The short answer
Buy by reader. Clients who need a finished, branded view need a dashboard. Analysts who rebuild numbers in Sheets, BI or a warehouse need a connector. People who ask new questions in Claude or ChatGPT need an agent data layer. MCP does not replace the other two. It sits beside them, and most teams need one or two layers, not three.
- Name your readers before you compare any vendor's feature list
- Write down what "conversion" means per site, whichever layer you buy
- DuckDive is the agent layer for Google marketing data; it builds no client reports
Start with who reads the numbers
The word muddying this purchase is MCP. The Model Context Protocol is, in its maintainers' words, "an open-source standard for connecting AI applications to external systems," and they compare it to "a USB-C port for AI applications." A dashboard can have that port, and so can a connector. So ask a different question first: who opens the numbers, and what do they do next?
- Readers who want a finished view
- Clients, franchise owners, a board. They will not write a query; they want the same layout monthly, with your logo. They need a reporting dashboard.
- Readers who want rows
- Analysts and finance people who join marketing data to revenue or CRM exports in Sheets, BI or a warehouse. They need a data connector, the pipe that lands platform data in their tool.
- Readers who want an answer
- Founders, account managers, operators and their AI agents, with questions that start with "why" or "which". They need an agent data layer: stored marketing data an agent reads through MCP tools.
A decision tree in five questions
Answer in order and keep a tally of the yeses. Each answer also marks where DuckDive, the agent layer sold here, stops.
-
Does anyone outside your team need to see the numbers?
Clients, investors, a mentor, a parent company. If yes, list a dashboard: you cannot hand an outsider your chat window. DuckDive will not serve this reader. It renders no branded screen, and the only scheduled email it sends goes to you, not your client.
-
Does someone rebuild the numbers in Sheets, BI or a warehouse?
If an analyst joins ad spend to invoices from billing, they need a connector. If nobody on payroll does that today, a connector is a pipe into an empty room. Need campaign rows or paid social in Sheets? That is the connector's job: DuckDive hands back written answers, not exportable tables, and reads Google Ads only as account totals.
-
Are most of your questions new each week?
"Why did enquiries fall while traffic held?" A fixed chart cannot anticipate that. If you ask questions like it most weeks, add an agent data layer. Check its clock first. Every DuckDive data tool counts back from today over 7, 14, 30 or 90 days, or over 6 or 12 months. If your questions always name a calendar month, a dashboard's date picker answers them more exactly.
-
How many websites do the questions cover, and through which sources?
With one site, almost anything works. With ten or forty, the hard question spans all of them, and dashboards are usually built per client. A layer that already stores each site's history handles that better; the stored history versus live MCP comparison shows why. Then read the source list: DuckDive covers GA4, Search Console, Google Ads, Local Services Ads, Yext and Local Falcon, so a roster living on social ads is a connector's job.
-
Who will keep it running?
A warehouse pipeline needs an owner for schemas and refresh jobs; a dashboard needs someone to fix broken widgets. An agent layer has no schema or widgets to maintain, but someone still owns each metric's meaning. It raises no alarms either; if you want a nudge when a number moves, pick a dashboard that lists alerting.
Reading your tally
Yes to 1 means a dashboard. Yes to 2 means a connector. Yes to 3, or many sites in 4, means an agent layer. Question 5 is the veto: if nobody can run a layer, cross it off.
What a dashboard does well
Dashboards win whenever the reader is not you. They look the same every month, carry your brand and open without anyone typing a question.
- A link a client can bookmark. DashThis' pricing page lists white-labeled dashboards on every plan: "We remove all of DashThis branding from all your dashboards" (as published on DashThis' pricing page, checked November 2026).
- Breadth at no cost. Google's free report builder is now titled Data Studio, and the old Looker Studio address redirects to it. Its product page promises to "directly connect to more than 1,400 data sources" and says it "is available at no charge for creators and report viewers" (checked November 2026).
