Analytics
Dale McGeorge
Head of Product
One URL connects your entire marketing stack to Claude, ChatGPT or any MCP client. What the eleven tools actually do, the five kinds of question that work, two worked examples with real numbers, and an honest account of where it still needs your judgement.
You can point Claude at your live marketing data by connecting the Jepto MCP server — one URL, https://mcp.jepto.com, that works with Claude, ChatGPT, Gemini, Cursor or any MCP-compatible client. Once connected, you can ask about campaign performance, budget pacing, KPI forecasts and client context in plain English, and get answers drawn from the same governed data model that powers Jepto's own reports.
This guide is written for agency people: account managers, strategists and owners who spend more of the week assembling data than acting on it. It covers how the connection works, what the tools actually do, the kinds of question that get good answers, and where a scheduled report is still the better tool.
Everything below was run against a live Jepto account. The numbers in the two worked examples are real returned values, not illustrations.
Model Context Protocol is an open standard for connecting AI assistants to external tools and data. Instead of pasting a CSV into a chat window, the assistant asks your data platform a question at the moment you ask yours, and gets a live answer back.
Jepto runs a remote MCP server at mcp.jepto.com. It's OAuth-secured and scoped to your account, so the assistant sees exactly the clients and data sources your login already has access to. There's nothing to install, no webhook to configure and no local process to keep running — you add the connector once and it persists across every future conversation in that assistant.
The important distinction is what sits behind the URL. The MCP server doesn't proxy you through to the Google Ads API. It queries the same semantic layer that Jepto's reporting runs on — a catalogued set of metrics and dimensions with agreed definitions, described in the Data Model docs. Your assistant discovers what's available before it queries, rather than guessing at field names.
Architecture
One URL. Any MCP client. One governed layer underneath.
The transport is standard. What makes the answers agree is what sits behind it.
OAuth-secured and scoped to your Jepto account. The assistant sees exactly what your login already sees — nothing more.
Eleven, in three groups. Your assistant picks the right one based on what you ask — you never name a tool yourself, and you don't need to know they exist. They're worth understanding anyway, because they tell you the shape of what's answerable.
The read tools cover both context and data. Client profile, business details, stored notes and connected sources on one side; schema discovery and analytics queries on the other. That schema step is the quiet one that matters: the assistant asks what fields exist before it queries them, which is what stops invented metric names before they happen.
Under the hood
Eleven tools. Three levels of consequence.
Read tools answer. Manage tools change things in Jepto. Act tools change things in your ad accounts — and always ask first.
The action catalogue is deliberately small today: Google Ads campaign enable and pause, both medium risk, both reversible, each declaring the other as its reverse. New actions appear automatically — nothing changes on your side.
You never name a tool. Knowing the three tiers exist is still useful, because it tells you what a question can reach.
Setup takes about a minute and there's no code involved. In Claude, go to Settings → Connectors → Browse directory and search for Jepto. In ChatGPT, open the apps directory and search there. For any client without a directory listing, find "Add custom connector" and paste https://mcp.jepto.com.
Sign in once with your Jepto account and the connection persists. In the composer, open the + menu → More → Jepto; the Jepto pill appears and stays active for that conversation. Then ask something.
If you don't have an account yet, start a free trial and connect a data source first — the server has nothing to query until at least one platform is connected. Full detail is in the MCP help doc.
There are five useful tiers, and they escalate in a way worth understanding, because most people stop at the first one and never reach where the leverage actually is.
Tier one is the daily check-in: one client, one platform, one metric. Tier two is where the unified data model earns its keep — because spend, clicks, conversions and ROAS are defined once and mapped across every connected platform, a single question can span Google Ads, Meta, Search Console and Google Business Profile without you reconciling anything by hand.
Tier three is the one specific to agencies, and the one generic AI connectors can't reach at all. Omit the client and the budget and KPI tools run account-wide, so one question covers everybody. Tier four adds the client's business context to the analysis. Tier five turns the answer into a deliverable — or into a change in the account.
Five tiers
Start narrow. The leverage is further up.
Each tier is available from day one. Most people never leave the first.
Tier 3 is the one that changes a Monday morning. It's also the one a per-platform connector structurally cannot do, because it has no concept of a client, let alone a portfolio.
Here is the tier-three question run against a live account, verbatim.
