. Label every number Tool, Calculated or Neither. If any source isn't ready, stop and tell me.
Tools it calls get_sync_status get_ga4_overview get_search_console_summary get_ga4_landing_pages
- Pin the company list. Name every company in the prompt and have the agent confirm each one, so a renamed site cannot fall out unnoticed.
- File each month's raw table beside the memo, so a question in March about a December figure has an answer.
If your exec team asks follow-ups between memos, the page for in-house teams answering execs with AI pairs well with these rules.
A reviewer's checklist to print
Before any agent-written number leaves your desk
- Sync is clean. Every source for every company reads ready, and none is days behind the rest.
- Windows are named and dated. No calendar words in the memo that the tools could not have read.
- The earlier period is the same length. Cut from a series, or flagged as approximate.
- Totals are summed, not averaged. Every portfolio percentage is rebuilt from summed counts.
- Words match the math. "Conversions rose" means the count rose, not the rate.
- No full list is read as complete. Rows returned are below rows requested, or the text says "below the cut".
- Every sentence has a label. Nothing marked Neither survives without a named human source.
- Headline figures redone by hand. The two or three numbers a reader will quote back are rebuilt in a spreadsheet.
Connecting the nine companies before the first check
Everything above depends on the agent having a source it can quote. DuckDive gives it one. Sign in with Google, link each company's GA4 property and Search Console site, and DuckDive copies their history into storage it keeps. The agent then reads each company through DuckDive's remote MCP server, one website per call.
DuckDive does not check the agent's arithmetic or make it right. It gives a reviewer a named tool behind every figure, so steps two to four have something to point at.
-
Sign in with Google
"Continue with Google" creates the account with no form. If your Google login cannot see a company's GA4 or Search Console yet, someone there adds you as a user.
-
Add each company as a website
Use the names the memo uses, so frozen prompts find them. Admins can let the agent type them in, provided "Also allow actions" stayed ticked when they connected it.
Give your agent this lineSet up DuckDive for my website. Follow https://analytics.duckdive.ai/
agents.md -
Link GA4 and Search Console for each
Every company gets a single GA4 property and a single Search Console site. A company that also pays for Google Ads or Local Services Ads should have those accounts connected as well. Each GA4 property can sit under only one DuckDive website.
-
Wait for the imports, then run step one
Linking starts the import. Draft nothing until
get_sync_statusshows every source ready. -
Connect the agent you write memos in
For Claude, the server address
https://api.duckdive.ai/mcpgoes under custom connectors; the Claude connection guide has the one-link version. Claude Code adds it from a terminal withclaude mcp add --transport http duckdive https://api.duckdive.ai/mcp, and/mcpinside a session starts the sign-in. Cursor installs it from the link in the agent setup guide. ChatGPT works on Business or Enterprise plans only. Each route lands on a DuckDive page; pick Continue with Google there. App menus get renamed, so look for "Connectors", "Apps" or "MCP".
The safe-connection guide for analytics agents covers who should hold which role.
Errors no protocol catches, and where DuckDive stops
- Bad tracking upstream. If a company's GA4 counts each form twice, every check passes on a wrong number.
- Exact search months and quarters. Weekly points blur the edges of a month, so a memo that must state "Q3 clicks" takes that figure from Search Console itself.
- Revenue, pipeline and CRM figures. DuckDive reads GA4, Search Console, Google Ads, Local Services Ads, Yext and Local Falcon, nothing else, so a revenue line has no tool to trace to.
- A fixed report nobody questions. If readers want the same layout monthly and never ask follow-ups, a dashboard may suit them better. The reporting dashboard versus agent data layer comparison helps you choose.
If you send one site's numbers to one boss a quarter, open GA4 and copy them. The protocol pays off when many companies meet a monthly deadline.
What reviewers ask before signing off
Does connecting DuckDive stop my agent from making numbers up?
No. It gives the agent stored data to read through named tools, which makes its numbers checkable. The agent can still pick the wrong window, average percentages or treat a capped list as complete. Anthropic's own guide says its techniques reduce hallucinations but do not eliminate them. That is why the checks exist.
Which numbers can I accept without recomputing?
Raw counts that match a row in the tool table from step two: sessions, conversions, clicks, impressions or spend for one company and one window. Anything divided, subtracted, summed or ranked gets rebuilt. In practice, rebuild every percentage that appears in a headline or a first sentence.
Why does my agent's figure differ from what I see in GA4?
Usually the window. The tools read fixed windows such as the last 30 days, while you may be looking at a calendar month in GA4. Recent days can also still change: Google says GA4 processing can take 24 to 48 hours. Ask the agent for the window's dates and check sync status before comparing.
Can I ask for October or the third quarter exactly?
Not as a window; every tool takes one of the six preset spans. For GA4 sessions and conversions, the agent can add up October's daily rows from a longer span, and you check that it used 31 of them. Search Console comes back in weekly points, so a calendar-month click figure from DuckDive is approximate. When a memo must state an exact quarter, take that number from GA4 or Search Console directly.
Does the reviewer need admin access in DuckDive?
No. Team members have read access, and reading is all the checking needs. Admin rights only matter for adding websites, linking sources, queuing a refresh or scheduling a report.
Where do the tool results go after my agent reads them?
Into your chat with the AI app. DuckDive keeps a record of each tool call (the tool, the website and the time), while the chat itself stays with the AI app. From there, that provider's own policies govern the results.
Sources
- Data freshness, Google Analytics Help (processing times table), checked November 16, 2026.
- What are impressions, position, and clicks?, Google Search Console Help, checked November 16, 2026.
- Reduce hallucinations, Claude documentation, Anthropic, checked November 16, 2026.
- Grounding overview, Google Cloud documentation, last updated November 12, 2026, checked November 16, 2026.
Last reviewed . Numbers in EXAMPLE DATA blocks are fictional.