- A report that arrives on its own. Google's scheduled delivery help page says it "lets you regularly send an email with the report attached as a PDF file," and that "You can include other recipients by entering individual addresses or group email aliases" (checked November 2026).
Dashboards are adding MCP too. DashThis carries the banner "MCP Connector available for all plans" and describes the feature as a way to "Connect Claude or ChatGPT to your dashboards in minutes" (DashThis pricing page, checked November 2026). That suits questions about data your dashboards already hold.
Where a dashboard beats DuckDive: whenever a client, franchisee or board member must see numbers without you in the room. DuckDive gives that reader no screen of their own.
What a connector does well
A connector moves platform data to where an analyst works. Its output is rows, not answers. Supermetrics' homepage puts the current pitch plainly: "Supermetrics connects your ad platforms, analytics and CRM to Claude, ChatGPT and Supermetrics Studio," and it names Google Sheets, Looker Studio, BigQuery and Snowflake as destinations (checked November 2026).
- Joins with data marketing does not own. Revenue, margin, CRM deals. Only rows in a shared place make those joins possible.
- Source breadth. Paid social, CRM and ecommerce are common reasons to buy one, and all sit outside what DuckDive reads.
The cost is a person to design tables and catch failed refreshes. See the Supermetrics comparison for AI agent work.
Where a connector beats DuckDive: when marketing numbers must sit beside finance or CRM data in a model you own, or most spend is on paid social.
What an agent data layer does well
An agent data layer serves the question nobody built a chart for. DuckDive is that layer for Google marketing data. You sign in with Google, then match each website to its GA4 property and Search Console site, and to its Google Ads and Local Services Ads accounts when it advertises there. Yext and Local Falcon sit outside that sign-in and get a link step of their own. DuckDive imports the history, keeps it, and answers agents over MCP from https://api.duckdive.ai/mcp.
Each tool, such as get_ga4_overview or get_search_console_summary, reads one website. For a question about many sites, the agent calls it once per site, and each call reads what DuckDive has already stored rather than asking Google again. Most questions it gets are small and arrive at awkward moments, like this one ten minutes before a call:
Ask your agentGive me one paragraph on how my main site did this month that I could read aloud on a call.
Tools it calls get_ga4_overview get_ga4_landing_pages get_search_console_summary
No dashboard tile writes a paragraph you can say out loud. Here the agent reads three tools and drafts it.
- No analyst needed for setup. Paste one prompt and the agent sets DuckDive up; you click Continue with Google once. The agent self-setup guide shows each step.
- Read-only by default. Actions need an admin account and your permission.
- A chat for people without an agent. Teammates who would rather not run their own assistant can ask Drake, the assistant inside DuckDive.
No rows out and no alerts. Nothing lands in a warehouse; the only push is a scheduled email report to your own address.
Metric definitions stay your job in all three
No layer knows what your business means by a conversion. DashThis makes this point in "MCP for Marketing Analytics: What It Actually Changes (And What It Doesn't)". Using return on ad spend, it warns that "It'll calculate ROAS using the formula it assumes, and hand you a number that looks correct" (as published on the DashThis blog, checked November 2026).
- Key event (GA4)
- What DuckDive's tools call conversions. Google's conversions and key events page calls it "an event that measures an action that's particularly important to the success of your business." You choose which events count.
- Counting method (GA4)
- Per Google's counting method page, "Once per event" counts every trigger and "Once per session" counts one per session. Changes apply to future data only, so a switch shows as a step in the trend.
- Conversion counting (Google Ads)
- Separate from GA4. Google's counting options page says "Every" counts each conversion after an ad interaction and "One" counts one per ad click, and suggests One for leads.
Through DuckDive, an agent sees GA4 counts by day, channel and landing page, not your key event settings. It can still spot the symptoms of an odd definition.
Ask your agentWhat does 'conversions' mean for each of my websites, and are any of them counting something odd?