Asking which budgets need attention, without naming a client, returns all seven budgets across the account with a summary layer on top: a combined target of $7,205, combined spend to date of $697.12, and a breakdown by platform and status. Drilling into the one actually burning money — a $1,500 Google Ads budget running over 63 days — produces the numbers that make the flag mean something. Thirteen days elapsed, fifty remaining. Spend of $697.12, which is 46% of budget in 21% of the period. Daily average of $53.62 against a $16.06 target. Projected finish: $3,378.35, or 225% of target.
That's a paragraph you'd otherwise assemble by opening a dashboard, reading a chart, and doing the pacing arithmetic yourself. Here it arrives as the answer to one question, alongside the other six budgets, before you've opened a tab.
Live example · real returned values
Monday morning, one question, whole portfolio
No client named. The budget tool runs account-wide and summarises before it lists.
One budget is meaningfully off pace. The rest are flagged for underspend or have no delivery yet.
Figures returned by the Jepto MCP against a live account. Pacing, projection and days remaining are computed by Jepto, not by the model.
Spend is at 46% of budget with only 21% of the period gone. That gap is the whole answer.
Knowing a budget is overspending is half a job. Asking for the same client's campaigns ranked by spend, with conversions and cost per conversion alongside, returns a table across 108 campaigns — and three rows carry the story.
The biggest spender, a services and repair campaign, has consumed roughly $34,500 at a $31.51 cost per conversion. A product-brand campaign has spent about $30,400 at $10.90 — three times more efficient on nearly the same money. And a remarketing campaign has spent roughly $12,400 for 160 conversions at $77.64 each, with a click-through rate of 0.20%.
The assistant doesn't just return the table. Because the client profile came back in the same response — a family-owned Victorian HVAC business competing on energy efficiency, system design and finance options — the recommendation is framed for that business rather than for a generic advertiser. And the obvious move is visible without any analysis at all: the remarketing spend is buying conversions at seven times the cost of the campaigns sitting next to it.
Live example · real returned values
Where the money is going, and where it should go
Three rows from a 108-campaign account, ranked by spend. Cost per conversion tells a different story than spend does.
Bar length is cost per conversion, not spend. Same table, different question — which is exactly the reframing that takes five seconds in conversation and twenty minutes in a dashboard.
Only with your explicit confirmation, and the mechanism is worth understanding rather than trusting.
Action tools are split in two on purpose. The catalogue tool returns what's available, including each action's input schema, its risk level, whether it's reversible and what the reverse is. The execute tool then runs one — but the first call never executes anything. It returns a preview: the exact action, the ad account, the campaigns, and the model's stated reason for proposing it. Nothing touches your ad account until you confirm, and you can cancel instead.
Today the catalogue holds Google Ads campaign enable and pause. Both are marked medium risk, both reversible, each naming the other as its reverse action. More are being added; because the catalogue is discovered at runtime, new actions appear without any change on your side.
Two-step execution
Propose. Confirm. Then it happens.
The first call to the execute tool cannot change anything. It can only ask.
Risk level and reversibility are properties of the action, declared in the catalogue — not judgements the model makes at the time.
You could connect a Google Ads MCP, a Meta MCP and a GA4 MCP straight to your assistant and skip the middle layer entirely. For a solo consultant on one ad account, that's genuinely fine. It's worth being clear about what you give up as soon as you're past that.
Definitions stop agreeing. Google Ads conversions and Meta actions are not the same unit, and neither is clicks versus link clicks. Ask an assistant to compare them and it will produce a number, because producing a number is what it does. We wrote about that at length in You don't need a more expensive LLM. You need a data model, and it's the single strongest argument for putting a governed layer in the middle.
There's also no client layer. Platform MCPs know about ad accounts. They don't know that three ad accounts, a GA4 property, a Search Console domain and a Business Profile location all belong to one client, or what that client sells. Nothing is portfolio-shaped, either — a Google Ads MCP can tell you about a campaign; it can't tell you which of your twenty clients needs you today.
And it's per-conversation. Every new chat starts from zero. Jepto's data model, client profiles and knowledgebase are identical in Claude, in ChatGPT, and inside Data Chat.
Output quality tracks input specificity, and the pattern doesn't vary. Name the client, name the period, name the metrics that matter, and state the output you want.
A weak prompt — "how are my campaigns doing?" — leaves the assistant guessing at client, date range, platform and success criterion. A strong one — "Pull this client's Google Ads campaigns for the last 30 days, rank by spend, and flag anything over $5,000 with a cost per conversion above $50" — leaves nothing to interpretation.