Tools it calls list_websites get_ga4_overview get_ga4_landing_pages
What does 'conversions' mean for each of my websites, and are any of them counting something odd?
list_websitesthe five client sitesget_ga4_overvieweach site, sessions and conversions by day, last 90 daysget_ga4_landing_pageseach site, which landing pages the conversions came from
I can't see which events each property marks as key, only the counts. Three sites look odd.
| Site | Conversions | Rate |
|---|---|---|
| Fernbrake Cycles | 2,650 | 14.4% |
| Saffron Lane Yoga | 391 | 8.0% |
| Quillmere Lettings | 0 | 0.0% |
| Pellow & Stone Joinery | 96 | 2.8% |
| Tollgate Pottery | 212 | 2.3% |
Most of Fernbrake's conversions land on its /newsletter page, which fits a newsletter signup marked as key. Saffron Lane's daily conversions jumped from about 2 to about 7 on October 14 with flat sessions. Quillmere Lettings recorded none in 90 days. Check Key events in GA4 Admin for all three.
The agent said what it could not see. Now write the answers down once:
A one-page metric sheet, per website
- Key events. List them by name, such as a quote form submit, and nothing else.
- Counting methods. GA4 and Google Ads settings, and when each last changed.
- Which "organic". Search Console clicks or GA4 organic sessions. They measure different things and never match.
- Date windows. Calendar months or rolling 30 days. DuckDive's tools use rolling windows.
- Which layer may quote it. The figure a client hears, and the one only your team sees.
Give the sheet to every layer you run. To check agent figures against the platforms, see how to check your AI agent's marketing numbers.
Worked example: an agency that needs two of the three
Brindlewood Agency is a fictional eight-person shop with five client sites. Question 1 is yes: clients want a monthly view with the agency's logo. Question 3 is yes: account managers field new questions daily. Question 2 is no, so it skips the connector.
Two layers only work if everyone knows which one answers what. Brindlewood's house rule fits on a sticky note: any number a client will read comes from the dashboard; anything the team asks for itself goes to the agent; anything neither holds goes to the platform's own screen.
| Question, and who asked | Answered in | Why there |
|---|---|---|
| "Can I see September beside October?" (Saffron Lane's owner) | Branded dashboard | A client reads it, in calendar months |
| "Has Quillmere logged one enquiry since we fixed its form?" (account manager) | Agent | Internal, one-off, daily counts |
| "Which clients would notice first if organic kept sliding?" (agency founder) | Agent | Spans all five sites at once |
| "Which Fernbrake campaign spent most last week?" (paid media lead) | Google Ads itself | DuckDive stops at account totals |
| "What margin did each client leave after ad spend?" (founder, quarterly) | A spreadsheet, by hand | Needs billing data; too rare for a connector |
| "Why is your organic figure lower than mine?" (Tollgate's owner) | Both, via the metric sheet | Two layers, two definitions |
Two of the six questions landed outside both layers, which confirms skipping the connector. The third is where the layers met.
Ask your agentWhich of my clients would notice first if their organic traffic kept falling at this month's rate?
Tools it calls list_websites get_search_console_summary get_ga4_channels
Which of my clients would notice first if their organic traffic kept falling at this month's rate?
list_websitesthe five client sitesget_search_console_summaryeach site, weekly clicks over 90 days; the newest four weeks set against the four before themget_ga4_channelseach site, last 30 days by medium, to see what share of conversions arrives from organic search
Tollgate Pottery would feel it first. Its Google clicks fell 18% over the newest four weeks, and organic search brings 71% of its conversions. Pellow & Stone comes next.
| Client | Click change | Organic share |
|---|---|---|
| Tollgate Pottery | -18% | 71% |
| Pellow & Stone Joinery | -12% | 66% |
| Fernbrake Cycles | -22% | 24% |
| Quillmere Lettings | -9% | n/a |
| Saffron Lane Yoga | +4% | 38% |
Fernbrake's clicks drop fastest, but paid traffic brings most of its conversions, so its owner would feel the slide later. Quillmere cannot be ranked until its key events record something. The table is in order of exposure.