Three habits help specifically here. Start with a verb: pull, compare, rank, draft, flag, forecast. Instructions outperform questions. Name the comparison, not just the metric — "CPA this month versus last" tells the assistant to run two queries and do the subtraction, rather than handing you one number to interpret alone. And ask for the check: "show me the query you ran" is a legitimate follow-up, and Jepto returns the resolved query alongside the data, so you can see exactly which sources and fields produced the answer.
They're for different jobs, and the fastest teams use both without thinking about it much.
Choosing a surface
Same data model. Three front doors.
The question isn't which is better. It's whether you need an answer, an artefact, or both.
Ad hoc questions where the answer is a decision, not a document. Monday triage, mid-meeting fact checks, late-night diagnosis.
Anything that must look the same every month, land in a client's inbox on a schedule, and carry your branding.
The conversational experience inside Jepto, with client switching, history, persistent memory and model choice per conversation.
All three resolve against the same unified data model, so the arithmetic underneath is identical. Nobody has to reconcile anything, because there was never more than one number.
Grounding an assistant in live data removes the largest source of error — invented numbers — but it doesn't remove the need for judgement, and it's worth being specific about where the remaining risk sits.
The retrieved figures are real. They come from your connected accounts through Jepto's data model, and the resolved query is returned with every response, so the retrieval step is auditable rather than opaque. If you want to see how a query is constructed, the Data Query doc walks through it.
Derived arithmetic is the assistant's. Percentage changes, blended averages and anything summed across rows are produced by the model, not by Jepto. Spot-check anything going in front of a client. Where a headline total matters, ask for it as a period total rather than letting the assistant add up a table.
Check the date range if a number looks large. Relative periods work well, but some query shapes drop the date dimension, and when that happens you get lifetime figures presented in a sentence about the last 30 days. It's a fast thing to verify and worth doing on any number you're about to quote.
The assistant can't see what isn't in the data. A campaign in a deliberate learning phase looks like an underperforming campaign. A CPA spike with a known cause looks like a problem. This is exactly what the client knowledgebase is for — the more brief and note context you store in Jepto, the less often the assistant reaches a conclusion that's obvious to you and invisible to it.
The working model is the one that applies to any competent junior analyst: let it do the retrieval and the first pass, then apply your own judgement before anything leaves the building.
No SQL, no API configuration, no understanding of MCP internals. If you can connect a data source in Jepto and paste a URL into a settings screen, you can do the whole setup.
What you do need is an account with at least one data source connected, an assistant that supports MCP, and the ability to tell whether an answer makes sense — which every experienced marketer already has. The one skill that measurably improves results is prompting, and it's learnable in an afternoon.
Any MCP-compatible client. Claude, ChatGPT and Gemini all have connector directories where Jepto is listed; for anything else, including custom internal tools, add https://mcp.jepto.com as a custom connector.
Everything connected to your Jepto account — Google Ads, Meta, Instagram, GA4, Search Console, Google Business Profile, Microsoft Ads, LinkedIn, TikTok and more. The full list is on the data sources page, and the assistant can tell you exactly which are connected for any given client.
Only with your explicit confirmation. Action tools return a preview first showing the exact action, account and campaigns; nothing executes until you confirm. Available actions currently cover Google Ads campaign enable and pause, both reversible, with more being added.
The connection is OAuth-secured and scoped to your Jepto account. Your assistant sees only what your Jepto login can already see, and the connection can be revoked from either side at any time.
No. It replaces the twenty minutes spent assembling an answer to a one-off question. Recurring, branded, client-facing reporting is still a job for Jepto Reports.
Same data, same MCP, different surface. Data Chat lives inside Jepto and adds persistent memory, client workspaces, conversation history and per-conversation model choice. The MCP brings your Jepto data into whichever assistant you already work in.
AI credits meter model usage inside Jepto's own Data Chat. When you use the MCP server from an external assistant, the model cost sits with that assistant's own plan.
Pick one client, one channel and one thing you're unsure about. Ask it. Look at what comes back, then ask the follow-up you'd have asked a colleague.
Most people find the shift isn't in the answers themselves. It's realising how much of the week was going into getting data into a shape where an answer was possible at all.
Connect Jepto to your AI assistant — or start a 14-day free trial and connect your first data source in minutes.
Thanks a lot for your feedback! If you’d like a member of our support team to respond to you, please send a message here
Oops! Something went wrong while submitting the form.