One client, two honest numbers
The founder then opened Tollgate's branded dashboard: October organic down 6%, not 18%. Nobody had made a mistake. The layers used different definitions, and only the metric sheet said so.
| Detail | Branded dashboard | Agent, via DuckDive |
|---|---|---|
| What it counts | GA4 sessions with medium "organic" | Search Console clicks |
| Window | October 1 to 31 | Newest four weekly points |
| Change shown | -6% | -18% |
| Who may quote it | Anyone, including the client | Brindlewood staff only |
Brindlewood fixed it on paper, not in either tool. Tollgate's sheet gained two lines: client-facing organic is GA4 organic sessions by calendar month, quoted from the dashboard; the internal early warning is Search Console clicks over four weeks, read by the agent and never quoted to the client. When the owner asked why the figures differed, the account manager read him those two lines.
The dashboard owns what Tollgate is told; the agent owns what Brindlewood notices first. Agencies already paying for a reporting platform with its own MCP should read the AgencyAnalytics comparison for AI agent work, and the wider picture is on DuckDive for agencies.
Keep your dashboard. Add the question layer beside it.
Start with one site at no cost and put the who-feels-it-first question to your agent.
Free for 2 sites · No credit card
Worked example: a founder who needs only one
Okapi Notes is a fictional one-person shop that sells hand-bound notebooks and paper refills from one website. Its founder packs orders with a part-time helper. Tally: no outside readers, no analyst, plenty of new questions. One yes, one layer. Here is a month of what she asked in Claude:
| Week | What the founder asked |
|---|---|
| 1 | Which landing pages turned the most visits into orders over the last 90 days? |
| 2 | Where did last week's orders come from: Google, my email list or the craft-fair flyer link? |
| 3 | Which Google searches show my shop but almost never get a click? |
| 4 | Is the A5 dot-grid notebook page still slipping in Google results? |
Each answer was used once and discarded, the pattern an agent layer suits. Then, in October, question 1 flipped.
The month a mentor asked for his own view
Her mentor, a retired bookshop owner she calls monthly, used to hear the read-aloud paragraph from earlier. He asked for something he could look at before each call instead. Now there is an outside reader, and she had two cheap ways to serve him.
| Detail | Free Data Studio report | Scheduled DuckDive email |
|---|---|---|
| Who receives it | Anyone she shares it with, or adds as an email recipient | Only her own DuckDive address; she forwards it |
| Setup | Build a report once with the Google Analytics and Search Console connectors | Ask her agent; it calls create_scheduled_report |
| Dates | Any range, calendar months included | Last 7, 10 or 30 days, or all time |
| Frequency | As often as daily without the Pro tier | Daily, weekly or monthly |
| Who maintains it | She does, when a chart breaks | She picks sources and range; DuckDive sets the layout |
Data Studio rows: scheduled delivery help page and Search Console connector page, checked November 2026. DuckDive rows: the create_scheduled_report tool, which needs an admin account with actions allowed.
She chose Data Studio, on the first row: a view the mentor receives directly beats one she must remember to forward, and it can say "October". For that reader Google's free tool is the better fit. She kept DuckDive for herself, plus a monthly scheduled email to her own inbox as a cue for follow-up questions.
Okapi Notes now runs two layers, and the second cost nothing. If the reader you want to serve is your own agent every morning rather than a mentor once a month, the guide to giving an AI chief of staff marketing data covers that setup.
Do I need all three?
Rarely: only with outside readers, an analyst who models data and people asking new questions weekly. The table includes rows where DuckDive is the wrong pick.
| Need | Reporting dashboard | Data connector | Agent data layer (DuckDive) |
|---|---|---|---|
| Branded view a client can open | Yes, its core job | Partly, through the report tool it feeds | No, it builds no client reports |
| Scheduled email to someone outside the team | Yes, Data Studio attaches a PDF | Partly, through its destination | No, it emails only your own address |
| Model joining marketing and finance data | Partly, blending inside the tool | Yes, rows land in Sheets or a warehouse | No, answers come back as text |
| A new "why" question on a Tuesday | Partly, if a chart already exists | Partly, once someone writes the query | Yes, ask in plain words |
| One question across many websites | Partly, often one client per dashboard | Partly, within plan account limits | Yes, reads each site's stored history |
| A named calendar month | Yes, date picker | Yes, set in the query | Partly, rolling windows only |
| Paid social and CRM data | Yes, varies by vendor | Yes, usually its main strength | No, six Google and local sources |
| Campaign-level Google Ads | Yes, varies by vendor | Yes, varies by vendor | No, account totals only |
| Running without an analyst | Yes, templates help | Partly, someone owns the destination | Yes, the agent can set itself up |
| Owns your metric definitions | No | No | No |
Sources: DashThis pricing, Google Cloud Data Studio page, Data Studio scheduled delivery help and Supermetrics homepage, checked November 2026. Category rows describe typical tools.
- Dashboard plus agent layer. Agencies and multi-location brands with outside readers and no analyst: the Brindlewood pattern, and Okapi Notes once its mentor asked.
- Connector plus agent layer. In-house teams whose analyst owns the models.
- Agent layer alone. Founders and operators whose readers all sit in chat, until someone outside asks to see.
- Dashboard alone. Questions never change. Fine, and cheap.
Questions buyers ask before choosing a category
Does MCP replace my reporting dashboard?
No. MCP is a standard way for an AI app to reach outside data. It changes how you ask, not who needs a finished view. If clients open a dashboard today, keep it.
My dashboard tool already has an MCP connector. Do I still need an agent data layer?
Maybe not. If your questions stay inside one client's dashboard at a time, the built-in MCP may cover them. An agent layer earns its place when questions span many websites or reach data no dashboard charts.
Is an agent data layer just a connector with a chat box?
The plumbing overlaps; the output differs. A connector mainly delivers rows for an analyst. An agent layer delivers small, named tool results, and the agent writes the answer. Many connectors now offer both.
Can my clients use DuckDive instead of a dashboard?
Not as a portal. DuckDive has no white-label view, and the only scheduled email it sends goes to you, not your client. Give clients a dashboard and let your team use DuckDive to prepare what they say about it.
What does adding the agent layer cost?
The dashboard or connector invoice you already pay does not change, because DuckDive sits beside those tools instead of taking a seat in them. The new line is DuckDive itself: the first website you link costs nothing and asks for no card, and each website after that moves you onto a paid plan, with today's figures in the pricing section below.
Sources
- What is the Model Context Protocol (MCP)?, Model Context Protocol, checked November 16, 2026.
- Pricing | DashThis, DashThis, checked November 16, 2026.
- Data Studio: Business Insights Visualizations, Google Cloud (the old Looker Studio address redirects here), checked November 16, 2026.
- Schedule automatic report delivery, Data Studio documentation, Google Cloud, checked November 16, 2026.
- Connect to Search Console, Data Studio documentation, Google Cloud, checked November 16, 2026.
- Marketing Intelligence Platform: Supermetrics, Supermetrics, checked November 16, 2026.
- MCP for Marketing Analytics: What It Actually Changes (And What It Doesn't), DashThis, checked November 16, 2026.
- Conversions vs. key events in Google Analytics, support.google.com (Analytics), checked November 16, 2026.
- Change the counting method of key events, support.google.com (Analytics), checked November 16, 2026.
- About conversion counting options, support.google.com (Google Ads), checked November 16, 2026.
Last reviewed . Numbers in EXAMPLE DATA blocks are fictional. DashThis, Supermetrics and Google product names are trademarks of their owners, named here only to describe published features.
